UC San Diego

This dissertation addresses the critical role of perception in autonomous navigation and driving systems, proposing a paradigm shift by focusing on radar as the primary sensor. Traditional light-based sensors, such as cameras and lidars, face challenges in adverse weather conditions, impacting the reliability of autonomous systems. The research explores four key aspects of radar-based sensing: high-quality perception, semantic understanding, contextual information, and scalable deployment.

The first aspect introduces a multi-radar system that mitigates specularity and sparsity issues in radar point clouds, enhancing perception quality. Spatial averaging techniques filter out noise, and a deep learning system accurately converts radar data into bounding boxes for objects. Semantic understanding is tackled by a novel sensor fusion system, integrating radar point clouds with missing texture and color information from cameras. A unique fusion philosophy breaks feature dependence, ensuring robust performance even in adverse conditions.

Recognizing the importance of environmental context, the dissertation addresses challenges in identifying traffic infrastructure using radar alone. Specialized millimeter-wave backscatter tags, affixed to infrastructure, enhance radar visibility. The Van-Atta array amplifies reflections, RF switches modulate backscattered signals, and unique gold codes enable precise localization without modifying radar hardware.

To facilitate deployment, an open-source radar simulation framework is introduced, simulating high-fidelity Multiple Input Multiple Output (MIMO) radar data using lidar point clouds and camera images. This framework allows researchers to efficiently evaluate algorithms with effectiveness comparable to real-world data, enabling rapid iterations across the radar parameter space.

The comprehensive exploration of multi-radar systems, radar-camera fusion, smart infrastructure augmentation, and the innovative radar simulation framework contributes to advancing autonomous systems, particularly in challenging environmental conditions. The dissertation aims to overcome current limitations and chart a path for future breakthroughs in autonomous perception technologies.