While markerless motion capture provided acceptable accuracy, no clear patterns emerged regarding the individual effects of surface properties on technique. This is most likely due to limitations such as sample size, lack of standardizing data set (players) across facilities, and limited control over player behavior. However, analyzing one individual's motion capture data across surfaces showed potential for distinguishing turning styles based on facility parameters.

The method in this thesis demonstrates the potential of markerless motion capture for injury prevention research in football. Despite inconclusive results on the individual facility parameter effects, the ability to distinguish player styles across surfaces suggests valuable future directions for investigating personalized risk factors and optimizing playing surfaces. Further research with larger, more diverse samples and a broader set of biomechanical and facility features could provide deeper insight into injury prevention strategies.

With over 53 million articles and 11 million images, Wikipedia is the greatest encyclopedia in history. The number of users is equally significant, with daily views surpassing 1 billion. Such an enormous system needs automation of tasks to make it possible for the volunteers to maintain. When it comes to textual data, there is a system based on machine learning called ORES providing automation to tasks such as article quality estimation and article topic routing. A visual counterpart system also needs to be developed to support tasks such as vandalism detection in images and for a better understanding of the visual data of Wikipedia. Researchers from the Wikimedia Foundation identified a hindrance to implementing the visual counterpart of ORES: the images of Wikipedia lack topical metadata. Thus, this work aims to develop a deep learning model that classifies images into a set of topics, which have been pre-determined in parallel work. State-of-the-art image classification models and other methods to mitigate the existing class imbalance are used. The conducted experiments show, among others, that: using the data that considers the hierarchy of labels performs better; resampling techniques are ineffective at mitigating imbalance due to the high label concurrence; sample-weighting improves metrics; and that initializing parameters as pre-trained on ImageNet rather than randomly yields better metrics. Moreover, we find interesting outlier labels that, despite having fewer samples, obtain better performance metrics, which is believed to be either due to bias from pre-training or simply more signal in the label. The distribution of the visual data predicted by the models displayed. Finally, some qualitative examples of the model predictions to some images are presented, proving the ability of the model to find correct labels that are missing in the ground truth

Point Cloud Registration with data measured from a photon-counting LIDAR sensor from a large distance (500 m - 1.5 km) is an expanding field. Data measuredfrom far is sparse and have low detail, which can make the registration processdifficult, and registering this type of data is fairly unexplored. In recent years,machine learning for point cloud registration has been explored with promisingresults. This work compares the performance of the point cloud registration algorithm Iterative Closest Point with state-of-the-art algorithms, with data froma photon-counting LIDAR sensor. The data was provided by the Swedish Defense Research Agency (FOI). The chosen state-of-the-art algorithms were thenon-learning-based Fast Global Registration and learning-based D3Feat and SpinNet. The results indicated that all state-of-the-art algorithms achieve a substantial increase in performance compared to the Iterative Closest Point method. Allthe state-of-the-art algorithms utilize their calculated features to obtain bettercorrespondence points and therefore, can achieve higher performance in pointcloud registration. D3Feat performed point cloud registration with the highestaccuracy of all the state-of-the-art algorithms and ICP.

In the area of Traffic Sign Recognition (TSR), deep learning models are trained to detect and classify images of traffic signs. The amount of data available to train these models is often limited, and collecting more data is time-consuming and expensive. A possible complement to traditional data acquisition, is to generate synthetic images with a generative machine learning model. This thesis investigates the use of denoising diffusion probabilistic models for generating synthetic data of one or multiple traffic sign classes, when providing different amount of real images for that class (classes). In the few-sample method, the number of images used was from 1 to 1000, and zero images were used in the zero-shot method. The results from the few-sample method show that combining synthetic images with real images when training a traffic sign classifier, increases the performance in 3 out of 6 investigated cases. The results indicate that the developed zero-shot method is useful if further refined, and potentially could enable generation of realistic images of signs not seen in the training data.

Detecting defects in industrially manufactured products is crucial to ensure their safety and quality. This process can be both expensive and error-prone if done manually, making automated solutions desirable. There is extensive research on industrial anomaly detection in images, but recent studies have shown that adding 3D information can increase the performance. This thesis aims to extend the 2D anomaly detection framework, PaDiM, to incorporate 3D information. The proposed methods combine RGB with depth maps or point clouds and the effects of using PointNet++ and vision transformers to extract features are investigated. The methods are evaluated on the MVTec 3D-AD public dataset using the metrics image AUROC, pixel AUROC and AUPRO, and on a small dataset collected with a Time-of-Flight sensor. This thesis concludes that the addition of 3D information improves the performance of PaDiM and vision transformers achieve the best results, scoring an average image AUROC of 86.2±0.2 on MVTec 3D-AD.

In the field of autonomous driving a common scenario is to apply deep learningmodels on camera feeds to provide information about the surroundings. A recenttrend is for such vision-based methods to be centralized, in that they fuse imagesfrom all cameras in one big model for a single comprehensive output. Designingand tuning such models is hard and time consuming, in both development andtraining. This thesis aims to reproduce the results of a paper about a centralizedvision-based model performing 3D object detection, called BEVDet. Additionalgoals are to ablate the technique of class balanced grouping and sampling usedin the model, to tune the model to improve generalization, and to change thedetection head of the model to a Transformer decoder-based head.

The findings include a successful reproduction of the results of the paper,while adding depth supervision to BEVDet establishes a baseline for the subsequentexperiments. An increasing validation loss during most of the training indicatesthat there is room for improvement in the generalization of the model. Severaldifferent methods are tested in order to resolve the increasing validation loss,but they all fail to do so. The ablation study shows that the class balanced groupingis important for the performance of the chosen configuration of the model,while the class balanced sampling does not contribute significantly. Without extensivetuning the replacement head gives performance similar to the PETR, themodel that the head is adapted from, but fails to match the performance of thebaseline model. In addition, the model with the Transformer decoder-based headshows a converging validation loss, unlike the baseline model.

The goal of this thesis is to use fringe-pattern phase analysis to calibrate the distortion of a camera lens. The benefit of using this method is that the distortion can be calculated using data from each individual pixel and the methodology does not need any model.

The phase used to calibrate the images is calculated in two different ways, either utilizing the monogenic signal or through fringe-pattern phase analysis.

The calibration approaches were also validated through different methods. Primarily by utilizing the Hough transform and calibrating simulated distortion. The thesis also introduces a validation approach utilizing the phase orientation calculated through the monogenic signal.

The thesis also implements different approaches such as flat field correction to limit the impact of the image sensor noise to mitigate the phase noise.

It is also investigated which fringe-pattern frequencies are best suited for calibration through comparative analysis. The comparative analysis identified problems with too high and low frequencies of the fringe-patterns when calibrating using fringe-pattern phase analysis.

This thesis explores the integration of deep learning-based depth estimation models with the ORB-SLAM3 framework to address challenges in monocular Simultaneous Localization and Mapping (SLAM), particularly focusing on pure rotational movements. The study investigates the viability of using pre-trained generic depth estimation networks, and hybrid combinations of these networks, to replace traditional depth sensors and improve scale accuracy in SLAM systems. A series of experiments are conducted outdoors, utilizing a custom camera setup designed to isolate pure rotational movements. The analysis involves assessing each model's impact on the SLAM process as well as performance indicators (KPIs) on both depth estimation and 3D tracking. Results indicate a correlation between depth estimation accuracy and SLAM performance, underscoring the potential of depth estimation models in enhancing SLAM systems. The findings contribute to the understanding of the role of monocular depth estimation in integrating with SLAM, especially in applications requiring precise spatial awareness for augmented reality.

When a camera system in a car is mounted behind the windshield, light rays will be refracted by the windshield. The distortion can be significant, especially for wide field-of-view cameras. Traditional approaches handle the windshield distortion along with the calibration that calculates the intrinsic and extrinsic parameters. However, these approaches do not handle the windshield distortion explicitly, and to understand the image formation, it brings to understand more about the windshield distortion effect. In this thesis, data is collected from a camera system viewed with and without the windshield. The windshield distortion effect has been studied by varying the windshield’s tilt and the camera’s setup. Points are then found in both images and matched. From this, a distortion difference is calculated and analyzed. Next, a preliminary model of the windshield distortion effect is presented and evaluated. The results show that the model works well for all cases and the two windshields considered in this thesis. 

Radiologists often have to look through many different patients and examinations in quick succession, and to aid in the workflow the different types of images should be presented for the radiologist in the same manner and order between each new examination. Thus decreasing the time needed for the radiologist to either find the correct image or rearrange the images to their liking. A step in thisprocess requires a comparison between two images to be made and produce a score between 0-1 describing how similar the images are. A similar algorithm already exists at Sectra, but that algorithm only uses the metadata from the images without considering the actual pixel data.

The aim of this thesis were to explore different methods of doing the same comparison as the previous algorithm but only using the pixel data. Considering only 3D volumes from CT examinations of the abdomen and thorax region, this thesis explores the possibility of using SSIM, SIFT and SIFT together with a histogram comparison using the Bhattacharyya distance for this task. It was deemed very important that the ranking produced when ordering the images in terms of similarity to one reference image followed a specific order. This order was determined by consulting personnel at Sectra that works closely with the clinical side of radiology.

SSIM were able to differentiate between different plane orientations since they usually had large resolution differences in each led, but it could not be made tofollow the desired ranking and was thus disregarded as a reliable option for this problem. The method using SIFT followed the desired ranking better, but struggled a lot with differentiating between the different contrast phases. A histogram component were also added to this method, which increased the accuracy and improved the ranking. Although, further development is still needed for thismethod to be a reliable option that could be used in a clinical setting. 

An autonomous vehicle is a complex system that requires a good perception of the surrounding environment to operate safely. One part of that is multiple object tracking, which is an essential component in camera-based perception whose responsibility is to estimate object motion from a sequence of images. This requires an association problem to be solved where newly estimated object positions are mapped to previously predicted trajectories, for which different solution strategies exist. 

In this work, a multiple hypothesis tracking algorithm is implemented. The purpose is to demonstrate that measurement associations are improved compared to less compute-intensive alternatives. It was shown that the implemented algorithm performed 13 percent better than an intersection over union tracker when evaluated using a standard evaluation metric.

Furthermore, this work also investigates the usage of abstraction layers to accelerate time-critical parallel operations on the GPU. It was found that the execution time of the tracking algorithm could be reduced by 42 percent by replacing four functions with implementations written in the purely functional array language Futhark. Finally, it was shown that a GPU code abstraction layer can reduce the knowledge barrier required to write efficient CUDA kernels.

Ceramic materials contain several defects, one of which is porosity. At the time of writing, porosity measurement is a manual and time-consuming process performed by a human operator. With advances in deep learning for computer vision, this thesis explores to what degree convolutional neural networks and semantic segmentation can reliably measure porosity from microscope images. Combining classical image processing techniques with deep learning, images were automatically labeled and then used for training semantic segmentation neural networks leveraging transfer learning. Deep learning-based methods were more robust and could more reliably identify porosity in a larger variety of images than solely relying on classical image processing techniques.

Unmanned aerial vehicles (UAVs) with high-resolution cameras are common in today’s society. Industries, such as the forestry industry, use drones to get a fast overview of tree populations. More advanced sensors, such as near-infrared light or depth data, can increase the amount of information that UAV images provide, providing information about the forest, such as; tree quantity or forest health. However, the fast-expanding field of deep learning could help expand the information acquired using only RGB cameras. Three deep learning models, FasterR-CNN, RetinaNet, and YOLOR were compared to investigate this. It was also investigated if initializing the models using transfer learning from the MS COCO dataset could increase the performance of the models. The dataset used was Swedish Forest Agency (2021): Forest Damages-Spruce Bark Beetle 1.0 National Forest Data Lab and drone images provided by IT-Bolaget Per & Per. The deep learning models were to detect five different tree species; spruce, pine, birch, aspen, and others. The results show potential for the usage of deep learning to detect tree species in images from UAVs.

In recent years, the EU has observed a decrease in the stocks of certain fish species due to unrestricted fishing. To combat the problem, many fisheries are investigating how to automatically estimate the catch size and composition using sensors onboard the vessels. Yet, measuring the size of fish in marine imagery is a difficult task. The images generally suffer from complex conditions caused by cluttered fish, motion blur and dirty sensors.

In this thesis, we propose a novel method for automatic measurement of fish size that can enable measuring both visible and occluded fish. We use a Mask R-CNN to segment the visible regions of the fish, and then fill in the shape of the occluded fish using a U-Net. We train the U-Net to perform shape completion in a semi-supervised manner, by simulating occlusions on an open-source fish dataset. Different to previous shape completion work, we teach the U-Net when to fill in the shape and not by including a small portion of fully visible fish in the input training data.

Our results show that our proposed method succeeds to fill in the shape of the synthetically occluded fish as well as of some of the cluttered fish in real marine imagery. We achieve an mIoU score of 93.9 % on 1 000 synthetic test images and present qualitative results on real images captured onboard a fishing vessel. The qualitative results show that the U-Net can fill in the shapes of lightly occluded fish, but struggles when the tail fin is hidden and only parts of the fish body is visible. This task is difficult even for a human, and the performance could perhaps be increased by including the fish appearance in the shape completion task. The simulation-to-reality gap could perhaps also be reduced by finetuning the U-Net on some real occlusions, which could increase the performance on the heavy occlusions in the real marine imagery.

In recent years, pictures from handheld devices such as smartphones have been increasingly utilized as a documentation tool by medical practitioners not trained to take professional photographs. Similarly to the other types of image modalities, the images should be taken in a way to capture the vital information in the region of interest. Nevertheless, image capturing cannot always be done as desired, so images may exhibit different blur types at the region of interest. Having blurry images does not serve medical purposes, therefore, the patients might have to schedule a second appointment several days later to retake the images. A solution to this problem is to create an algorithm which immediately after capturing an image determines if it is medically useful and notifies the user of the result. The algorithm needs to perform the analysis at a reasonable speed, and at best, with a limited number of operations to make the calculations directly in the smartphone device. A large number of medical images must be available to create such an algorithm. Medical images are difficult to acquire, and it is specifically difficult to acquire blurry images since they are usually deleted.

The main objective of this thesis is to determine the medical usefulness of images taken with smartphone cameras, using both machine learning and handcrafted algorithms, with a low number of floating point operations and a high performance. Seven different algorithms (one hand-crafted and six machine learned) are created and compared regarding both number of floating point operations and performance. Fast Walsh-Hadamard transforms are the basis of the hand-crafted algorithm. The employed machine learning algorithms are both based on common convolutional neural networks (MobileNetV3 and ResNet50) and on our own designs. The issue with the low number of medical images acquired is solved by training the machine learning models on a synthetic dataset, where the non-medically useful images are generated by applying blur on the medically useful images. These models do, however, undergo evaluation using a real dataset, containing medically useful images as well as non-medically useful images.

Our results implicate that a real-time determination of the medical usefulness of images is possible on handheld devices, since our machine learned model DeepLAD-Net reaches the highest accuracy with 42 · 10⁶ floating point operations. In terms of accuracy, MobileNetV3-large is the second best model with31 times as many floating point operations as our best model.

This master thesis project was done together with Saab Dynamics in Linköping the spring of 2022 and aims to perform an online IMU-camera calibration using an AprilTag board. Experiments are conducted on two different types of datasets, the public dataset Euroc and internal datasets from Saab. The calibration is done iteratively by solving a series of nonlinear optimization problems without any initial knowledge of the sensor configuration. The method is largely based on work by Huang and collaborators. Other than just finding the transformation between the IMU and the camera, the biases in the IMU, and the time delay between the two sensors are also explored. By comparing the resulting transformation with Kalibr, the current state of the art offline calibration toolbox, it is possible to conclude that the model can find and correct for the biases in the gyroscope. Therefore it is important to include these biases in the model. The model is able to roughly find the time shift between the two sensors but has more difficulties correcting for it. The thesis also aims to explore ways of compiling a good dataset for calibration. Results show that it is desirable to avoid rapid movements as well as images gathered at distances from the AprilTag board that very a lot. Also, having a shorter exposure time is useful to not lose AprilTag detections.

