Roy, SumitGao, Xiangyu2023-09-272023-09-272023-09-272023Gao_washington_0250E_26072.pdfhttp://hdl.handle.net/1773/50783Thesis (Ph.D.)--University of Washington, 2023Millimeter-wave radars are increasingly integrated into commercial vehicles to support advanced driver-assistance systems, enabling robust object detection, localization, and recognition as a crucial component of environmental perception in autonomous driving systems. This thesis focuses on radar perception algorithm design, incorporating fundamental signal processing, and novel deep learning applications to address open challenges observed in autonomous driving. To tackle challenging conditions for autonomous driving, where optical sensing may be limited, we propose a novel radar multiple perspectives convolutional neural network (RAMP-CNN). This model extracts object location and class information from range-velocity-angle heatmap sequences. To reduce complexity, we combine lower-dimension network models within our RAMP-CNN, achieving significant performance enhancement. Experimental results demonstrate superior average recall and average precision compared to prior works in all testing scenarios. Notably, the RAMP-CNN model exhibits robust performance during nighttime, showcasing potential for low-cost radars as substitutes for optical sensing under adverse conditions. Current vehicular radar imaging suffers from poor azimuth resolution for side-looking operation due to antenna size constraints. To address this limitation, we propose a multiple-input and multiple-output synthetic aperture radar (MIMO-SAR) imaging technique. By applying coherent SAR principles to vehicular MIMO radar, we enhance the side-view angular resolution. The proposed MIMO-SAR algorithm employs a 2-stage hierarchical workflow, significantly reducing computation load while preserving image resolution. Coherent processing over the synthetic aperture is enabled by integrating a radar odometry algorithm to estimate the trajectory of the ego radar. Validation of the MIMO-SAR algorithm is conducted through simulations and real experiment data collected from a vehicle-mounted radar platform. Anti-collision assistance (as part of the current push towards increasing vehicular autonomy) critically depends on accurate detection/localization of {\em moving targets} in vicinity. An effective solution pathway involves removing background or static objects from the scene, so as to enhance the detection/localization of moving targets as a key component for improving overall system performance. We present an efficient algorithm for background removal for automotive scenarios, applicable to commodity frequency-modulated continuous wave (FMCW)-based radars. Our proposed algorithm follows a three-step approach: a) preprocessing of back-scattered received radar signal for 4-dimensional (4D) point clouds generation, b) 3-dimensional (3D) radar ego-motion estimation, and c) notch filter-based background removal in the azimuth-elevation-Doppler domain. The performance of our algorithm is evaluated using both simulated data and experiments with real-world data. By offering a fast and computationally efficient solution, our approach contributes to a potential pathway for challenges posed by non-homogeneous environments and real-time processing requirements. Overall, this thesis contributes to the advancement of autonomous driving systems by introducing efficient and enhanced radar perception techniques. The proposed algorithms address critical challenges, paving the way for safer and more reliable autonomous vehicles in diverse and complex driving environments.application/pdfen-USCC BY-NC-SAautonomous driving systemsmachine learningmillimeter-wave radarradar perceptionsignal processingElectrical engineeringArtificial intelligenceCivil engineeringElectrical and computer engineeringThesis