Teacher-Student Knowledge Distillation for Radar Perception on Embedded Accelerators

With the steady advances in autonomous driving, advanced safety features using one or more sensors are highly desirable. In order to avoid collisions and unintended breaking maneuvers, it is crucial to detect potential road obstacles accurately. Although camera and LiDAR-based object detection have been studied in the literature [1, 2], it's only recently that interest in radar-based object detection using ML methods has begun, primarily because of its low cost, long-range detection capability, and robustness to poor weather conditions. Traditionally, automotive radar-based object detection is performed through peak detection using simple local thresholding methods such as the Constant False-Alarm Rate (CFAR) algorithm [3]. With the breakthroughs of ML in numerous applications [4, 5, 6, 7], radar-based object perception using ML has attracted attention [8, 9, 10, 11, 12, 13].