Perception of depth, ego-motion and robust keypoints is critical for SLAM andstructure from motion applications. Neural networks have achieved great perfor-mance in perception tasks in recent years. But collecting labeled data for super-vised training is labor intensive and costly. This thesis explores recent methodsin unsupervised training of neural networks that can predict depth, ego-motion,keypoints and do geometric consensus maximization. The benefit of unsuper-vised training is that the networks can learn from raw data collected from thecamera sensor, instead of labeled data. The thesis focuses on training on imagesfrom a monocular camera, where no stereo or LIDAR data is available. The exper-iments compare different techniques for depth and ego-motion prediction fromprevious research, and shows how the techniques can be combined successfully.A keypoint prediction network is evaluated and its performance is comparedwith the ORB detector provided by OpenCV. A geometric consensus network isalso implemented and its performance is compared with the RANSAC algorithmin OpenCV. The consensus maximization network is trained on the output of thekeypoint prediction network. For future work it is suggested that all networkscould be combined and trained jointly to reach a better overall performance. Theresults show (1) which techniques in unsupervised depth prediction are most ef-fective, (2) that the keypoint predicting network outperformed the ORB detector,and (3) that the consensus maximization network was able to classify outlierswith comparable performance to the RANSAC algorithm of OpenCV.

Instance segmentation has a great potential for improving the current state of littering by autonomously detecting and segmenting different categories of litter. With this information, litter could, for example, be geotagged to aid litter pickers or to give precise locational information to unmanned vehicles for autonomous litter collection. Land-based litter instance segmentation is a relatively unexplored field, and this study aims to give a comparison of the instance segmentation models Mask R-CNN and DetectoRS using the multiclass litter dataset called Trash Annotations in Context (TACO) in conjunction with the Common Objects in Context precision and recall scores. TACO is an imbalanced dataset, and therefore imbalanced data-handling is addressed, exercising a second-order relation iterative stratified split, and additionally oversampling when training Mask R-CNN. Mask R-CNN without oversampling resulted in a segmentation of 0.127 mAP, and with oversampling 0.163 mAP. DetectoRS achieved 0.167 segmentation mAP, and improves the segmentation mAP of small objects most noticeably, with a factor of at least 2, which is important within the litter domain since small objects such as cigarettes are overrepresented. In contrast, oversampling with Mask R-CNN does not seem to improve the general precision of small and medium objects, but only improves the detection of large objects. It is concluded that DetectoRS improves results compared to Mask R-CNN, as well does oversampling. However, using a dataset that cannot have an all-class representation for train, validation, and test splits, together with an iterative stratification that does not guarantee all-class representations, makes it hard for future works to do exact comparisons to this study. Results are therefore approximate considering using all categories since 12 categories are missing from the test set, where 4 of those were impossible to split into train, validation, and test set. Further image collection and annotation to mitigate the imbalance would most noticeably improve results since results depend on class-averaged values. Doing oversampling with DetectoRS would also help improve results. There is also the option to combine the two datasets TACO and MJU-Waste to enforce training of more categories.

Reconstruction of sonar images is an inverse problem, which is normally solved with model-based methods. These methods may introduce undesired artifacts called angular and range leakage into the reconstruction. In this thesis, a method called Learned Primal-Dual Reconstruction, which combines a data-driven and a model-based approach, is used to investigate the use of data-driven methods for reconstruction within sonar imaging. The method uses primal and dual variables inspired by classical optimization methods where parts are replaced by convolutional neural networks to iteratively find a solution to the reconstruction problem. The network is trained and validated with synthetic data on eight models with different architectures and training parameters. The models are evaluated on measurement data and the results are compared with those from a purely model-based method. Reconstructions performed on synthetic data, where a ground truth image is available, show that it is possible to achieve reconstructions with the data-driven method that have less leakage than reconstructions from the model-based method. For reconstructions performed on measurement data where no ground truth is available, some variants of the learned model achieve a good result with less leakage.

Generic visual object tracking is the task of tracking one or several objects in all frames in a video, knowing only the location and size of the target in the initial frame. Visual tracking can be carried out in both the infrared and the visual spectrum simultaneously, this is known as multi-modal tracking. Utilizing both spectra can result in a more diverse tracker since visual tracking in infrared imagery makes it possible to detect objects even in poor visibility or in complete darkness. However, infrared imagery lacks the number of details that are present in visual images. A common method for visual tracking is to use discriminative correlation filters (DCF). These correlation filters are then used to detect an object in every frame of an image sequence. This thesis focuses on investigating aspects of a DCF based tracker, operating in the two different modalities, infrared and visual imagery. First, it was investigated whether the tracking benefits from using two channels instead of one and what happens to the tracking result if one of those channels is degraded by an external cause. It was also investigated if the addition of image features can further improve the tracking. The result shows that the tracking improves when using two channels instead of only using a single channel. It also shows that utilizing two channels is a good way to create a robust tracker, which is still able to perform even though one of the channels is degraded. Using deep features, extracted from a pre-trained convolutional neural network, was the image feature improving the tracking the most, although the implementation of the deep features made the tracking significantly slower. 

Learning-based multi-view stereo (MVS) has shown promising results in the domain of general 3D reconstruction. However, no work before this thesis has applied learning-based MVS to urban 3D reconstruction from satellite images. In this thesis, learning-based MVS is used to infer depth maps from satellite images. Models are trained on both synthetic and real satellite images from Las Vegas with ground truth data from a high-resolution aerial-based 3D model. This thesis also evaluates different methods for reconstructing digital surface models (DSM) and compares them to existing satellite-based 3D models at Maxar Technologies. The DSMs are created by either post-processing point clouds obtained from predicted depth maps or by an end-to-end approach where the depth map for an orthographic satellite image is predicted. 

This thesis concludes that learning-based MVS can be used to predict accurate depth maps. Models trained on synthetic data yielded relatively good results, but not nearly as good as for models trained on real satellite images. The trained models also generalize relatively well to cities not present in training. This thesis also concludes that the reconstructed DSMs achieve better quantitative results than the existing 3D model in Las Vegas and similar results for the test sets from other cities. Compared to ground truth, the best-performing method achieved an L1 and L2 error of 14 % and 29 % lower than Maxar's current 3D model, respectively. The method that uses a point cloud as an intermediate step achieves better quantitative results compared to the end-to-end system. Very promising qualitative results are achieved with the proposed methods, especially when utilizing an end-to-end approach. 

Image segmentation through neural networks and deep learning have, in the recent decade, become a successful tool for automated decision-making. For Luossavaara-Kiirunavaara Aktiebolag (LKAB), this means identifying the amount of slag inside a furnace through computer vision. 

There are many prominent convolutional neural network architectures in the literature, and this thesis explores two: a modified U-Net and the PSPNet. The architectures were combined with three loss functions and three class weighting schemes resulting in 18 model configurations that were evaluated and compared. This thesis also explores transfer learning techniques for neural networks tasked with identifying slag in images from inside a furnace. The benefit of transfer learning is that the network can learn to find features from already labeled data of another context. Finally, the thesis explored how temporal information could be utilised by adding an LSTM layer to a model taking pairs of images as input, instead of one.

The results show (1) that the PSPNet outperformed the U-Net for all tested configurations in all relevant metrics, (2) that the model is able to find more complex features while converging quicker by using transfer learning, and (3) that utilising temporal information reduced the variance of the predictions, and that the modified PSPNet using an LSTM layer showed promise in handling images with outlying characteristics. 

This thesis investigates the possibility of utilizing data from multiple modalities to enable an automated recycling system to separate ferrous from non-ferrous debris. The two methods sensor fusion and hallucinogenic sensor fusion were implemented in a four-step approach of deep CNNs. Sensor fusion implies that multiple modalities are run simultaneously during the operation of the system.The individual outputs are further fused, and the joint performance expects to be superior to having only one of the sensors. In hallucinogenic sensor fusion, the goal is to achieve the benefits of sensor fusion in respect to cost and complexity even when one of the modalities is reduced from the system. This is achieved by leveraging data from a more complex modality onto a simpler one in a student/teacher approach. As a result, the teacher modality will train the student sensor to hallucinate features beyond its visual spectra. Based on the results of a performed prestudy involving multiple types of modalities, a hyperspectral sensor was deployed as the teacher to complement a simple RGB camera. Three studies involving differently composed datasets were further conducted to evaluate the effectiveness of the methods. The results show that the joint performance of a hyperspectral sensor and an RGB camera is superior to both individual dispatches. It can also be concluded that training a network with hyperspectral images can improve the classification accuracy when operating with only RGB data. However, the addition of a hyperspectral sensor might be considered as superfluous as this report shows that the standardized shapes of industrial debris enable a single RGB to achieve an accuracy above 90%. The material used in this thesis can also be concluded to be suboptimal for hyperspectral analysis. Compared to the vegetation scenes, only a limited amount of additional data could be obtained by including wavelengths besides the ones representing red, green and blue. 

Deep learning methods for medical image segmentation are hindered by the lack of training data. This thesis aims to develop a method that overcomes this problem. Basic U-net trained on XCAT phantom data was tested first. The segmentation results were unsatisfactory even when artificial quantum noise was added. As a workaround, CycleGAN was used to add tissue textures to the XCAT phantom images by analyzing patient CT images. The generated images were used totrain the network. The textures introduced by CycleGAN improved the segmentation, but some errors remained. Basic U-net was replaced with Attention U-net, which further improved the segmentation. More work is needed to fine-tune and thoroughly evaluate the method. The results obtained so far demonstrate the potential of this method for the segmentation of medical images. The proposed algorithms may be used in iterative image reconstruction algorithms in multi-energy computed tomography.

In the glass wool industry, the molten glass flow is monitored for regulation purposes. Given the progress in the computer vision field, the current monitoring solution might be replaced by a camera based solution. The aim of this thesis is to investigate the possibility of using optical flow techniques for estimation of the molten glass flow displacement.

Three glass melt flow datasets were recorded, as well as two additional melt flow datasets, using a NIR camera. The block matching techniques Full Search (FS) and Adaptive Rood Pattern Search (ARPS), as well as the local feature methods ORB and A-KAZE were considered. These four techniques were compared to RAFT, the state-of-the-art approach for optical flow estimation, using available pre-trained models, as well as an approach of using the tracking method ECO for the optical flow estimation.

The methods have been evaluated using the metrics MAE, MSE, and SSIM to compare the warped flow to the target image. In addition, ground truth for 50 frames from each dataset was manually annotated as to use the optical flow metric End-Point Error. To investigate the computational complexity the average computational time per frame was calculated.

The investigation found that RAFT does not perform well on the given data, due to the large displacements of the flows. For simulated displacements of up to about 100 pixels at full resolution, the performance is satisfactory, with results comparable to the traditional methods.

Using ECO for optical flow estimation encounters similar problems as RAFT, where the large displacement proved challenging for the tracker. Simulating smaller motions of up to 60 pixels resulted in good performance, though computation time of the used implementation is much too high for a real-time implementation.

The four traditional block matching and local feature approaches examined in this thesis outperform the state-of-the-art approaches. FS, ARPS, A-KAZE, and ORB all have similar performance on the glass flow datasets, whereas the block matching approaches fail on the alternative melt flow data as the template extraction approach is inadequate. The two local feature approaches, though working reasonably well on all datasets given full resolution, struggle to identify features on down-sampled data. This might be mitigated by fine-tuning the settings of the methods. Generally, ORB mostly outperforms A-KAZE with respect to the evaluation metrics, and is considerably faster.

In Storytel’s application on which a user can read and listen to digitalized literature, a user is displayed a list of books where the first thing the user encounters is the book title and cover. A book cover is therefore essential to attract a consumer’s attention. In this study, we take a data-driven approach to investigate the design principles for book covers through deep learning models and explainable AI. The first aim is to explore how well a Convolutional Neural Network (CNN) can interpret and classify a book cover image according to its genre in a multi-class classification task. The second aim is to increase model interpretability and investigate model feature to genre correlations. With the help of the explanatory artificial intelligence method Gradient-weighted Class Activation Map (Grad-CAM), we analyze the pixel-wise contribution to the model prediction. In addition, object detection by YOLOv3 was implemented to investigate which objects are detectable and reoccurring in the book covers. An interplay between Grad-CAM and YOLOv3 was used to investigate how identified objects and features correlate to a specific book genre and ultimately answer what makes a good book cover. Using a State-of-the-Art CNN model architecture we achieve an accuracy of 48% with the best class-wise accuracies for genres Erotica, Economy & Business and Children with accuracies 73%, 67% and 66%. Quantitative results from the Grad-CAM and YOLOv3 interplay show some strong associations between objects and genres, while indicating weak associations between abstract design principles and genres. Furthermore, a qualitative analysis of Grad-CAM visualizations show strong relevance of certain objects and text fonts for specific book genres. It was also observed that the portrayal of a feature was relevant for the model prediction of certain genres.

Point set registration is a well-researched yet still not a very exploited area in computer vision. As the field of machine learning grows, the possibilities of application expand. This thesis investigates the possibility to expand an already implemented probabilistic machine learning approach to point set registration to more complex, larger datasets gathered in a forest environment. The system used as a starting point was created by Järemo Lawin et. al. [10]. The aim of the thesis was to investigate the possibility to register the forest data with the existing system, without ground-truth poses, with different optimizers, and to implement a SLAM pipeline. Also, older methods were used as a benchmark for evaluation, more specifically iterative closest point(ICP) and fast global registration(FGR).To enable the gathered data to be processed by the registration algorithms, preprocessing was required. Transforming the data points from the coordinate system of the sensor to world relative coordinates via LiDAR base coordinates. Subsequently, the registration was performed with different approaches. Both the KITTI odometry dataset, which RLLReg originally was evaluated with[10], and the gathered forest data were used. Data augmentation was utilized to enable ground-truth-independent training and to increase diversity in the data. In addition, the registration results were used to create a SLAM-pipeline, enabling mapping and localization in the scanned areas. The results showed great potential for using RLLReg to register forest scenes compared to other, older, approaches. Especially, the lack of ground-truth was manageable using data augmentation to create training data. Moreover, there was no evidence that AdaBound improves the system when replacing the Adam-optimizer. Finally, forest models with sensor paths plotted were generated with decent results. However, a potential for post-processing with further refinement is possible. Nevertheless, the possibility of point set registration and LiDAR-SLAM using machine learning has been confirmed.

This thesis investigates methods for automatic colour transfer when working withgeodata and possible metrics to evaluate the results. Several methods for colourtransfer as well as methods to create an objective measurement were tested. Themethod was evaluated by using a subjective score which was generated by surveyingeight people working with geodata. In the survey the participants were askedto “Rank the images from most similar to least similar, with what you imagine theresult would have been if you would have made the colour transfer manually”.The method with the best overall performance in this study was using colourtransfer in the CIEl   colour space. This method was only matched by a methodsegmenting the image first based on colour information. As the method had thehighest average subjective score but a larger standard deviation than other methods.This was suspected to be largely due to the deviation in quality of the segmentationalgorithm. Using a different method for segmenting the image thismethod might perform even better.The objective measurements proposed in this study were not found to have aconsistent correlation with the subjective measurement, with the exception ofgradient structural similarity. Other methods could have a use in some cases butnot as general colour transfer objective measurement, though a larger study andmore data would be needed to confirm the findings.

The performance of conventional deep neural networks tends to degrade when a domain shift is introduced, such as collecting data from a new site. Model-Agnostic Meta-Learning, or MAML, has achieved state-of-the-art performance in few-shot learning by finding initial parameters that adapt easily for new tasks.

This thesis studies MAML in a digital pathology setting. Experiments show that a conventional model generalises poorly to data collected from another site. By annotating a few samples during inference however, a model with initial parameters obtained through MAML training can adapt to achieve better generalisation performance. It is also demonstrated that a simple transfer learning approach using a kNN classifier on features extracted from a conventional model yields good generalisation, but the variance caused by random sampling is higher.

The results indicate that meta learning can lead to a lower annotation effort for machine learning in digital pathology while maintaining accuracy.

In today’s society, we experience an increasing challenge to provide healthcare to everyone in need due to the increasing number of patients and the shortage of medical staff. Computers have contributed to mitigating this challenge by offloading the medical staff from some of the tasks. With the rise of deep learning, countless new possibilities have opened to help the medical staff even further. One domain where deep learning can be applied is analysis of ultrasound images. In this thesis we investigate the problem of classifying standard views of the heart in ultrasound images with the help of deep learning. We conduct mainly three experiments. First, we use NasNet mobile, InceptionV3, VGG16 and MobileNet, pre-trained on ImageNet, and finetune them to ultrasound heart images. We compare the accuracy of these networks to each other and to the baselinemodel, a CNN that was proposed in [23]. Then we assess a neural network’s capability to generalize to images from ultrasound machines that the network is not trained on. Lastly, we test how the performance of the networks degrades with decreasing amount of training data. Our first experiment shows that all networks considered in this study have very similar performance in terms of accuracy with Inception V3 being slightly better than the rest. The best performance is achieved when the whole network is finetuned to our problem instead of finetuning only apart of it, while gradually unlocking more layers for training. The generalization experiment shows that neural networks have the potential to generalize to images from ultrasound machines that they are not trained on. It also shows that having a mix of multiple ultrasound machines in the training data increases generalization performance. In our last experiment we compare the performance of the CNN proposed in [23] with MobileNet pre-trained on ImageNet and MobileNet randomly initialized. This shows that the performance of the baseline model suffers the least with decreasing amount of training data and that pre-training helps the performance drastically on smaller training datasets.

