Toward Automotive Safety and Autonomy with Machine-Learning-Assisted Radar Perception

Recent years have seen a surge of interest in active automotive safety and autonomous driving systems. Such systems rely on robust and affordable perception platforms, requiring advanced processing techniques that derive higher-level environment descriptions. Increasingly, such processing is performed through data-driven machine learning methods. Radar especially has the potential to benefit greatly from such methods. It is a comparatively cheap sensor and resilient to adverse weather, but provides a set of challenges related to the fidelity of its data. Deep artificial neural networks excel at the non-trivial processing such data requires. This thesis contributes innovations in the space of neural-network-driven automotive radar perception. The work presented herein develops mainly two complementary perception systems operating on data from likewise complementary sets of radar sensors. The first system performs short-range surround-view detection and tracking of dynamic objects. Its final iteration is capable of adaptively memorising and contextualising relevant information from the raw radar data, time-continuously predicting the near-future trajectory of close-by vehicles, and quantifying the system uncertainty. It achieves this using a novel architecture for a recurrent convolutional pyramid network. The second system performs medium-range occupancy segmentation of the three-dimensional static environment. It is capable of inferring the geometry of scene elements from highly incomplete raw data and mapping the location of traffic infrastructure such as the road and overhead objects. This is achieved using a novel scheme for the machine-learning-assisted refinement of occupancy grid maps.