Object Detection for Automotive Radar Perception

Automated vehicles are among the biggest trends in the automotive industry. The desired level of automation slowly progresses from advanced driver assistance system functions to fully autonomous driving. Excellent environmental perception is a critical requirement in this development. This thesis focuses on solutions to the challenges that come with the utilization of automotive radar systems for road user recognition. Therefore, several machine learning techniques are applied and compared to detect and classify moving road users in automotive radar point clouds. An overview of radar processing is given to provide information on how to utilize and interpret the data properly. All methods are evaluated on publicly available real-world data sets. To facilitate the creation of such data sets, a system for automating the associated labeling process is introduced. The detection and classification concepts start with classical modularized approaches that use a clustering algorithm, followed by a feature extraction stage and a conventional classifier. Several techniques that improve these traditional methods are proposed and evaluated, e.g., by utilizing recurrent neural network ensembles or advanced multi-stage clustering. Then, a transition is made from modularized concepts to more self-contained models enabled by modern end-to-end deep learning methods that combine the localization, the feature extraction, and the classification stages in a single model. The developed methods are applied in two case studies, which show how automotive radar can detect non-line-of-sight objects around corners and how next-generation radar sensors impact the accuracy of radar detection systems.