The process of locating moving objects through video sequences is a fundamental computer vision problem. This process is referred to as video tracking and has a broad range of applications. Even though video tracking is an open research topic that have received much attention during recent years, developing accurate and robust algorithms that can handle complicated tracking tasks and scenes is still challenging. One challenge in computer vision is to develop systems that like humans can understand, interpret and recognize visual information in different situations.

In this master thesis work, a tracking algorithm based on eye tracking data is proposed. The aim was to compare the tracking performance of the proposed algorithm with a state-of-the-art video tracker. The algorithm was tested on gaze signals from five participants recorded with an eye tracker while the participants were exposed to dynamic stimuli. The stimuli were moving objects displayed on a stationary computer screen. The proposed algorithm is working offline meaning that all data is collected before analysis.

The results show that the overall performance of the proposed eye tracking algorithm is comparable to the performance of a state-of-the-art video tracker. The main weaknesses are low accuracy for the proposed eye tracking algorithm and handling of occlusion for the video tracker. We also suggest a method for using eye tracking as a complement to object tracking methods. The results show that the eye tracker can be used in some situations to improve the tracking result of the video tracker. The proposed algorithm can be used to help the video tracker to redetect objects that have been occluded or for some other reason are not detected correctly. However, ATOM brings higher accuracy.

Automatic Face Recognition (AFR) can be useful in the forensic field when identifying people in surveillance footage. In AFR systems it is common to use deep neural networks which perform well if the quality of the images keeps a certain level. This is a problem when applying AFR on surveillance data since the quality of those images can be very poor. In this thesis the CNN FaceNet has been used to evaluate how different quality parameters influence the accuracy of the face recognition. The goal is to be able to draw conclusions about how to improve the recognition by using and avoiding certain parameters based on the conditions. Parameters that have been experimented with are angle of the face, image quality, occlusion, colour and lighting. This has been achieved by using datasets with different properties or by alternating the images. The parameters are meant to simulate different situations that can occur in surveillance footage that is difficult for the network to recognise. Three different models have been evaluated with different amount of embeddings and different training data. The results show that the two models trained on the VGGFace2 dataset performs much better than the one trained on CASIA-WebFace. All models performance drops on images with low quality compared to images with high quality because of the training data including mostly high-quality images. In some cases, the recognition results can be improved by applying some alterations in the images. This could be by using one frontal and one profile image when trying to identify a person or occluding parts of the shape of the face if it gets recognized as other persons with similar face shapes. One main improvement would be to extend the training datasets with more low-quality images. To some extent, this could be achieved by different kinds of data augmentation like artificial occlusion and down-sampled images.

This thesis investigates the use of Generative Adversarial Networks (GANs) for detecting images containing non-natural objects in natural environments and if the introduction of stereo data can improve the performance. The state-of-the-art GAN-based anomaly detection method presented by A. Berget al. in [5] (BergGAN) was the base of this thesis. By modifiying BergGAN to not only accept three channel input, but also four and six channel input, it was possible to investigate the effect of introducing stereo data in the method. The input to the four channel network was an RGB image and its corresponding disparity map, and the input to the six channel network was a stereo pair consistingof two RGB images. The three datasets used in the thesis were constructed froma dataset of aerial video sequences provided by SAAB Dynamics, where the scene was mostly wooded areas. The datasets were divided into training and validation data, where the latter was used for the performance evaluation of the respective network. The evaluation method suggested in [5] was used in the thesis, where each sample was scored on the likelihood of it containing anomalies, Receiver Operating Characteristics (ROC) analysis was then applied and the area under the ROC-curve was calculated. The results showed that BergGAN was successfully able to detect images containing non-natural objects in natural environments using the dataset provided by SAAB Dynamics. The adaption of BergGAN to also accept four and six input channels increased the performance of the method, showing that there is information in stereo data that is relevant for GAN-based anomaly detection. There was however no substantial performance difference between the network trained with two RGB images versus the one trained with an RGB image and its corresponding disparity map.

Object detection is a classical computer vision task, encountered in many practical applications such as robotics and autonomous driving. The latter involves serious consequences of failure and a multitude of challenging demands, including high computational efficiency and detection accuracy. Distant objects are notably difficult to detect accurately due to their small scale in the image, consisting of only a few pixels. This is especially problematic in autonomous driving, as objects should be detected at the earliest possible stage to facilitate handling of hazardous situations. Previous work has addressed small objects via use of feature pyramids and super-resolution techniques, but the efficiency of such methods is limited as computational cost increases with image resolution. Therefore, a trade-off must be made between accuracy and cost. Opportunely though, a common characteristic of driving scenarios is the predominance of distant objects in the centre of the image. Thus, the full-frame image can be downsampled to reduce computational cost, and a crop can be extracted from the image centre to preserve resolution for distant vehicles. In this way, short- and long-range images are generated. This thesis investigates the fusion of such images in a convolutional neural network, particularly the fusion level, fusion operation, and spatial alignment. A novel framework — DetSLR — is proposed for the task and examined via the aforementioned aspects. Through adoption of the framework for the well-established SSD detector and MobileNetV2 feature extractor, it is shown that the framework significantly improves upon the original detector without incurring additional cost. The fusion level is shown to have great impact on the performance of the framework, favouring high-level fusion, while only insignificant differences exist between investigated fusion operations. Finally, spatial alignment of features is demonstrated to be a crucial component of the framework.

In digital image correlation, an optical full-field analysis method that can determine displacements of an object under load, high-resolution images are preferable. One way to improve the resolution is to improve the camera hardware. This can be expensive, hence another way to enhance the image is by various image processing techniques increase the resolution of the image. There are several ways of doing this and these techniques are called super-resolution. In this thesisthe theory behind several different approaches to super-resolution is presented and discussed. The goal of this Thesis has been to investigate if super-resolutionis possible in a scene with moving objects as well as movement of the camera. It became clear early on that image registration, a step in many super-resolution methods that will be explained in this thesis, was of utmost importance, and a major part of the work became comparing image registration methods. Data has been recorded and then two different super-resolution algorithms have been evaluated on a data set showing that super-resolution is possible.

Light Detection and Ranging (LiDAR) sensors have many different application areas, from revealing archaeological structures to aiding navigation of vehicles. However, it is challenging to interpret and fully use the vast amount of unstructured data that LiDARs collect. Automatic classification of LiDAR data would ease the utilization, whether it is for examining structures or aiding vehicles.

In recent years, there have been many advances in deep learning for semantic segmentation of automotive LiDAR data, but there is less research on aerial LiDAR data. This thesis investigates the current state-of-the-art deep learning architectures, and how well they perform on LiDAR data acquired by an Unmanned Aerial Vehicle (UAV). It also investigates different training techniques for class imbalanced and limited datasets, which are common challenges for semantic segmentation networks. Lastly, this thesis investigates if pre-training can improve the performance of the models.

The LiDAR scans were first projected to range images and then a fully convolutional semantic segmentation network was used. Three different training techniques were evaluated: weighted sampling, data augmentation, and grouping of classes. No improvement was observed by the weighted sampling, neither did grouping of classes have a substantial effect on the performance. Pre-training on the large public dataset SemanticKITTI resulted in a small performance improvement, but the data augmentation seemed to have the largest positive impact. The mIoU of the best model, which was trained with data augmentation, was 63.7% and it performed very well on the classes Ground, Vegetation, and Vehicle. The other classes in the UAV dataset, Person and Structure, had very little data and were challenging for most models to classify correctly. In general, the models trained on UAV data performed similarly as the state-of-the-art models trained on automotive data.

Multi-pedestrian tracking (MPT) is the task of localizing and following the trajectory of pedestrians in a sequence. Using an MPT algorithm is an important part in preventing pedestrian-vehicle collisions in Automated Driving (AD) and Advanced Driving Assistance Systems (ADAS). It has benefited greatly from the advances in computer vision and machine learning in the last decades. Using a pedestrian detector, the tracking consists of associating the detections between frames and maintaining pedestrian identities throughout the sequence. This can be a challenging task due to occlusions, missed detections and complex scenes. The number of pedestrians is unknown, and it varies with time. Finding new methods for improving MPT is an active research field and there are many approaches found in the literature. This work focuses on improving the detection-to-track association, the data association, with the help of extracted color features for each pedestrian. Utilizing the recent improvements in object detection this work shows that classical color features still is relevant in pedestrian tracking for real time applications with limited computational resources. The appearance is not only used in the data association but also integrated in a new proposed method to avoid tracking errors due to missed detections. The results show that even with simple models the color appearance can be used to improve the tracking results. Evaluation on the commonly used Multi-Object Tracking-benchmark shows an improvement in the Multi-Object Tracking Accuracy and identity switches, while keeping other measures essentially unchanged.

CNN-based (Convolutional Neural Network) visual object detectors often reach human level of accuracy but need to be trained with large amounts of manually annotated data. Collecting and annotating this data can frequently be time-consuming and financially expensive. Using generative models to augment the data can help minimize the amount of data required and increase detection per-formance. Many state-of-the-art generative models are Generative Adversarial Networks (GANs). This thesis investigates if and how one can utilize image data to generate new data through GANs to train a YOLO-based (You Only Look Once) object detector, and how CAD (Computer-Aided Design) models can aid in this process.

In the experiments, different models of GANs are trained and evaluated by visual inspection or with the Fréchet Inception Distance (FID) metric. The data provided by Ericsson Research consists of images of antenna and baseband equipment along with annotations and segmentations. Ericsson Research supplied the YOLO detector, and no modifications are made to this detector. Finally, the YOLO detector is trained on data generated by the chosen model and evaluated by the Average Precision (AP).

The results show that the generative models designed in this work can produce RGB images of high quality. However, the quality reduces if binary segmentation masks are to be generated as well. The experiments with CAD input data did not result in images that could be used for the training of the detector.

The GAN designed in this work is able to successfully replace objects in images with the style of other objects. The results show that training the YOLO detector with GAN-modified data compared to training with real data leads to the same detection performance. The results also show that the shapes and backgrounds of the antennas contributed more to detection performance than their style and colour.

In one of the facilities at the Stena Recycling plant in Halmstad, Sweden, about 300 tonnes of metallic waste is processed each day with the aim of sorting out all non-ferrous material. At the end of this process, non-ferrous materials are

manually sorted out from the ferrous materials. This thesis investigates a computer vision based approach to identify and localize the non-ferrous materials

and eventually automate the sorting.Images were captured of ferrous and non-ferrous materials. The images areprocessed and segmented to be used as annotation data for a deep convolutionalneural segmentation network. Network models have been trained on different

kinds and amounts of data. The resulting models are evaluated and tested in ac-cordance with different evaluation metrics. Methods of creating advanced train-ing data by merging imaging information were tested. Experiments with using

classifier prediction confidence to identify objects of unknown classes were per-formed.

This thesis shows that it is possible to discern ferrous from non-ferrous mate-rial with a purely vision based system. The thesis also shows that it is possible to

automatically create annotated training data. It becomes evident that it is possi-ble to create better training data, tailored for the task at hand, by merging image

data. A segmentation network trained on more than two classes yields lowerprediction confidence for objects unknown to the classifier.Substituting manual sorting with a purely vision based system seems like aviable approach. Before a substitution is considered, the automatic system needsto be evaluated in comparison to the manual sorting.

This thesis presents and evaluates different methods to semantically segment 3D-models by rendered 2D-views. The 2D-views are segmented separately and then merged together. The thesis evaluates three different merge strategies, two different classification architectures, how many views should be rendered and how these rendered views should be arranged. The results are evaluated both quantitatively and qualitatively and then compared with the current classifier at Vricon presented in [30].

The conclusion of this thesis is that there is a performance gain to be had using this method. The best model was using two views and attains an accuracy of 90.89% which can be compared with 84.52% achieved by the single view network from [30]. The best nine view system achieved a 87.72%. The difference in accuracy between the two and the nine view system is attributed to the higher quality mesh on the sunny side of objects, which typically is the south side.

The thesis provides a proof of concept and there are still many areas where the system can be improved. One of them being the extraction of training data which seemingly would have a huge impact on the performance.

Visual object detection is a popular computer vision task that has been intensively investigated using deep learning on real data. However, data from virtual environments have not received the same attention. A virtual environment enables generating data for locations that are not easily reachable for data collection, e.g. aerial environments. In this thesis, we study the problem of object detection in virtual environments, more specifically an aerial virtual environment. We use a simulator, to generate a synthetic data set of 16 different types of vehicles captured from an airplane.

To study the performance of existing methods in virtual environments, we train and evaluate two state-of-the-art detectors on the generated data set. Experiments show that both detectors, You Only Look Once version 3 (YOLOv3) and Single Shot MultiBox Detector (SSD), reach similar performance quality as previously presented in the literature on real data sets.

In addition, we investigate different fusion techniques between detectors which were trained on two different subsets of the dataset, in this case a subset which has cars with fixed colors and a dataset which has cars with varying colors. Experiments show that it is possible to train multiple instances of the detector on different subsets of the data set, and combine these detectors in order to boost the performance.

Traffic sign recognition is an important problem for autonomous cars and driver assistance systems. With recent developments in the field of machine learning, high performance can be achieved, but typically at a large computational cost.

This thesis aims to investigate the relation between classification accuracy and computational complexity for the visual recognition problem of classifying traffic signs. In particular, the benefits of partitioning the classification problem into smaller sub-problems using prior knowledge in the form of shape or current region are investigated.

In the experiments, the convolutional neural network (CNN) architecture MobileNetV2 is used, as it is specifically designed to be computationally efficient. To incorporate prior knowledge, separate CNNs are used for the different subsets generated when partitioning the dataset based on region or shape. The separate CNNs are trained from scratch or initialized by pre-training on the full dataset.

The results support the intuitive idea that performance initially increases with network size and indicate a network size where the improvement stops. Including shape information using the two investigated methods does not result in a significant improvement. Including region information using pretrained separate classifiers results in a small improvement for small complexities, for one of the regions in the experiments.

In the end, none of the investigated methods of including prior knowledge are considered to yield an improvement large enough to justify the added implementational complexity. However, some other methods are suggested, which would be interesting to study in future work.

The interest for autonomous driving assistance, and in the end, self-driving cars, has increased vastly over the last decade. Automotive safety continues to be a priority for manufacturers, politicians and people alike. Visual-based systems aiding the drivers have lately been boosted by advances in computer vision and machine learning. In this thesis, we evaluate the concept of an end-to-end machine learning solution for detecting and classifying road lane markings, and compare it to a more classical semantic segmentation solution. The analysis is based on the frame-by-frame scenario, and shows that our proposed end-to-end system has clear advantages when it comes detecting the existence of lanes and producing a consistent, lane-like output, especially in adverse conditions such as weak lane markings. Our proposed method allows the system to predict its own confidence, thereby allowing the system to suppress its own output when it is not deemed safe enough. The thesis finishes with proposed future work needed to achieve optimal performance and create a system ready for deployment in an active safety product.

Semantic segmentation is a key approach to comprehensive image data analysis. It can be applied to analyze 2D images, videos, and even point clouds that contain 3D data points. On the first two problems, CNNs have achieved remarkable progress, but on point cloud segmentation, the results are less satisfactory due to challenges such as limited memory resource and difficulties in 3D point annotation. One of the research studies carried out by the Computer Vision Lab at Linköping University was aiming to ease the semantic segmentation of 3D point cloud. The idea is that by first projecting 3D data points to 2D space and then focusing only on the analysis of 2D images, we can reduce the overall workload for the segmentation process as well as exploit the existing well-developed 2D semantic segmentation techniques. In order to improve the performance of CNNs for 2D semantic segmentation, the study has used input data derived from different modalities. However, how different modalities can be optimally fused is still an open question. Based on the above-mentioned study, this thesis aims to improve the multistream framework architecture. More concretely, we investigate how different singlestream architectures impact the multistream framework with a given fusion method, and how different fusion methods contribute to the overall performance of a given multistream framework. As a result, our proposed fusion architecture outperformed all the investigated traditional fusion methods. Along with the best singlestream candidate and few additional training techniques, our final proposed multistream framework obtained a relative gain of 7.3\% mIoU compared to the baseline on the semantic3D point cloud test set, increasing the ranking from 12th to 5th position on the benchmark leaderboard.

