Personal variations severely limit the performance of appearance-based gaze tracking. Adapting to these variations using standard neural network model adaptation methods is difficult. The problems range from overfitting, due to small amounts of training data, to underfitting, due to restrictive model architectures. In this thesis, these problems are tackled by introducing the SPatial Adaptive GaZe Estimator (\spaze{}). By modeling personal variations as a low-dimensional latent parameter space, \spaze{} provides just enough adaptability to capture the range of personal variations without being prone to overfitting. Calibrating \spaze{} for a new person reduces to solving a small optimization problem. \spaze{} achieves an error of \ang{2.70} with \num{9} calibration samples on MPIIGaze, improving on the state-of-the-art by \SI{14}{\percent}. In the introductory chapters the history, methods and applications of eye tracking are reviewed, with focus on video-based eye tracking and the use of personal calibration in these methods. Emphasis is placed on methods using neural networks and the strengths and weaknesses of how these methods implement personal calibration. / <p>QC 20210528</p>

Autonomous systems have progressed from theory to application especially in the last decade, thanks to the recent technological evolution. The number of autonomous systems in our surroundings is increasing rapidly. Since these systems in most cases handle privacy-sensitive data, the privacy concerns are also increasing at a similar rate. However, privacy research has not been in sync with these developments. Moreover, the systems are heterogeneous in nature and continuously evolving which makes the privacy problem even more challenging. The domain poses some unique privacy challenges which are not always possible to solve using existing solutions from other related fields. In this thesis, we identify open privacy challenges of autonomous systems and later propose solutions to some of the most prominent challenges. We investigate the privacy challenges in the context of smart home-based systems including Ambient Assisted Living (AAL) systems as well as autonomous vehicles. In the case of smart home, we propose a framework to enhance the privacy of owners during ownership change of IoT devices and conduct a systematic literature review to identify the privacy challenges of home-based health monitoring systems. For autonomous vehicles, we quantify, improve, and tune the privacy utility trade-off of the image de-identification process. Our investigation reveals that there is a lack of consideration when it comes to the privacy of autonomous systems and there are several open research questions in the domain regarding, for instance, privacy-preserving data management, quantification of privacy utility trade-off, and compliance with privacy laws. Since the field is evolving, this work can be seen as a step towards privacy preserving autonomous systems. The identified privacy concerns and their corresponding solutions presented in this thesis will help the research community to identify and address existing privacy concerns of autonomous systems. Solving the concerns will encourage the end-users to adopt the systems and enjoy the benefits without bothering about privacy. / <p>QC 20201116</p>

No description available.

No description available.

In industrial processes, products are often visually inspected for defects inorder to verify their quality. Many automated visual inspection algorithms exist, and in many cases humans still perform the inspections. Advances in machine learning have showed that deep learning methods lie at the forefront of reliability and accuracy in such inspection tasks. In order to detect defects, most deep learning methods need large amounts of training data to learn from. This makes demonstrating such methods to a new customer problematic, since such data often does not exist beforehand, and has to be gathered specifically for the task. The aim of this thesis is to develop a method to perform such demonstrations. With access to only a small dataset, the method should be able to analyse an image and return a map of binary values, signifying which pixels in the original image belong to a defect and which do not. A method was developed that divides an image into overlapping patches, and analyses each patch individually for defects, using a deep learning method. Three different deep learning methods for classifying the patches were evaluated; a convolutional neural network, a transfer learning model based on the VGG19 network, and an autoencoder. The three methods were first compared in a simple binary classification task, without the patching method. They were then tested together with the patching method on two sets of images. The transfer learning model was able to identify every defect across both tests, having been trained using only four training images, proving that defect detection with deep learning can be done successfully even when there is not much training data available.

Digital image analysis is the field of extracting relevant information from digital images. Recent developments in imaging techniques have made 3-dimensional volume images more common. This has created a need to extend existing 2D image analysis tools to handle images of higher dimensions. Such extensions are usually not straightforward. In many cases, the theoretical and computational complexity of a problem increases dramatically when an extra dimension is added. A fundamental problem in image analysis is image segmentation, i.e., identifying and separating relevant objects and structures in an image. Accurate segmentation is often required for further processing and analysis of the image can be applied. Despite years of active research, general image segmentation is still seen as an unsolved problem. This mainly due to the fact that it is hard to identify objects from image data only. Often, some high-level knowledge about the objects in the image is needed. This high-level knowledge may be provided in different ways. For fully automatic segmentation, the high-level knowledge must be incorporated in the segmentation algorithm itself. In interactive applications, a human user may provide high-level knowledge by guiding the segmentation process in various ways. The aim of the work presented here is to develop segmentation and analysistools for volume images. To limit the scope, the focus has been on two specic capplications of volume image analysis: analysis of volume images of fibrousmaterials and interactive segmentation of medical images. The respective image analysis challenges of these two applications will be discussed. While the work has been focused on these two applications, many of the results presented here are applicable to other image analysis problems.