One fundamental task in robotics is random bin-picking, where it is important to be able to detect an object in a bin and estimate its pose to plan the motion of a robotic arm. For this purpose, this thesis work aimed to investigate and evaluate algorithms for 6D pose estimation when the object was given by a CAD model. The scene was given by a point cloud illustrating a partial 3D view of the bin with multiple instances of the object. Two algorithms were thus implemented and evaluated. The first algorithm was an approach based on Point Pair Features, and the second was Fast Global Registration. For evaluation, four different CAD models were used to create synthetic data with ground truth annotations.

It was concluded that the Point Pair Feature approach provided a robust localization of objects and can be used for bin-picking. The algorithm appears to be able to handle different types of objects, however, with small limitations when the object has flat surfaces and weak texture or many similar details. The disadvantage with the algorithm was the execution time. Fast Global Registration, on the other hand, did not provide a robust localization of objects and is thus not a good solution for bin-picking.

Finding disparity maps between stereo images is a well studied topic within computer vision. While both classical and machine learning approaches exist in the literature, they frequently struggle to correctly solve the disparity in regions with low texture, sharp edges or occlusions. Finding approximate solutions to these problem areas is frequently referred to as disparity refinement, and is usually carried out separately after an initial disparity map has been generated.

In the recent literature, the use of Normalized Convolution in Convolutional Neural Networks have shown remarkable results when applied to the task of stereo depth completion. This thesis investigates how well this approach performs in the case of disparity refinement. Specifically, we investigate how well such a method can improve the initial disparity maps generated by the stereo matching algorithm developed at Saab Dynamics using a rectified stereo rig.

To this end, a dataset of ground truth disparity maps was created using equipment at Saab, namely a setup for structured light and the stereo rig cameras. Because the end goal is a dataset fit for training networks, we investigate an approach that allows for efficient creation of significant quantities of dense ground truth disparities.

The method for generating ground truth disparities generates several disparity maps for every scene measured by using several stereo pairs. A densified disparity map is generated by merging the disparity maps from the neighbouring stereo pairs. This resulted in a dataset of 26 scenes and 104 dense and accurate disparity maps.

Our evaluation results show that the chosen Normalized Convolution Network based method can be adapted for disparity map refinement, but is dependent on the quality of the input disparity map.

Given satellite images with accompanying pixel classifications and elevation data, we propose different solutions to object detection. The first method uses hierarchical clustering for segmentation and then employs different methods of classification. One of these classification methods used domain knowledge to classify objects while the other used Support Vector Machines. Additionally, a combination of three Support Vector Machines were used in a hierarchical structure which out-performed the regular Support Vector Machine method in most of the evaluation metrics. The second approach is more conventional with different types of Convolutional Neural Networks. A segmentation network was used as well as a few detection networks and different fusions between these. The Convolutional Neural Network approach proved to be the better of the two in terms of precision and recall but the clustering approach was not far behind. This work was done using a relatively small amount of data which potentially could have impacted the results of the Machine Learning models in a negative way.

The organization International Aid Services (IAS) provides people in East Africawith clean water through well drilling. The wells are located in surroundingsfar away for the investors to inspect and therefore IAS wishes to be able to monitortheir wells to get a better overview if different types of improvements needto be made. To see the load on different water sources at different times of theday and during the year, and to know how many people that are visiting thewells, is of particular interest. In this paper, a method is proposed for countingpeople around the wells. The goal is to choose a suitable method for detectinghumans in images and evaluate how it performs. The area of counting humansin images is not a new topic, though it needs to be taken into account that thesituation implies some restrictions. A Raspberry Pi with an associated camerais used, which is a small embedded system that cannot handle large and complexsoftware. There is also a limited amount of data in the project. The methodproposed in this project uses a pre-trained convolutional neural network basedobject detector called the Single Shot Detector, which is adapted to suit smallerdevices and applications. The pre-trained network that it is based on is calledMobileNet, a network that is developed to be used on smaller systems. To see howgood the chosen detector performs it will be compared with some other models.Among them a detector based on the Inception network, a significantly larger networkthan the MobileNet. The base network is modified by transfer learning.Results shows that a fine-tuned and modified network can achieve better result,from a F1-score of 0.49 for a non-fine-tuned model to 0.66 for the fine-tuned one.

Deep learning has been intensively researched in computer vision tasks like im-age classification. Collecting and labeling images that these neural networks aretrained on is labor-intensive, which is why alternative methods of collecting im-ages are of interest. Virtual environments allow rendering images and automaticlabeling,  which could speed up the process of generating training data and re-duce costs.This  thesis  studies  the  problem  of  transfer  learning  in  image  classificationwhen the classifier has been trained on rendered images using a game engine andtested on real images. The goal is to render images using a game engine to createa classifier that can separate images depicting people wearing civilian clothingor camouflage.  The thesis also studies how domain adaptation techniques usinggenerative  adversarial  networks  could  be  used  to  improve  the  performance  ofthe classifier.  Experiments show that it is possible to generate images that canbe used for training a classifier capable of separating the two classes.  However,the experiments with domain adaptation were unsuccessful.  It is instead recom-mended to improve the quality of the rendered images in terms of features usedin the target domain to achieve better results.

Visual object tracking is one of the fundamental problems in computer vision, with a wide number of practical applications in e.g.\ robotics, surveillance etc. Given a video sequence and the target bounding box in the first frame, a tracker is required to find the target in all subsequent frames. It is a challenging problem due to the limited training data available. An object tracker is generally evaluated using two criterias, namely robustness and accuracy. Robustness refers to the ability of a tracker to track for long durations, without losing the target. Accuracy, on the other hand, denotes how accurately a tracker can estimate the target bounding box.

Recent years have seen significant improvement in tracking robustness. However, the problem of accurate tracking has seen less attention. Most current state-of-the-art trackers resort to a naive multi-scale search strategy which has fundamental limitations. Thus, in this thesis, we aim to develop a general target estimation component which can be used to determine accurate bounding box for tracking. We will investigate how bounding box estimators used in object detection can be modified to be used for object tracking. The key difference between detection and tracking is that in object detection, the classes to which the objects belong are known. However, in tracking, no prior information is available about the tracked object, other than a single image provided in the first frame. We will thus investigate different architectures to utilize the first frame information to provide target specific bounding box predictions. We will also investigate how the bounding box predictors can be integrated into a state-of-the-art tracking method to obtain robust as well as accurate tracking.

Photos captured in the shortwave infrared (SWIR) spectrum are interesting in military applications because they are independent of what time of day the pic- ture is captured because the sun, moon, stars and night glow illuminate the earth with short-wave infrared radiation constantly. A major problem with today’s SWIR cameras is that they are very expensive to produce and hence not broadly available either within the military or to civilians. Using a relatively new tech- nology called compressive sensing (CS), enables a new type of camera with only a single pixel sensor in the sensor (a SPC). This new type of camera only needs a fraction of measurements relative to the number of pixels to be reconstructed and reduces the cost of a short-wave infrared camera with a factor of 20. The camera uses a micromirror array (DMD) to select which mirrors (pixels) to be measured in the scene, thus creating an underdetermined linear equation system that can be solved using the techniques described in CS to reconstruct the im- age. Given the new technology, it is in the Swedish Defence Research Agency (FOI) interest to evaluate the potential of a single pixel camera. With a SPC ar- chitecture developed by FOI, the goal of this thesis was to develop methods for sampling, reconstructing images and evaluating their quality. This thesis shows that structured random matrices and fast transforms have to be used to enable high resolution images and speed up the process of reconstructing images signifi- cantly. The evaluation of the images could be done with standard measurements associated with camera evaluation and showed that the camera can reproduce high resolution images with relative high image quality in daylight.

Deep learning has been rapidly growing in recent years obtaining excellent results for many computer vision applications, such as image classification and object detection. One aspect for the increased popularity of deep learning is that it mitigates the need for hand-crafted features. This thesis work investigates deep learning as a methodology to solve the problem of autonomous collision avoidance for a small robotic car. To accomplish this, transfer learning is used with the VGG16 deep network pre-trained on ImageNet dataset. A dataset has been collected and then used to fine-tune and validate the network offline. The deep network has been used with the robotic car in a real-time manner. The robotic car sends images to an external computer, which is used for running the network. The predictions from the network is sent back to the robotic car which takes actions based on those predictions. The results show that deep learning has great potential in solving the collision avoidance problem.

3D reconstruction is the process of constructing a three-dimensional model from images. It contains multiple steps where each step can induce errors. When doing 3D reconstruction of outdoor scenes, there are some types of scene content that regularly cause problems and affect the resulting 3D model. Two of these are water, due to its fluctuating nature, and sky because of it containing no useful (3D) data. These areas cause different problems throughout the process and do generally not benefit it in any way. Therefore, masking them early in the reconstruction chain could be a useful step in an outdoor scene reconstruction pipeline. Manual masking of images is a time-consuming and boring task and it gets very tedious for big data sets which are often used in large scale 3D reconstructions. This master thesis explores if this can be done automatically using Convolutional Neural Networks for semantic segmentation, and to what degree the masking would benefit a 3D reconstruction pipeline.

The cost and environmental damage of reclaims is a large problem within thepaper industry. With certain types of paper, so called crepe marks on the paper’ssurface is a common issue, leading to printing defects and consequentlyreclaims. This thesis compares four different image analysis methods for evaluatingcrepe marks and predicting printing results. The methods evaluated consistsof one established methods, two adaptations of established methods andone novel method. All methods were evaluated on the same data, topographicheight images of paper samples from 4 paper rolls of similar type but differingin roughness. The method based on 1D Fourier analysis and the method basedon fully convolutional networks performs best, depending on if speed or detailedcharacteristics is a priority.

Volume measurements of timber loads is done in conjunction with timber trade. When dealing with goods of major economic values such as these, it is important to achieve an impartial and fair assessment when determining price-based volumes.

With the help of Saab’s missile targeting technology, CIND AB develops products for digital volume measurement of timber loads. Currently there is a system in operation that automatically reconstructs timber trucks in motion to create measurable images of them. Future iterations of the system is expected to fully automate the scaling by generating a volumetric representation of the timber and calculate its external gross volume. The first challenge towards this development is to separate the timber load from the truck.

This thesis aims to evaluate and implement appropriate method for semantic pixel-wise segmentation of timber loads in real time. Image segmentation is a classic but difficult problem in computer vision. To achieve greater robustness, it is therefore important to carefully study and make use of the conditions given by the existing system. Variations in timber type, truck type and packing together create unique combinations that the system must be able to handle. The system must work around the clock in different weather conditions while maintaining high precision and performance.

Robotic bin picking is the problem of emptying a bin of randomly distributedobjects through a robotic interface. This thesis examines an SVM approach to ex-tract grasping points for a vacuum-type gripper. The SVM is trained on syntheticdata and used to classify the points of a non-synthetic 3D-scanned point cloud aseither graspable or non-graspable. The classified points are then clustered intograspable regions from which the grasping points are extracted.

The SVM models and the algorithm as a whole are trained and evaluated againstcubic and cylindrical objects. Separate SVM models are trained for each type ofobject in addition to one model being trained on a dataset containing both typesof objects. It is shown that the performance of the SVM in terms accuracy isdependent on the objects and their geometrical properties. Further, it is shownthat the algorithm is reasonably robust in terms of successfully picking objects,regardless of the scale of the objects.

Recently, sensors such as radars and cameras have been widely used in automotives, especially in Advanced Driver-Assistance Systems (ADAS), to collect information about the vehicle's surroundings. Stereo cameras are very popular as they could be used passively to construct a 3D representation of the scene in front of the car. This allowed the development of several ADAS algorithms that need 3D information to perform their tasks. One interesting application is Road Surface Preview (RSP) where the task is to estimate the road height along the future path of the vehicle. An active suspension control unit can then use this information to regulate the suspension, improving driving comfort, extending the durabilitiy of the vehicle and warning the driver about potential risks on the road surface. Stereo cameras have been successfully used in RSP and have demonstrated very good performance. However, the main disadvantages of stereo cameras are their high production cost and high power consumption. This limits installing several ADAS features in economy-class vehicles. A less expensive alternative are monocular cameras which have a significantly lower cost and power consumption. Therefore, this thesis investigates the possibility of solving the Road Surface Preview task using a monocular camera. We try two different approaches: structure-from-motion and Convolutional Neural Networks.The proposed methods are evaluated against the stereo-based system. Experiments show that both structure-from-motion and CNNs have a good potential for solving the problem, but they are not yet reliable enough to be a complete solution to the RSP task and be used in an active suspension control unit.

This master thesis explores the possibility of using generative Adversarial Networks (GANs) to refine labeled synthetic code images to resemble real code images while preserving label information. The GAN used in this thesis consists of a refiner and a discriminator. The discriminator tries to distinguish between real images and refined synthetic images. The refiner tries to fool the discriminator by producing refined synthetic images such that the discriminator classify them as real. By updating these two networks iteratively, the idea is that they will push each other to get better, resulting in refined synthetic images with real image characteristics.

The aspiration, if the exploration of GANs turns out successful, is to be able to use refined synthetic images as training data in Semantic Segmentation (SS) tasks and thereby eliminate the laborious task of gathering and labeling real data. Starting off from a foundational GAN-model, different network architectures, hyperparameters and other design choices are explored to find the best performing GAN-model.

As is widely acknowledged in the relevant literature, GANs can be difficult to train and the results in this thesis are varying and sometimes ambiguous. Based on the results from this study, the best performing models do however perform better in SS tasks than the unrefined synthetic set they are based on and benchmarked against, with regards to Intersection over Union.

Training data is the bottleneck for training Convolutional Neural Networks. A larger dataset gives better accuracy though also needs longer training time. It is shown by finetuning neural networks on synthetic rendered images, that the mean average precision increases. This method was applied to two different datasets with five distinctive objects in each. The first dataset consisted of random objects with different geometric shapes. The second dataset contained objects used to assemble IKEA furniture. The neural network with the best performance, trained on 5400 images, achieved a mean average precision of 0.81 on a test which was a sample of a video sequence. Analysis of the impact of the factors dataset size, batch size, and numbers of epochs used in training and different network architectures were done. Using synthetic images to train CNN’s is a promising path to take for object detection where access to large amount of annotated image data is hard to come by.

In many situations after a big catastrophe such as the one in Fukushima, the disaster area is highly dangerous for humans to enter. It is in such environments that a semi-autonomous robot could limit the risks to humans by exploring and mapping the area on its own. This thesis intends to design and implement a software based SLAM system which has potential to run in real-time using a Kinect 2 sensor as input.

The focus of the thesis has been to create a system which allows for efficient storage and representation of the map, in order to be able to explore large environments. This is done by separating the map in different abstraction levels corresponding to local maps connected by a global map.

During the implementation, this structure has been kept in mind in order to allow modularity. This makes it possible for each sub-component in the system to be exchanged if needed.

The thesis is broad in the sense that it uses techniques from distinct areas to solve the sub-problems that exist. Some examples being, object detection and classification, point-cloud registration and efficient 3D-based occupancy trees.

This thesis investigates if support vector machine classification is a suitable approach when performing automatic segmentation of knee cartilage using quantitative magnetic resonance imaging data. The data sets used are part of a clinical project that investigates if patients that have suffered recent knee damage will develop cartilage damage. Therefore the thesis also investigates if the segmentation results can be used to predict the clinical outcome of the patients.

Two methods that perform the segmentation using support vector machine classification are implemented and evaluated. The evaluation indicates that it is a good approach for the task, but the implemented methods needs to be further improved and tested on more data sets before clinical use.

It was not possible to relate the cartilage properties to clinical outcome using the segmentation results. However, the investigation demonstrated good promise of how the segmentation results, if they are improved, can be used in combination with quantitative magnetic resonance imaging data to analyze how the cartilage properties change over time or vary between knees.

There is a need for quantitative CT data in radiation therapy. Currently there are only few algorithms that address this issue, for instance the commercial DirectDensity algorithm. In scientific literature, an example of such an algorithm is DIRA. DIRA is an iterative model-based reconstruction method for dual-energy CT whose goal is to determine the material composition of the patient from accurate linear attenuation coefficients. It has been implemented in a two dimensional geometry, i.e., it could process axial scans only.  There was a need to extend DIRA so that it could process projection data generated in helical scanning geometries. The newly developed algorithm (DIRA-3D) implemented (i) polyenergetic semi-parallel projection generation, (ii) mono-energetic parallel projection generation and (iii) the PI-method for image reconstruction. The computation experiments showed that the accuracies of the resulting LAC and mass fractions were comparable to the ones of the original DIRA. The results converged after 10 iterations.

Modern cars are often equipped with sensors like radar, infrared cameras and stereo cameras that collect information about its surroundings. By using a stereo camera, it is possible to receive information about the distance to points in front of the car. This information can be used to estimate the height of the predicted path of the car. An application which does this is the stereo based Road surface preview (RSP) algorithm. By using the output from the RSP algorithm it is possible to use active suspension control, which controls the vertical movement of the wheels relative to the chassis. This application primarily makes the driving experience more comfortable, but also extends the durability of the vehicle. The idea behind this Master’s thesis is to create an evaluation tool for the RSP algorithm, which can be used at arbitrary roads. 

The thesis describes the proposed evaluation tool, where focus has been to make an accurate comparison of camera data received from the RSP algorithm and laser data used as ground truth in this thesis. Since the tool shall be used at the company proposing this thesis, focus has also been on making the tool user friendly. The report discusses the proposed methods, possible sources to errors and improvements. The evaluation tool considered in this thesis shows good results for the available test data, which made it possible to include an investigation of a possible improvement of the RSP algorithm.

All dairy cows in Europe wear unique identification tags in their ears. These eartags are standardized and contains the cows identification numbers, today only used for visual identification by the farmer. The cow also needs to be identified by an automatic identification system connected to milk machines and other robotics used at the farm. Currently this is solved with a non-standardized radio transmitter which can be placed on different places on the cow and different receivers needs to be used on different farms. Other drawbacks with the currently used identification system are that it is expensive and unreliable. This thesis explores the possibility to replace this non standardized radio frequency based identification system with a standardized computer vision based system. The method proposed in this thesis uses a color threshold approach for detection, a flood fill approach followed by Hough transform and a projection method for segmentation and evaluates template matching, k-nearest neighbour and support vector machines as optical character recognition methods. The result from the thesis shows that the quality of the data used as input to the system is vital. By using good data, k-nearest neighbour, which showed the best results of the three OCR approaches, handles 98 % of the digits.

Extracting foreground objects from an image is a hot research topic. Doing thisfor high quality real world images in real-time on limited hardware such as asmart phone, is a demanding task. This master thesis shows how this problemcan be addressed using Otsu’s method together with Gaussian probability dis-tributions to create classifiers in different colour channels. We also show howclassifiers can be combined resulting in higher accuracy than using only the indi-vidual classifiers. We also propose using inter-class variance together with imagevariance to estimate classifier quality.A data set was produced to evaluate performance. The data set featuresreal-world images captured by a smart phone and objects of varying complex-ity against plain backgrounds that can be found in a typical office or urban space.

The ability to automatically estimate the volume of timber is becoming increasingly important within the timber industry. The large number of timber trucks arriving each day at Swedish timber terminals fortifies the need for a volume estimation performed in real-time and on-the-go as the trucks arrive.

This thesis investigates if a volumetric integration of disparity maps acquired from a Multi-View Stereo (MVS) system is a suitable approach for automatic volume estimation of timber loads. As real-time execution is preferred, efforts were made to provide a scalable method. The proposed method was quantitatively evaluated on datasets containing two geometric objects of known volume. A qualitative comparison to manual volume estimates of timber loads was also made on datasets recorded at a Swedish timber terminal.

The proposed method is shown to be both accurate and precise under specific circumstances. However, robustness is poor to varying weather conditions, although a more thorough evaluation of this aspect needs to be performed. The method is also parallelizable, which means that future efforts can be made to significantly decrease execution time.

The Exponential Linear Unit (ELU) has been proven to speed up learning and improve the classification performance over activation functions such as ReLU and Leaky ReLU for convolutional neural networks. The reasons behind the improved behavior are that ELU reduces the bias shift, it saturates for large negative inputs and it is continuously differentiable. However, it remains open whether ELU has the optimal shape and we address the quest for a superior activation function.

We use a new formulation to tune a piecewise linear activation function during training, to investigate the above question, and learn the shape of the locally optimal activation function. With this tuned activation function, the classification performance is improved and the resulting, learned activation function shows to be ELU-shaped irrespective if it is initialized as a RELU, LReLU or ELU. Interestingly, the learned activation function does not exactly pass through the origin indicating that a shifted ELU-shaped activation function is preferable. This observation leads us to introduce the Shifted Exponential Linear Unit (ShELU) as a new activation function.

Experiments on Cifar-100 show that the classification performance is further improved when using the ShELU activation function in comparison with ELU. The improvement is achieved when learning an individual bias shift for each neuron.

Barcodes are ubiquituous in modern society and they have had industrial application for decades. However, for noisy images modern methods can underperform. Poor lighting conditions, occlusions and low resolution can be problematic in decoding. This thesis aims to solve this problem by using neural networks, which have enjoyed great success in many computer vision competitions the last years. We investigate how three different networks perform on data sets with noisy images. The first network is a single classifier, the second network is an ensemble classifier and the third is based on a pre-trained feature extractor. For comparison, we also test two baseline methods that are used in industry today. We generate training data using software and modify it to ensure proper generalization. Testing data is created by photographing barcodes in different settings, creating six image classes - normal, dark, white, rotated, occluded and wrinkled. The proposed single classifier and ensemble classifier outperform the baseline as well as the pre-trained feature extractor by a large margin. The thesis work was performed at SICK IVP, a machine vision company in Linköping in 2017.

The purpose of this document is to reect on novel and upcoming methods for computer vision that might have relevance for application in robot vision and video analytics. The document covers many dierent sub-elds of computer vision, most of which have been addressed by our research activity at the computer vision laboratory. The report has been written based on a request of, and supported by, FOI.

The objective of this master’s thesis work is to evaluate the potential benefit of a superpixel preprocessing step for general object detection in a traffic environment. The various effects of different superpixel parameters on object detection performance, as well as the benefit of including depth information when generating the superpixels are investigated.

In this work, three superpixel algorithms are implemented and compared, including a proposal for an improved version of the popular Spectral Linear Iterative Clustering superpixel algorithm (SLIC). The proposed improved algorithm utilises a coarse-to-fine approach which outperforms the original SLIC for high-resolution images. An object detection algorithm is also implemented and evaluated. The algorithm makes use of depth information obtained by a stereo camera to extract superpixels corresponding to foreground objects in the image. Hierarchical clustering is then applied, with the segments formed by the clustered superpixels indicating potential objects in the input image.

The object detection algorithm managed to detect on average 58% of the objects present in the chosen dataset. It performed especially well for detecting pedestrians or other objects close to the car. Altering the density distribution of the superpixels in the image yielded an increase in detection rate, and could be achieved both with or without utilising depth information. It was also shown that the use of superpixels greatly reduces the amount of computations needed for the algorithm, indicating that a real-time implementation is feasible.

This work investigates the landscape of aerial image stereo matching (AISM) methods suitable for large scale forest variable estimation. AISM methods are an important source of remotely collected information used in modern forestry to keep track of a growing forest's condition.

A total of 17 AISM methods are investigated, out of which 4 are evaluated by processing a test data set consisting of three aerial images. The test area is located in southern Sweden, consisting of mainly Norway Spruce and Scots Pine. From the resulting point clouds and height raster images, a total of 30 different metrics of both height and density types are derived. Linear regression is used to fit functions from metrics derived from AISM data to a set of forest variables including tree height (HBW), tree diameter (DBW), basal area, volume. As ground truth, data collected by dense airborne laser scanning is used. Results are presented as RMSE and standard deviation concluded from the linear regression.

For tree height, tree diameter, basal area, volume the RMSE ranged from 7.442% to 10.11%, 11.58% to 13.96%, 32.01% to 35.10% and 34.01% to 38.26% respectively. The results concluded that all four tested methods achieved comparable estimation quality although showing small differences among them. Keystone and SURE performed somewhat better while MicMac placed third and Photoscan achieved the less accurate result.

When forensic examiners try to identify the perpetrator of a felony, they use individual facial marks when comparing the suspect with the perpetrator. Facial marks are often used for identification and they are nowadays found manually. To speed up this process, it is desired to detect interesting facial marks automatically. This master thesis describes a method to automatically detect and separate permanent and non-permanent marks. It uses a fast radial symmetry algorithm as a core element in the mark detector. After candidate skin mark extraction, the false detections are removed depending on their size, shape and number of hair pixels. The classification of the skin marks is done with a support vector machine and the different features are examined. The results show that the facial mark detector has a good recall while the precision is poor. The elimination methods of false detection were analysed as well as the different features for the classifier. One can conclude that the color of facial marks is more relevant than the structure when classifying them into permanent and non-permanent marks.

Measurements performed from stereo reconstruction can be obtained with a high accuracy with correct calibrated cameras. A stereo camera rig mounted in an outdoor environment is exposed to temperature changes, which has an impact of the calibration of the cameras.

The aim of the master thesis was to investigate the thermal impact of a calibrated stereo camera rig. This was performed by placing a stereo rig in a temperature chamber and collect data of a calibration board at different temperatures. Data was collected with two different cameras and lensesand used for calibration of the stereo camera rig for different scenarios. The obtained parameters were plotted and analyzed.

The result from the master thesis gives that the thermal variation has an impact of the accuracy of the calibrated stereo camera rig. A calibration obtained in one temperature can not be used for a different temperature without a degradation of the accuracy. The plotted parameters from the calibration had a high noise level due to problems with the calibration methods, and no visible trend from temperature changes could be seen.

Face recognition is the problem of identifying individuals in images. This thesis evaluates two methods used to determine if pairs of face images belong to the same individual or not. The first method is a combination of principal component analysis and a neural network and the second method is based on state-of-the-art convolutional neural networks. They are trained and evaluated using two different data sets. The first set contains many images with large variations in, for example, illumination and facial expression. The second consists of fewer images with small variations.

Principal component analysis allowed the use of smaller networks. The largest network has 1.7 million parameters compared to the 7 million used in the convolutional network. The use of smaller networks lowered the training time and evaluation time significantly. Principal component analysis proved to be well suited for the data set with small variations outperforming the convolutional network which need larger data sets to avoid overfitting. The reduction in data dimensionality, however, led to difficulties classifying the data set with large variations. The generous amount of images in this set allowed the convolutional method to reach higher accuracies than the principal component method.

Lung cancer is the most common type of cancer in the world and always manifests as lung nodules. Nodules are small tumors that consist of lung tissue. They are usually spherical in shape and their cores can be either solid or subsolid. Nodules are common in lungs, but not all of them are malignant. To determine if a nodule is malignant or benign, attributes like nodule size and volume growth are commonly used. The procedure to obtain these attributes is time consuming, and therefore calls for tools to simplify the process.

The purpose of this thesis work was to investigate  the feasibility of a semi-automatic lungnodule segmentation pipeline including volume estimation. This was done by implementing, tuning and evaluating image processing algorithms with different characteristics to create pipeline candidates. These candidates were compared using a similarity index between their segmentation results and ground truth markings to determine the most promising one.

The best performing pipeline consisted of a fixed region of interest together with a level set segmentation algorithm. Its segmentation accuracy was not consistent for all nodules evaluated, but the pipeline showed great potential when dynamically adapting its parameters for each nodule. The use of dynamic parameters was only brie y explored, and further research would be necessary to determine its feasibility.

Detection and positioning of anatomical landmarks, also called points of interest(POI), is often a concept of interest in medical image processing. Different measures or automatic image analyzes are often directly based upon positions of such points, e.g. in organ segmentation or tissue quantification. Manual positioning of these landmarks is a time consuming and resource demanding process. In this thesis, a general method for positioning of anatomical landmarks is outlined, implemented and evaluated. The evaluation of the method is limited to three different POI; left femur head, right femur head and vertebra T9. These POI are used to define the range of the abdomen in order to measure the amount of abdominal fat in 3D data acquired with quantitative magnetic resonance imaging (MRI). By getting more detailed information about the abdominal body fat composition, medical diagnoses can be issued with higher confidence. Examples of applications could be identifying patients with high risk of developing metabolic or catabolic disease and characterizing the effects of different interventions, i.e. training, bariatric surgery and medications. The proposed method is shown to be highly robust and accurate for positioning of left and right femur head. Due to insufficient performance regarding T9 detection, a modified method is proposed for T9 positioning. The modified method shows promises of accurate and repeatable results but has to be evaluated more extensively in order to draw further conclusions.

Cardiovascular diseases are among the most common causes of death worldwide. One of the recently developed flow analysis technique called 4D flow magnetic resonance imaging (MRI) allows an early detection of such diseases. Due to the limited resolution and contrast between blood pool and myocardium of 4D flow images, cine MR images are often used for cardiac segmentation. The delineated structures are then transferred to the 4D Flow images for cardiovascular flow analysis. Cine MR images are however acquired with multiple breath-holds, which can be challenging for some people, especially, when a cardiovascular disease is present. Consequently, unexpected breathing motion by a patient may lead to misalignments between the acquired cine MR images.

The goal of the thesis is to test the feasibility of an automatic image registration method to correct the misalignment caused by respiratory motion in morphological 2D cine MR images by using the 4D Flow MR as the reference image. As a registration method relies on a set of optimal parameters to provide desired results, a comprehensive investigation was performed to find such parameters. Different combinations of registration parameters settings were applied on 20 datasets from both healthy volunteers and patients. The best combinations, selected on the basis of normalized cross-correlation, were evaluated using the clinical gold-standard by employing widely used geometric measures of spatial correspondence. The accuracy of the best parameters from geometric evaluation was finally validated by using simulated misalignments.

Using a registration method consisting of only translation improved the results for both datasets from healthy volunteers and patients and the simulated misalignment data. For the datasets from healthy volunteers and patients, the registration improved the results from 0.7074 ± 0.1644 to 0.7551 ± 0.0737 in Dice index and from 1.8818 ± 0.9269 to 1.5953 ± 0.5192 for point-to-curve error. These values are a mean value for all the 20 datasets.

The results from geometric evaluation on the data from both healthy volunteers and patients show that the developed correction method is able to improve the alignment of the cine MR images. This allows a reliable segmentation of 4D flow MR images for cardiac flow assessment.

Since most people now have a high-performing computing device with an attached camera in their pocket, in the form of a smartphone, robotics and computer vision researchers are thrilled about the possibility this creates. Such devices have previously been used in robotics to create 3D maps of environments and objects by feeding the camera data to a 3D reconstruction algorithm.

The big downside with smartphones is that their cameras use a different sensor than what is usually used in robotics, namely a rolling shutter camera.These cameras are cheaper to produce but are not as well suited for general 3D reconstruction algorithms as the global shutter cameras typically used in robotics research. One recent, accurate and performance effective 3D reconstruction method which could be used on a mobile device, if tweaked, is LSD-SLAM.

This thesis uses the LSD-SLAM method developed for global shutter cameras and incorporates additional methods developed allow the usage of rolling shutter data.The developed method is evaluated by calculating numbers of failed 3D reconstructions before a successful one is obtained when using rolling shutter data.The result is a method which improves this metric with about 70\% compared to the unedited LSD-SLAM method.

The usage of 3D modelling is increasing fast, both for civilian and military areas, such as navigation, targeting and urban planning. When creating a 3D model from satellite images, clouds canbe problematic. Thus, automatic detection ofclouds inthe imagesis ofgreat use. This master thesis was carried out at Vricon, who produces 3D models of the earth from satellite images.This thesis aimed to investigate if Support Vector Machines could classify pixels into cloud or non-cloud, with a combination of texture and color as features. To solve the stated goal, the task was divided into several subproblems, where the first part was to extract features from the images. Then the images were preprocessed before fed to the classifier. After that, the classifier was trained, and finally evaluated.The two methods that gave the best results in this thesis had approximately 95 % correctly classified pixels. This result is better than the existing cloud segmentation method at Vricon, for the tested terrain and cloud types.