This thesis introduces a GPU-accelerated algorithm for path planning under kinodynamic constraints, focusing on navigation of flying vehicles within a high-resolution cost map. The algorithm operates by creating dynamically feasible initial paths, and a non-reversible parallel tempering Markov chain Monte Carlo scheme to optimize the paths while adhering to the nonholonomic kinodynamical constraints. The algorithm efficiently generates high quality dynamically feasible paths. An analysis demonstrates the algorithm's robustness, stability and scalability. The approach used for this algorithm is versatile, allowing for straightforward adaptation to different dynamic conditions and cost maps. The algorithm's applicability also extends to various path planning problems, signifying the potential advantages of GPU-accelerated algorithms in the domain of path planning.

MDU, in collaboration with several other universities, plans to join the World Robotic Sailing Championship (WRSC), where in certain sub-challenges some object detection is necessary. Such as for detecting objects such as boats, buoys, and possibly other items. Utilizing a camera system could significantly aid in these tasks, and in this research, an omnidirectional camera is proposed. This is a camera that offers a wide field of view of 360 degrees and could provide comprehensive information about the boat’s surroundings. However, these images use a spherical camera model, which projects the image on a sphere and, when saved to a 2D format, becomes very distorted. To be able to use state-of-the-art vision algorithms for object detection and tracking, this research proposes to project these images to other formats. As such, four systems using object detection and tracking are made that uses different image representation projected from the spherical images. One system uses spherical images and is used as a baseline, while the three remaining systems use some form of projection. The first is cubemap projection, which projects the spherical image to a cube and unfolds this image on a 2D plane. The two other image representations used perspective projections, which are when the spherical image is projected to small sub-images. The two image representations that used perspective projections had 4 and 8 perspective images. None of the systems ultimately performed very well but did have some advantages and disadvantages.

<p>Inefficient traffic operations have far-reaching consequences in not just travel mobility but also public health, social equity, and economic prosperity. Congestion, a key symptom of inefficient operations, can lead to increased emissions, accidents, and productivity loss. Therefore, advancements and policies in transportation engineering require careful scrutiny to not only prevent unintended adverse consequences but also capitalize on opportunities for improvement. In spite of significant efforts to enhance traffic mobility and safety, human control of vehicles remains prone to errors such as inattention, impatience, and intoxication. Connected and autonomous vehicles (CAVs) are seen as a great opportunity to address human-related inefficiencies. This dissertation focuses on connectivity and automation and investigates the synergies between technologies. First, a deep reinforcement learning based strategy is proposed herein to enable agents to address the dynamic nature of inputs in traffic environments and to capture proximal and distant information, and to facilitate learning in rapidly changing traffic. The strategy is applied to alleviate congestion at highway bottlenecks by training a small number of CAVs to cooperatively reduce congestion through deep reinforcement learning. Secondly, to address congestion at intersections, the dissertation introduces a fog-based graphic RL (FG-RL) framework. This approach allows traffic signals across multiple intersections to form a cooperative coalition, sharing information for signal timing prescriptions. Large-scale traffic signal optimization is computationally inefficient, so the proposed FG-RL approach breaks down networks into smaller fog nodes that function as local centralization points within a decentralized system Doing so allows for a bottom-up solution approach for decomposing large traffic networks. Furthermore, the dissertation pioneers the notion of using a small CAV fleet, selected from any existing shared autonomous mobility services (SAMSs) to assist city road agencies to achieve string-stable driving in locally congested urban traffic. These vehicles are dispersed throughout the network to perform their primary function of providing ride-share services. However, when they encounter severe congestion, they act cooperatively with each other to be rerouted and to undertake traffic-stabilizing maneuvers to smoothen the traffic and reduce congestion. The smoothing behavior is learned through DRL, while the rerouting is facilitated through the proposed constrained entropy-based dynamic AV routing algorithm (CEDAR).</p>

No description available.