Automatisk travmätning är ett mätsystem som mäter vedvolymen på virkeslastbilar. Systemet består av sex stycken sensor-system. Varje sensor kalibreras först individuellt och sedan ihop för att ge ett sammanfogat världskoordinat system. Varje sensor genererar en djupbild och en reflektansbild, där värdena i djupbilden representerar avståndet från kameran. Uppdragsgivaren har utvecklat en algoritm som utifrån mätdatat(bilderna)  uppskattar vedvolymen till en viss noggrannhet som uppfyller kraven ställda av skogsindustrin för automatisk mätning av travar på virkeslastbil. I den här rapporten undersöks om bättre mätresultat kan uppnås exempelvis med andra metoder eller kombinationer av dem.Till förfogande finns ca 125 dataset av travar där facit finns. Facit består av manuella stickprovsmätningar där varje enskild stock mätts för sig. Initialt valdes aktivt att inte sätta sig in i uppdragsgivarens algoritm för att inte bli färgad av hur de kommit fram till sina resultat. Främst används fram- och baksidebilderna av entrave för att hitta stockarna. Därefter interpoleras de funna stockarna in till mitten av traven eller så paras stockarna ihop från de båda sidorna. Ibland finns vissa problem med bilderna. Oftast är minst en av sidorna ockluderade av lastbilshytten, kranen eller en annan trave. Då gäller det att hitta uppskattning utifrån det data man ser för fylla upp de skymda områdena.I början av examensarbetet användes två metoder(MSER och Punktplanmetoden) för undersöka om man kunde uppnå bra resultat utifrån att enbart mäta datat och användadet som initial gissning till volymen. Dock upptäcktes det att värdefulla detaljer   i dataseten missades för att mer noggrant bestämma vedvolymen. Exempel på sådan data är fördelningen av diametern på de funna stockändarna. Tillika tenderades kraftig överestimering när travarna innehöll en viss mängd ris och eller dåligt kvistade stockar. Därefter konstruerades en geometrisk metod, och det var den här metoden som det lades mest tid på.I figurerna nedan visas en tabell och en graf där alla tre metoders resultat under bark(UB) visas och intervall gränserna för att uppfylla kraven ställda av skogsindustrin.

The Kinect v2 is a RGB-D sensor manufactured as a gesture interaction tool for the entertainment console XBOX One. In this thesis we will use it to perform 3D reconstruction and investigate its ability to measure depth. In order to sense both color and depth the Kinect v2 has two cameras: one RGB camera and one infrared camera used to produce depth and near infrared images. These cameras need to be calibrated if we want to use them for 3D reconstruction. We present a calibration procedure for simultaneously calibrating the cameras and extracting their relative pose. This enables us to construct colored meshes of the environment. When we know the camera parameters of the infrared camera, the depth images could be used to perform the Kinectfusion algorithm. This produces well-formed meshes of the environment by combining many depth frames taken from several camera poses.The Kinect v2 uses a time-of-flight technology were the phase shifts are extracted from amplitude modulated infrared light signals produced by an emitter. The extracted phase shifts are then converted to depth values. However, the extraction of phase shifts includes a phase unwrapping procedure, which is sensitive to noise and can result in large depth errors.By utilizing the ability to access the raw phase measurements from the device we managed to modify the phase unwrapping procedure. This new procedure includes an extraction of several hypotheses for the unwrapped phase and a spatial propagation to select amongst them. This proposed method has been compared with the available drivers in the open source library libfreenect2 and the Microsoft Kinect SDK v2. Our experiments show that the depth images of the two available drivers have similar quality and our proposed method improves over libfreenect2. The calculations in the proposed method are more expensive than those in libfreenect2 but it still runs at 2.5× real time. However, contrary to libfreenect2 the proposed method lacks a filter that removes outliers from the depth images. It turned out that this is an important feature when performing Kinect fusion and future work should thus be focused on adding an outlier filter.

Autonomous Driving or self driving vehicles are concepts of vehicles knowing their environment and making driving manoeuvres without instructions from a driver. The concepts have been around for decades but has improved significantly in the last years since research in this area has made significant progress. Benefits of autonomous driving include the possibility to decrease the number of accidents in traffic and thereby saving lives.

A major challenge in autonomous driving is to acquire 3D information and relations between all objects in surrounding traffic. This is referred to as \textit{spatial perception}. Stereo camera systems have become a central sensor module for advanced driver assistance systems and autonomous driving. For object detection and measurements at large distances stereo vision encounter difficulties. This includes objects being small, having low contrast and the presence of image noise. Having an accurate perception of the environment at large distances is however of high interest for many applications, especially autonomous driving.

This thesis proposes a method which tries to increase the range to where generic objects are first detected using a given stereo camera setup. Objects are represented by planes in 3D space. The input image is segmented into the various objects and the 3D plane parameters are estimated jointly. The 3D plane parameters are estimated directly from the stereo image pairs. In particular, this thesis investigates methods to introduce geometric constraints to the segmentation or labeling task, i.e assigning each considered pixel in the image to a plane.

The methods provided in this thesis show that despite the difficulties at large distances it is possible to exploit planar primitives in 3D space for obstacle detection at distances where other methods fail.

In the steel industry, laser triangulation based measurement systems can be utilizedfor evaluating the flatness of the steel products. Shapeline is a company in Linköpingthat manufactures such measurement systems. This thesis work will present a series ofexperiments on a Shapeline measurement system in a relatively untested environment, thehot rolling mill at SSAB in Borlänge.The purpose of this work is to evaluate how the conditions at a hot rolling mill affectsthe measurement performance. It has been anticipated that measuring in high temperatureenvironment would introduce difficulties that do not exist when measuring in cold environments.A number of different experiments were conducted, where equipment such as laserand camera bandpass filter were alternated. Via the experiments, information about noisedue to the environment in the hot rolling mill was gained. The most significant noise wascaused by heat shimmering. Using the presented methods, the magnitude and frequencyspectrum of the heat shimmering noise could be determined. The results also indicates thatheat shimmering cause large errors and is quite troublesome to counter. In addition to this,the quality of the line detections under the hot rolling mill circumstances was examined. Itcould be observed that the line detections did not introduce any significant errors despitethe harmful conditions.

In the field of industrial automation large savings can be realized if position andorientation of an object is known. Knowledge about an objects position and orien-tation can be used by advanced robotic systems to be able to work with complexitems. Specifically 2D-objects are a big enough sub domain to motivate specialattention. Traditionally this problem has been solved with large mechanical sys-tems that forces the objects into specific configurations. Besides being expensive,taking up a lot of space and having great difficulty handling fragile items, thesemechanical systems have to be constructed for each particular type of object. Thisthesis explores the possibility of using registration algorithms from computer vi-sion based on 3D-data to find flat objects. While systems for locating 3D objectsalready exists they have issues with locating essentially flat objects since theirpositioning is mostly a function of their contour. The thesis consists of a briefexamination of 2D-algorithms and their extension to 3D as well as results fromthe most suitable algorithm.

Anomaly detection is a general theory of detecting unusual patterns or events in data. This master thesis investigates the subject of anomaly detection in two different applications. The first application is product inspection using a camera and the second application is surveillance using a 2D laser scanner.

The first part of the thesis presents a system for automatic visual defect inspection. The system is based on aligning the images of the product to a common template and doing pixel-wise comparisons. The system is trained using only images of products that are defined as normal, i.e. non-defective products. The visual properties of the inspected products are modelled using three different methods. The performance of the system and the different methods have been evaluated on four different datasets.

The second part of the thesis presents a surveillance system based on a single laser range scanner. The system is able to detect certain anomalous events based on the time, position and velocities of individual objects in the scene. The practical usefulness of the system is made plausible by a qualitative evaluation using unlabelled data.

Generic visual tracking is one of the classical problems in computer vision. In this problem, no prior knowledge of the target is available aside from a bounding box in the initial frame of the sequence. The generic visual tracking is a difficult task due to a number of factors such as momentary occlusions, target rotations, changes in target illumination and variations in the target size. In recent years, discriminative correlation filter (DCF) based trackers have shown promising results for visual tracking. These DCF based methods use the Fourier transform to efficiently calculate detection and model updates, allowing significantly higher frame rates than competing methods. However, existing DCF based methods only estimate translation of the object while ignoring changes in size.This thesis investigates the problem of accurately estimating the scale variations within a DCF based framework. A novel scale estimation method is proposed by explicitly constructing translation and scale filters. The proposed scale estimation technique is robust and significantly improve the tracking performance, while operating at real-time. In addition, a comprehensive evaluation of feature representations in a DCF framework is performed. Experiments are performed on the benchmark OTB-2015 dataset, as well as the VOT 2014 dataset. The proposed methods are shown to significantly improve the performance of existing DCF based trackers.

Machine learning can be utilized in many different ways in the field of automatic manufacturing and logistics. In this thesis supervised machine learning have been utilized to train a classifiers for detection and recognition of objects in images. The techniques AdaBoost and Random forest have been examined, both are based on decision trees.

The thesis has considered two applications: barcode detection and optical character recognition (OCR). Supervised machine learning methods are highly appropriate in both applications since both barcodes and printed characters generally are rather distinguishable.

The first part of this thesis examines the use of machine learning for barcode detection in images, both traditional 1D-barcodes and the more recent Maxi-codes, which is a type of two-dimensional barcode. In this part the focus has been to train classifiers with the technique AdaBoost. The Maxi-code detection is mainly done with Local binary pattern features. For detection of 1D-codes, features are calculated from the structure tensor. The classifiers have been evaluated with around 200 real test images, containing barcodes, and shows promising results.

The second part of the thesis involves optical character recognition. The focus in this part has been to train a Random forest classifier by using the technique point pair features. The performance has also been compared with the more proven and widely used Haar-features. Although, the result shows that Haar-features are superior in terms of accuracy. Nevertheless the conclusion is that point pairs can be utilized as features for Random forest in OCR.

The interest of using GPU:s as general processing units for heavy computations (GPGPU) has increased in the last couple of years. Manufacturers such as Nvidia and AMD make GPU:s powerful enough to outrun CPU:s in one order of magnitude, for suitable algorithms. For embedded systems, GPU:s are not as popular yet. The embedded GPU:s available on the market have often not been able to justify hardware changes from the current systems (CPU:s and FPGA:s) to systems using embedded GPU:s. They have been too hard to get, too energy consuming and not suitable for some algorithms. At SICK IVP, advanced computer vision algorithms run on FPGA:s. This master thesis optimizes two such algorithms for embedded GPU:s and evaluates the result. It also evaluates the status of the embedded GPU:s on the market today. The results indicates that embedded GPU:s perform well enough to run the evaluatedd algorithms as fast as needed. The implementations are also easy to understand compared to implementations for FPGA:s which are competing hardware.

High resolution 3D images are of high interest in military operations, where data can be used to classify and identify targets. The Swedish defence research agency (FOI) is interested in the latest research and technologies in this area. A draw- back with normal 3D-laser systems are the lack of high resolution for long range measurements. One technique for high long range resolution laser radar is based on time correlated single photon counting (TCSPC). By repetitively sending out short laser pulses and measure the time of flight (TOF) of single reflected pho- tons, extremely accurate range measurements can be done. A drawback with this method is that it is hard to create single photon detectors with many pixels and high temporal resolution, hence a single detector is used. Scanning an entire scene with one detector is very time consuming and instead, as this thesis is all about, the entire scene can be measured with less measurements than the number of pixels. To do this a technique called compressed sensing (CS) is introduced. CS utilizes that signals normally are compressible and can be represented sparse in some basis representation. CS sets other requirements on the sampling compared to the normal Shannon-Nyquist sampling theorem. With a digital micromirror device (DMD) linear combinations of the scene can be reflected onto the single photon detector, creating scalar intensity values as measurements. This means that fewer DMD-patterns than the number of pixels can reconstruct the entire 3D-scene. In this thesis a computer model of the laser system helps to evaluate different CS reconstruction methods with different scenarios of the laser system and the scene. The results show how many measurements that are required to reconstruct scenes properly and how the DMD-patterns effect the results. CS proves to enable a great reduction, 85 − 95 %, of the required measurements com- pared to pixel-by-pixel scanning system. Total variation minimization proves to be the best choice of reconstruction method. 

In recent years several depth cameras have emerged on the consumer market, creating many interesting possibilities forboth professional and recreational usage. One example of such a camera is the Microsoft Kinect sensor originally usedwith the Microsoft Xbox 360 game console. In this master thesis a system is presented that utilizes this device in order to create an as accurate as possible 3D reconstruction of an indoor environment. The major novelty of the presented system is the data structure based on signed distance fields and voxel octrees used to represent the observed environment.

Many methods have been developed for visual tracking of generic objects. The vast majority of these assume the world is two-dimensional, either ignoring the third dimension or only dealing with it indirectly. This causes difficulties for the tracker when the target approaches or moves away from the camera, is occluded or moves out of the camera frame.

Unmanned aerial vehicles (UAVs) are increasingly used in civilian applications and some of these will undoubtedly carry tracking systems in the future. As they move around, these trackers will encounter both scale changes and occlusions. To improve the tracking performance in these cases, the third dimension should be taken into account.

This thesis extends the capabilities of a 2D tracker to three dimensions, with the assumption that the target moves on a ground plane.

The position of the tracker camera is established by matching the video it produces to a sparse point-cloud map built with off-the-shelf structure-from-motion software. A target is tracked with a generic 2D tracker and subsequently positioned on the ground. Should the target disappear from view, its motion on the ground is predicted. In combination, these simple techniques are shown to improve the robustness of a tracking system on a moving platform under target scale changes and occlusions.

Visual simultaneous localization and mapping (SLAM) as field has been researched for ten years, but with recent advances in mobile performance visual SLAM is entering the consumer market in a completely new way. A visual SLAM system will however be sensitive to non cautious use that may result in severe motion, occlusion or poor surroundings in terms of visual features that will cause the system to temporarily fail. The procedure of recovering from such a fail is called relocalization. Together with two similar problems localization, to find your position in an existing SLAM session, and loop closing, the online reparation and perfection of the map in an active SLAM session, these can be grouped as visual location recognition (VLR).

This thesis presents novel results by combining the scalability of FabMap and the precision of 13th Lab's tracking yielding high-precision VLR, +/- 10 cm, while maintaining above 99 % precision and 60 % recall for sessions containing thousands of images. Everything functional purely on a normal mobile phone.

The applications of VLR are many. Indoors, where GPS is not functioning, VLR can still provide positional information and navigate you through big complexes like airports and museums. Outdoors, VLR can improve the precision of GPS tenfold yielding a new level of navigational experience. Virtual and augmented reality applications are other areas that benefit from improved positioning and localization.

Recording a video sequence with a camera during movement often produces blurred results. This is mainly due to motion blur which is caused by rapid movement of objects in the scene or the camera during recording. By correcting for changes in the orientation of the camera, caused by e.g. uneven terrain, it is possible to minimize the motion blur and thus, produce a stabilized video.

In order to do this, data gathered from a gyroscope and the camera itself can be used to measure the orientation of the camera. The raw data needs to be processed, synchronized and filtered to produce a robust estimate of the orientation. This estimate can then be used as input to some automatic control system in order to correct for changes in the orientation

This thesis focuses on examining the possibility of such a stabilization. The actual stabilization is left for future work. An evaluation of the hardware as well as the implemented methods are done with emphasis on speed, which is crucial in real time computing.

A laser triangulating camera system projects a laser line onto an object to create height curveson the object surface. By moving the object, height curves from different parts of the objectcan be observed and combined to produce a three dimensional representation of the object.The calibration of such a camera system involves transforming received data to get real worldmeasurements instead of pixel based measurements.

The calibration method presented in this thesis focuses specifically on small fields ofview. The goal is to provide an easy to use and robust calibration method that can complementalready existing calibration methods. The tool should get as good measurementsin metric units as possible, while still keeping complexity and production costs of the calibrationobject low. The implementation uses only data from the laser plane itself making itusable also in environments where no external light exist.

The proposed implementation utilises a complete scan of a three dimensional calibrationobject and returns a calibration for three dimensions. The results of the calibration havebeen evaluated against synthetic and real data.

Examensarbetet har gått ut på att undersöka om det är möjligt att jämföra foto- grafier på havsbotten tagna med en kamera monterad på SAAB Dynamics farkost AUV-62, här kallad Sapphires, med SONAR-bilder tagna från samma farkost men vid ett annat tillfälle. Föremål avbildade med kamera och sidriktade SONAR:er delar i normalfallet inte visuellt utseende och är därför svåra att jämföra. Meto- den som valts för att jämföra kamera- och SONAR-bilderna grundar sig av denna anledning inte på föremåls individuella utseede utan på mönster skapade av flera föremål. Föremål i bilderna identifieras som objekt, vilka beskrivs med en posi- tion i long. lat. och radie. I kamerabilderna identifieras objekt genom att segmen- tera bilderna med hjälp av MSER, där stenar och andra föremål har ett avvikand utseende från bakgrunden bestående av sand. I SONAR-bilden identifieras om- råden med objekt genom att studera hög intensitet på ekosvaren vilket motsva- rar föremål som reflekterat ljudpulserna bra, från dessa skapas objekten genom att använda MSER på dessa områden. De två uppsättningarna med objekt, från kamera- och SONAR-bilden, jämförs sedan genom att alla objekt i kamera-bilden jämförs med alla objekt i SONAR-bilden genom att translatera efter hypotesen att de är samma objekt och se hur många av dem övriga som passar in i antagandet.

In Sweden and many other northern countries, it is common for heat to be distributed to homes and industries through district heating networks. Such networks consist of pipes buried underground carrying hot water or steam with temperatures in the range of 90-150 C. Due to bad insulation or cracks, heat or water leakages might appear.

A system for large-scale monitoring of district heating networks through remote thermography has been developed and is in use at the company Termisk Systemteknik AB. Infrared images are captured from an aircraft and analysed, finding and indicating the areas for which the ground temperature is higher than normal. During the analysis there are, however, many other warm areas than true water or energy leakages that are marked as detections. Objects or phenomena that can cause false alarms are those who, for some reason, are warmer than their surroundings, for example, chimneys, cars and heat leakages from buildings.

During the last couple of years, the system has been used in a number of cities. Therefore, there exists a fair amount of examples of different types of detections. The purpose of the present master’s thesis is to evaluate the reduction of false alarms of the existing analysis that can be achieved with the use of a learning system, i.e. a system which can learn how to recognize different types of detections. 

A labelled data set for training and testing was acquired by contact with customers. Furthermore, a number of features describing the intensity difference within the detection, its shape and propagation as well as proximity information were found, implemented and evaluated. Finally, four different classifiers and other methods for classification were evaluated.

The method that obtained the best results consists of two steps. In the initial step, all detections which lie on top of a building are removed from the data set of labelled detections. The second step consists of classification using a Random forest classifier. Using this two-step method, the number of false alarms is reduced by 43% while the percentage of water and energy detections correctly classified is 99%.

Identification of individuals has been solved with many different solutions around the world, either using biometric data or external means of verification such as id cards or RFID tags. The advantage of using biometric measurements is that they are directly tied to the individual and are usually unalterable. Acquiring dependable measurements is however challenging when the individuals are uncooperative. A dependable system should be able to deal with this and produce reliable identifications.

The system proposed in this thesis can autonomously classify uncooperative specimens from depth data. The data is acquired from a depth camera mounted in an uncontrolled environment, where it was allowed to continuously record for two weeks. This requires stable data extraction and normalization algorithms to produce good representations of the specimens. Robust descriptors can therefore be extracted from each sample of a specimen and together with different classification algorithms, the system can be trained or validated. Even with as many as 138 different classes the system achieves high recognition rates. Inspired by the research field of face recognition, the best classification algorithm, the method of fisherfaces, was able to accurately recognize 99.6% of the validation samples. Followed by two variations of the method of eigenfaces, achieving recognition rates of 98.8% and 97.9%. These results affirm that the capabilities of the system are adequate for a commercial implementation.

Because of the increase in the number of security cameras, there is more video footage available than a human could efficiently process. In combination with the fact that computers are getting more efficient, it is getting more and more interesting to solve the problem of detecting and recognizing people automatically.

Therefore a method is proposed for estimating a 3D-path of a person of interest in multiple, non overlapping, monocular cameras. This project is a collaboration between two master theses. This thesis will focus on recognizing a person of interest from several possible candidates, as well as estimating the 3D-position of a person and providing a graphical user interface for the system. The recognition of the person of interest includes keeping track of said person frame by frame, and identifying said person in video sequences where the person of interest has not been seen before.

The final product is able to both detect and recognize people in video, as well as estimating their 3D-position relative to the camera. The product is modular and any part can be improved or changed completely, without changing the rest of the product. This results in a highly versatile product which can be tailored for any given situation.

Visual tracking is a classical computer vision problem with many important applications in areas such as robotics, surveillance and driver assistance. The task is to follow a target in an image sequence. The target can be any object of interest, for example a human, a car or a football. Humans perform accurate visual tracking with little effort, while it remains a difficult computer vision problem. It imposes major challenges, such as appearance changes, occlusions and background clutter. Visual tracking is thus an open research topic, but significant progress has been made in the last few years.

The first part of this thesis explores generic tracking, where nothing is known about the target except for its initial location in the sequence. A specific family of generic trackers that exploit the FFT for faster tracking-by-detection is studied. Among these, the CSK tracker have recently shown obtain competitive performance at extraordinary low computational costs. Three contributions are made to this type of trackers. Firstly, a new method for learning the target appearance is proposed and shown to outperform the original method. Secondly, different color descriptors are investigated for the tracking purpose. Evaluations show that the best descriptor greatly improves the tracking performance. Thirdly, an adaptive dimensionality reduction technique is proposed, which adaptively chooses the most important feature combinations to use. This technique significantly reduces the computational cost of the tracking task. Extensive evaluations show that the proposed tracker outperform state-of-the-art methods in literature, while operating at several times higher frame rate.

In the second part of this thesis, the proposed generic tracking method is applied to human tracking in surveillance applications. A causal framework is constructed, that automatically detects and tracks humans in the scene. The system fuses information from generic tracking and state-of-the-art object detection in a Bayesian filtering framework. In addition, the system incorporates the identification and tracking of specific human parts to achieve better robustness and performance. Tracking results are demonstrated on a real-world benchmark sequence.

The introduction of dual energy CT, DECT, in the field of medical healthcare has made it possible to extract more information of the scanned objects. This in turn has the potential to improve the accuracy in radiation therapy dose planning. One problem that remains before successful material decomposition can be achieved however, is the presence of beam hardening and scatter artifacts that arise in a scan. Methods currently in clinical use for removal of beam hardening often bias the CT numbers. Hence, the possibility for an appropriate tissue decomposition is limited.

Here a method for successful decomposition as well as removal of the beam hardening artifact is presented. The method uses effective linear attenuations for the five base materials, water, protein, adipose, cortical bone and marrow, to perform the decomposition on reconstructed simulated data. This is performed inside an iterative loop together with the polychromatic x-ray spectra to remove the beam hardening

This is a master thesis of the Master of Science degree program in Applied Physics and Electrical Engineering at Linköping University. The goal of this thesis is to find out how the Microsoft Kinect can be used as a part of a camera rig to create accurate 3D-models of an indoor environment. The Microsoft Kinect is marketed as a touch free game controller for the Microsoft Xbox 360 game console. The Kinect contains a color and a depth camera. The depth camera works by constantly projecting a near infrared dot pattern that is observed with a near infrared camera. In this thesis it is described how to model the near infrared projector pattern to enable external near infrared cameras to be used to improve the measurement precision. The depth data that the Kinect output have been studied to determine what types of errors it contains. The finding was that the Kinect uses an online calibration algorithm that changes the depth data.

This is a master thesis of the Master of Science degree program in Applied Physics and Electrical Engineering (Y) at Linköping University. The goal of the projectis to develop an application for creating a map in real time from a video camera on a miniature unmanned aerial vehicle. This thesis project and report is a first exploratory study for this application. It implements a prototype method and evaluates it on sample sequences from an on-board video camera. The method first looks for good points to follow in the image and then tracks them in a sequence.The image is then pasted, or merged, together with previous images so that points from the different images align.

Two methods to find good points to follow are examined with focus on real-time performance. The result is that the much faster FAST detector method yielded satisfactory results good enough to replace the slower standard method of the Harris-Stephens corner detector.

It is also examined whether it is possible to assume that the ground is a flat surface in this application or if a computationally more expensive method estimating altitude information has to be used. The result is that at high altitudes or when the ground is close to flat in reality and the camera points straight downwards a two-dimensional method will do. If flying lower or with high objects in the picture, which is often the case in this application, it must to be taken into account that the points really are at different heights, hence the ground can not be assumed to be flat.

When patients move during an MRI examination, severe artifacts arise in the reconstructed image and motion correction is therefore often desired. An in-plane motion correction algorithm suitable for PRESTO-CAN, a new 3D functional MRI method where sampling of k-space is radial in kx-direction and kz-direction and Cartesian in ky-direction, was implemented in this thesis work.

Rotation and translation movements can be estimated and corrected for sepa- rately since the magnitude of the data is only affected by the rotation. The data were sampled in a radial pattern and the rotation was estimated by finding the translation in angular direction using circular correlation. Correlation was also used when finding the translation in x-direction and z-direction.

The motion correction algorithm was evaluated on computer simulated data, the motion was detected and corrected for, and this resulted in images with greatly reduced artifacts due to patient movements. 

Today, 3D models of cities are created from aerial images using a camera rig. Images, together with sensor data from the flights, are stored for further processing when building 3D models. However, there is a market demand for a more mobile solution of satisfactory quality. If the camera position can be calculated for each image, there is an existing algorithm available for the creation of 3D models.

This master thesis project aims to investigate whether the iPhone 4 offers good enough image and sensor data quality from which 3D models can be created. Calculations on movements and rotations from sensor data forms the foundation of the image processing, and should refine the camera position estimations.

The 3D models are built only from image processing since sensor data cannot be used due to poor data accuracy. Because of that, the scaling of the 3D models are unknown and a measurement is needed on the real objects to make scaling possible. Compared to a test algorithm that calculates 3D models from only images, already available at the SBD’s system, the quality of the 3D model in this master thesis project is almost the same or, in some respects, even better when compared with the human eye.

A common computer vision task is navigation and mapping. Many indoor navigation tasks require depth knowledge of flat, unstructured surfaces (walls, floor, ceiling). With passive illumination only, this is an ill-posed problem. Inspired by small children using a torchlight, we use a spotlight for active illumination. Using our torchlight approach, depth and orientation estimation of unstructured, flat surfaces boils down to estimation of ellipse parameters. The extraction of ellipses is very robust and requires little computational effort.

MRI (Magnetic Resonance Imaging) is a medical imaging method that uses magnetic fields in order to retrieve images of the human body. This thesis revolves around a novel acquisition method of 3D fMRI (functional Magnetic Resonance Imaging) called PRESTO-CAN that uses a radial pattern in order to sample the (kx,kz)-plane of k-space (the frequency domain), and a Cartesian sample pattern in the ky-direction. The radial sample pattern allows for a denser sampling of the central parts of k-space, which contain the most basic frequency information about the structure of the recorded object. This allows for higher temporal resolution to be achieved compared with other sampling methods since a fewer amount of total samples are needed in order to retrieve enough information about how the object has changed over time. Since fMRI is mainly used for monitoring blood flow in the brain, increased temporal resolution means that we can be able to track fast changes in brain activity more efficiently.The temporal resolution can be further improved by reducing the time needed for scanning, which in turn can be achieved by applying parallel imaging. One such parallel imaging method is SENSE (SENSitivity Encoding). The scan time is reduced by decreasing the sampling density, which causes aliasing in the recorded images. The aliasing is removed by the SENSE method by utilizing the extra information provided by the fact that multiple receiver coils with differing sensitivities are used during the acquisition. By measuring the sensitivities of the respective receiver coils and solving an equation system with the aliased images, it is possible to calculate how they would have looked like without aliasing.In this master thesis, SENSE has been successfully implemented in PRESTO-CAN. By using normalized convolution in order to refine the sensitivity maps of the receiver coils, images with satisfying quality was able to be reconstructed when reducing the k-space sample rate by a factor of 2, and images of relatively good quality also when the sample rate was reduced by a factor of 4. In this way, this thesis has been able to contribute to the improvement of the temporal resolution of the PRESTO-CAN method.

This report introduces some simplications to the method by Fitzgibbon et al. that allows for 3D model construction from turn-table sequences. It is assumed that the reader has previously read in order to fully understand this report.

Fitzgibbon et al. presents a method for 3D model construction that utilizes the extra constraints imposed by turn-table sequences. Restricting the scenario to a turn-table sequence with a single camera with xed settings produces these extra constraints:

C1. The internal parameters for the camera are the same for all images

C2. The motion of the camera can be described by a rotation around a singleaxis

It is shown that in the uncalibrated case the number of parameters to estimate is m + 8 where m is the number of images.

We further simplify the problem by using extra constrains given from the fact that we know:

C3. The internal parameters of the camera, i.e the K matrix

C4. That the angle between each pair of consecutive cameras is the same

Using these extra simplications makes it possible to create a 3D model from realistic data without using Bundle Adjustment.

Most mobile video-recording devices of today, e.g. cell phones and music players, make use of a rolling shutter camera. A rolling shutter camera captures video by recording every frame line-by-line from top to bottom of the image, leading to image distortions in situations where either the device or the target is moving. Recording video by hand also leads to visible frame-to-frame jitter.

In this thesis, methods to decrease distortion caused by the motion of a video-recording device with a rolling shutter camera are presented. The methods are based on estimating the orientation of the camera from gyroscope and accelerometer measurements.

The algorithms are implemented on the iPod Touch 4, and the resulting videos are compared to those of competing stabilization software, both commercial and free, in a series of blind experiments. The results from this user study shows that the methods presented in the thesis perform equal to or better than the others.

3D cameras delivering height data can be used for quality inspection of goods on a conveyor.

It is then of interest to distinguish the important parts of the image from background and noise and further to divide these interesting parts into segments that have a strong correlation to objects on the conveyor belt.

Segmentation can easily be done by thresholding in the simple case. However, in more complex situations, for example when objects touch or overlap, this does not work well.

In this thesis, research and evaluation of a few different methods for segmentation of height image data are presented. The focus is to find an accurate method for segmentation of smooth irregularly shaped organic objects such as vegetables or shellfish.

For evaluative purposes a database consisting of height images depicting a variety of such organic objects has been collected.

We show in the thesis that a conventional gradient magnitude method is hard to beat in the general case. If, however, the objects to be segmented are heavily non-convex with a lot of crests and valleys within themselves one could be better off choosing a normalized least squares method.

This thesis treats topics within the area of object recognition. A real-time view matching method has been developed to compute the transformation between two different images of the same scene. This method uses a color based region detector called MSCR and affine transformations of these regions to create affine-invariant patches that are used as input to the SIFT algorithm. A parallel method to compute the SIFT descriptor has been created with relaxed constraints so that the descriptor size and the number of histogram bins can be adjusted. Additionally, a matching step to deduce correspondences and a parallel RANSAC method have been created to estimate the undergone transformation between these descriptors. To achieve real-time performance, the implementation has been targeted to use the parallel nature of the GPU with CUDA as the programming language. Focus has been put on the architecture of the GPU to find the best way to parallelize the different processing steps. CUDA has also been combined with OpenGL to be able to use the hardware accelerated anisotropic sampling method for affine transformations of regions. Parts of the implementation can also be used individually from either Matlab or by using the provided C++ library directly. The method was also evaluated in terms of accuracy and speed. It was shown that our algorithm has similar or better accuracy at finding correspondences than SIFT when the 3D geometry changes are large but we get a slightly worse result on images with flat surfaces.

Man portable air defence systems, MANPADS, pose a big threat to civilian and military aircraft. This thesis aims to find methods that could be used in a missile approach warning system based on infrared cameras.

The two main tasks of the completed system are to classify the type of missile, and also to estimate its position and velocity from a sequence of images.

The classification is based on hidden Markov models, one-class classifiers, and multi-class classifiers.

Position and velocity estimation uses a model of the observed intensity as a function of real intensity, image coordinates, distance and missile orientation. The estimation is made by an extended Kalman filter.

We show that fast classification of missiles based on radiometric data and a hidden Markov model is possible and works well, although more data would be needed to verify the results.

Estimating the position and velocity works fairly well if the initial parameters are known. Unfortunately, some of these parameters can not be computed using the available sensor data.

This master thesis investigates the difficulties of constructing a depth map using one low resolution grayscale camera mounted in the front of a car. The goal is to produce a depth map in real-time to assist other algorithms in the safety system of a car. This has been shown to be difficult using the evaluated combination of camera position and choice of algorithms.

The main problem is to estimate an accurate optical flow. Another problem is to handle moving objects. The conclusion is that the implementations, mainly triangulation of corresponding points tracked using a Lucas Kanade tracker, provide information of too poor quality to be useful for the safety system of a car.

Foreground segmentation is a common first step in tracking and surveillance applications.  The purpose of foreground segmentation is to provide later stages of image processing with an indication of where interesting data can be found.  This thesis is an investigation of how foreground segmentation can be performed in two contexts: as a pre-step to trajectory tracking and as a pre-step in indoor surveillance applications.

Three methods are selected and detailed: a single Gaussian method, a Gaussian mixture model method, and a codebook method.  Experiments are then performed on typical input video using the methods.  It is concluded that the Gaussian mixture model produces the output which yields the best trajectories when used as input to the trajectory tracker.  An extension is proposed to the Gaussian mixture model which reduces shadow, improving the performance of foreground segmentation in the surveillance context.

In this work we present a region detector, an adaptation to range data of the popular Maximally Stable Extremal Regions (MSER) region detector. We call this new detector Maximally Robust Range Regions (MRRR). We apply the new detector to real range data captured by a commercially available laser range camera. Using this data we evaluate the repeatability of the new detector and compare it to some other recently published detectors. The presented detector shows a repeatability which is better or the same as the best of the other detectors. The MRRR detector also offers additional data on the detected regions. The additional data could be crucial in applications such as registration or recognition.

Within this thesis an algorithm for object recognition called Cluster Matching has been developed, implemented and evaluated. The image information is sampled at arbitrary sample points, instead of interest points, and local image features are extracted. These sample points are used as a compact representation of the image data and can quickly be searched for prior known objects. The algorithm is evaluated on a test set of images and the result is surprisingly reliable and time efficient.

Gaze tracking is the estimation of the point in space a person is “looking at”. This is widely used in both diagnostic and interactive applications, such as visual attention studies and human-computer interaction. The most common commercial solution used to track gaze today uses a combination of infrared illumination and one or more cameras. These commercial solutions are reliable and accurate, but often expensive. The aim of this thesis is to construct a simple single-camera gaze tracker from off-the-shelf components. The method used for gaze tracking is based on infrared illumination and a schematic model of the human eye. Based on images of reflections of specific light sources in the surfaces of the eye the user’s gaze point will be estimated. Evaluation is also performed on both the software and hardware components separately, and on the system as a whole. Accuracy is measured in spatial and angular deviation and the result is an average accuracy of approximately one degree on synthetic data and 0.24 to 1.5 degrees on real images at a range of 600 mm.

This Master Thesis has been conducted at the National Laboratory of Forensic Science (SKL) in Linköping. When images that are to be analyzed at SKL, presenting an interesting object, are of bad quality there may be a need to enhance them. If several images with the object are available, the total amount of information can be used in order to estimate one single enhanced image. A program to do this has been developed by studying methods for image registration and high resolution image estimation. Tests of important parts of the procedure have been conducted. The final results are satisfying and the key to a good high resolution image seems to be the precision of the image registration. Improvements of this part may lead to even better results. More suggestions for further improvementshave been proposed.

an Internal Navigational System and a Global Navigational Satellite System (GNSS). In navigational warfare the GNSS can be jammed, therefore are a third navigational system is needed. The system that has been tried in this thesis is camera based navigation. Through a video camera and a sensor reference the position is determined. This thesis will process the matching between the sensor reference and the video image.

Two methods have been implemented: normalized cross correlation and position determination through a homography. Normalized cross correlation creates a correlation matrix. The other method uses point correspondences between the images to determine a homography between the images. And through the homography obtain a position. The more point correspondences the better the position determination will be.

The results have been quite good. The methods have got the right position when the Euler angles of the UAV have been known. Normalized cross correlation has been the best method of the tested methods.

The purpose of this master thesis was to study computer vision algorithms for vehicle detection in monochrome images captured by mono camera. The work has mainly been focused on detecting rear-view cars in daylight conditions. Previous work in the literature have been revised and algorithms based on edges, shadows and motion as vehicle cues have been modified, implemented and evaluated. This work presents a combination of a multiscale edge based detection and a shadow based detection as the most promising algorithm, with a positive detection rate of 96.4% on vehicles at a distance of between 5 m to 30 m. For the algorithm to work in a complete system for vehicle detection, future work should be focused on developing a vehicle classifier to reject false detections.

The PRESTO sequence is a well-known 3-D fMRI imaging sequence. In this sequence the echo planar imaging technique is merged with the echo-shift technique. This combination results in a very fast image acquisition, which is required for fMRI examinations of neural activation in the human brain. The aim of this work was to use the basic Cartesian PRESTO sequence as a framework when developing a novel trajectory using a non-Cartesian grid.

Our new pulse sequence, PRESTO CAN, rotates the k-space profiles around the ky-axis in a non-Cartesian manner. This results in a high sampling density close to the centre of the k-space, and at the same time it provides sparser data collection of the part of the k-space that contains less useful information. This "can- or cylinder-like" pattern is expected to result in a much faster k-space acquisition without loosing important spatial information.

A new reconstruction algorithm was also developed. The purpose was to be able to construct an image volume from data obtained using the novel PRESTO CAN sequence. This reconstruction algorithm was based on the gridding technique, and a Kaiser-Bessel window was also used in order to re-sample the data onto a Cartesian grid. This was required to make 3-D Fourier transformation possible. In addition, simulations were also performed in order to verify the function of the reconstruction algorithm. Furthermore, in vitro tests showed that the development of the PRESTO CAN sequence and the corresponding reconstruction algorithm were highly successful.

In the future, the results can relatively easily be extended and generalized for in vivo investigations. In addition, there are numerous exciting possibilities for extending the basic techniques described in this thesis.

In this thesis spacetime analysis is applied to laser triangulation in an attempt to eliminate certain artifacts caused mainly by reflectance variations of the surface being measured. It is shown that spacetime analysis do eliminate these artifacts almost completely, it is also shown that the shape of the laser beam used no longer is critical thanks to the spacetime analysis, and that in some cases the laser probably even could be exchanged for a non-coherent light source. Furthermore experiments of running the derived algorithm on a GPU (Graphics Processing Unit) are conducted with very promising results.

The thesis starts by deriving the theory needed for doing spacetime analysis in a laser triangulation setup taking perspective distortions into account, then several experiments evaluating the method is conducted.

A common problem when using background models to segment moving objects from video sequences is that objects cast shadow usually significantly differ from the background and therefore get detected as foreground. This causes several problems when extracting and labeling objects, such as object shape distortion and several objects merging together. The purpose of this thesis is to explore various possibilities to handle this problem.

Three methods for statistical background modeling are reviewed. All methods work on a per pixel basis, the first is based on approximating the median, the next on using Gaussian mixture models, and the last one is based on channel representation. It is concluded that all methods detect cast shadows as foreground.

A study of existing methods to handle cast shadows has been carried out in order to gain knowledge on the subject and get ideas. A common approach is to transform the RGB-color representation into a representation that separates color into intensity and chromatic components in order to determine whether or not newly sampled pixel-values are related to the background. The color spaces HSV, IHSL, CIELAB, YCbCr, and a color model proposed in the literature (Horprasert et al.) are discussed and compared for the purpose of shadow detection. It is concluded that Horprasert's color model is the most suitable for this purpose.

The thesis ends with a proposal of a method to combine background modeling using Gaussian mixture models with shadow detection using Horprasert's color model. It is concluded that, while not perfect, such a combination can be very helpful in segmenting objects and detecting their cast shadow.

One major goal of the COSPAL project is to develop an artificial cognitive system architecture with the capability of exploratory learning. Exploratory learning is a strategy that allows to apply generalization on a conceptual level, resulting in an extension of competences. Whereas classical learning methods aim at best possible generalization, i.e., concluding from a number of samples of a problem class to the problem class itself, exploration aims at applying acquired competences to a new problem class. Incremental or online learning is an inherent requirement to perform exploratory learning.

Exploratory learning requires new theoretic tools and new algorithms. In the COSPAL project, we mainly investigate reinforcement-type learning methods for exploratory learning and in this paper we focus on its algorithmic aspect. Learning is performed in terms of four nested loops, where the outermost loop reflects the user-reinforcement-feedback loop, the intermediate two loops switch between different solution modes at symbolic respectively sub-symbolic level, and the innermost loop performs the acquired competences in terms of perception-action cycles. We present a system diagram which explains this process in more detail.

We discuss the learning strategy in terms of learning scenarios provided by the user. This interaction between user (’teacher’) and system is a major difference to most existing systems where the system designer places his world model into the system. We believe that this is the key to extendable robust system behavior and successful interaction of humans and artificial cognitive systems.

We furthermore address the issue of bootstrapping the system, and, in particular, the visual recognition module.We give some more in-depth details about our recognition method and how feedback from higher levels is implemented. The described system is however work in progress and no final results are available yet. The available preliminary results that we have achieved so far, clearly point towards a successful proof of the architecture concept.

The objective of this master thesis was to study the performance of an active triangulation system for 3-D imaging in underwater applications. Structured light from a 20 mW laser and a conventional video camera was used to collect data for generation of 3-D images. Different techniques to locate the laser line and transform it into spatial coordinates were developed and evaluated. A field- and a laboratory trial were performed.

From the trials we can conclude that the distance resolution is much higher than the lateral- and longitudinal resolution. The lateral resolution can be improved either by using a high frame rate camera or simply by using a low scanning speed. It is possible to obtain a range resolution of less than a millimeter. The maximum range of vision was 5 meters under water measured on a white target and 3 meters for a black target in clear sea water. These results are however dependent on environmental and system parameters such as laser power, laser beam divergence and water turbidity. A higher laser power would for example increase the maximum range.

This report describes a method to detect and recognize objects from 3D laser radar data. The method is based on local descriptors computed from triplets of planes that are estimated from the data set. Each descriptor that is computed on query data is compared with descriptors computed on object model data to get a hypothesis of object class and pose. An hypothesis is either verified or rejected using a similarity measure between the model data set and the query data set.

This report develops a method for probabilistic conceptual sensor modeling. The idea is to generate probabilities for detection, recognition and identification based on a few simple factors. The

focus lies on FLIR sensors and thermal radiation, even if discussions of other wavelength bands are made. The model can be used as a hole or some or several parts can be used to create a simpler model. The core of the model is based on the Johnson criteria that uses resolution as the input parameter. Some extensions that models other factors are also implemented. In the end a short discussion of the possibility to use this model for other sensors than FLIR is made.

This report adresses the problem of software correction of spatially variant blur in digital images. The problem arises when the camera optics contains flaws, when the scene contains multiple moving objects with different relative motion or the camera itself is i.e. rotated. Compensation through deconvolving is impossible due to the shift-variance in the PSF hence alternative methods are required. There are a number of suggested methods published. This report evaluates two methods

This report introduces a robust contour descriptor for view-based object recognition. In recent years great progress has been made in the field of view based object recognition mainly due to the introduction of texture based features such as SIFT and MSER. Although these are remarkably successful for textured objects, they have problems with man-made objects with little or no texture. For such objects, either explicit geometrical models, or contour and shading based features are also needed. This report introduces a robust contour descriptor which we hope can be combined with texture based features to obtain object recognition systems that work in a wider range of situations. Each detected contour is described as a sequence of line and ellipse segments, both which have well defined geometrical transformations to other views. The feature detector is also quite fast, this is mainly due to the idea of first detecting chains of contour points, these chains are then split into line segments, which are later either kept, grouped into ellipses or discarded. We demonstrate the robustness of the feature detector with a repeatability test under general homography transformations of a planar scene. Through the repeatability test, we find that using ellipse segments instead of lines, where this is appropriate improves repeatability. We also apply the features in a robotic setting where object appearances are learned by manipulating the objects.

The MATLAB/C program take version 3.1 is a program for simulation of X-ray projections from 3D volume data. It is based on an older C version by Muller-Merbach as well as an extended C version by Turbell. The program can simulate 2D X-ray projections from 3D objects. These data can then be input to 3D reconstruction algorithms. Here however, we only demonstrate a couple of 2D reconstruction algorithms, written in MATLAB. Simple MATLAB examples show how to generate the take projections followed by subsequent reconstruction. Compared to the old take version, the C code have been carefully revised. A preliminary, rather untested feature of using a polychromatic X-ray source with different energy levels was already included in the old take version. The current polychromatic feature X-ray is however carefully tested. For example, it has been compared with the results from the program described by Malusek et al. We also demonstrate experiments with a polychromatic X-ray source and a Plexiglass object giving the beam-hardening artefact. Detector sensitivity for different energy levels is not included in take. However, in section~\refsec:realexperiment, we describe a technique to include the detector sensitivity into the energy spectrum. Finally, an experiment with comparison of real and simulated data were performed. The result wasn't completely successful, but we still demonstrate it. Contemporary analytical reconstruction methods for helical cone-beam CT have to be designed to handle the Long Object Problem. Normally, a moderate amount of over-scanning is sufficient for reconstruction of a certain Region-of-interest (ROI). Unfortunately, for iterative methods, it seems that the useful ROI will diminish for every iteration step. The remedies proposed here are twofold. Firstly, we use careful extrapolation and masking of projection data. Secondly, we generate and utilize projection data from incompletely reconstructed volume parts, which is rather counter-intuitive and contradictory to our initial assumptions. The results seem very encouraging. Even voxels close to the boundary in the original ROI are as well enhanced by the iterative loop as the middle part.

To improve the control of a steel casting process ABB has developed an Electro Magnetic Brake (EMBR). This product is designed to improve steel quality, i.e. reduce non-metallic inclusions and blisters as well as risk of surface cracks. There is a demand of increasing the steel quality and in order to optimize the steel casting, simulations and experiments play an important role in achieving this. An advanced CFD simulation model has been created to carry out this task.

The validation of the simulation model is performed on a water model that has been built for this purpose. This water model also makes experiments possible. One step to the simulation model is to measure the velocity and motion pattern of the seeding particles and the air bubbles in the water model to see if it corresponds to the simulation results.

Since the water is transparent, seeding particles have been added to the liquid in order to observe the motion of the water. They have the same density as water. Hence the particles will follow the flow accurately. The motions of the air bubbles that are added into the water model need also to be observed since they influence the flow pattern.

An algorithm - ”Transparent motions” - is thoroughly inspected and implemented. ”Transparent motions” was originally designed to post process x-ray images. However in this thesis, it is investigated whether the algorithm might be applicable to the water model and the image sequences containing seeding particles and air bubbles that are going to be used for motion estimation.

The result show satisfying results for image sequences of particles only, however with a camera with a faster sampling interval, these results would improve. For image sequences with both bubbles and particles no results have been achieved.

Detta examensarbete utreder avståndsbedömning med hjälp av bildbehandling och stereoseende för känd kamerauppställning.

Idag existerar ett stort antal beräkningsmetoder för att få ut avstånd till objekt, men metodernas prestanda har knappt mätts. Detta arbete tittar huvudsakligen på olika blockbaserade metoder för avståndsbedömning och tittar på möjligheter samt begränsningar då man använder sig av känd kunskap inom bildbehandling och stereoseende för avståndsbedömning. Arbetet är gjort på Bofors Defence AB i Karlskoga, Sverige, i syfte att slutligen användas i ett optiskt sensorsystem. Arbetet utreder beprövade

Resultaten pekar mot att det är svårt att bestämma en närmask, avstånd till samtliga synliga objekt, men de testade metoderna bör ändå kunna användas punktvis för att beräkna avstånd. Den bästa metoden bygger på att man beräknar minsta absolutfelet och enbart behåller de säkraste värdena.

Just how far is it possible to make learning of new parts for recognition and robot picking autonomous? This thesis initially gives the prerequisites for the steps in learning and calibration that are to be automated. Among these tasks are to select a suitable part model from numerous candidates with the help of a new part segmenter, as well as computing the spatial extent of this part, facilitating robotic collision handling. Other tasks are to analyze the part model in order to highlight correct and suitable edge segments for increasing pattern matching certainty, and to choose appropriate acceptance levels for pattern matching. Furthermore, tasks deal with simplifying camera calibration by analyzing the calibration pattern, as well as compensating for differences in perspective at great depth variations, by calculating the centre of perspective of the image. The image processing algorithms created in order to solve the tasks are described and evaluated thoroughly. This thesis shows that simplification of steps of learning and calibration, by the help of advanced image processing, really is possible.

Den här rapporten beskriver och utvärderar ett antal algoritmer för multisensordatafusion av radar och IR/TV-data på rådatanivå. Med rådatafusion menas att fusionen ska ske innan attribut- eller objektextrahering. Attributextrahering kan medföra att information går förlorad som skulle kunna förbättra fusionen. Om fusionen sker på rådatanivå finns mer information tillgänglig och skulle kunna leda till en förbättrad attributextrahering i ett senare steg. Två tillvägagångssätt presenteras. Den ena metoden projicerar radarbilden till IR-vyn och vice versa. Fusionen utförs sedan på de par av bilder med samma dimensioner. Den andra metoden fusionerar de två ursprungliga bilderna till en volym. Volymen spänns upp av de tre dimensionerna representerade i ursprungsbilderna. Metoden utökas också genom att utnyttja stereoseende. Resultaten visar att det kan vara givande att utnyttja stereoseende då den extra informationen underlättar fusionen samt ger en mer generell lösning på problemet.