Edge AI for Automotive Vulnerable Road User Safety: Deployable Detection via Knowledge Distillation

Researchers Akshay Karjol and Darrin M. Hanna from Oakland University have published a paper demonstrating that knowledge distillation (KD) is essential for deploying accurate, safety-critical object detection for Vulnerable Road Users (VRUs) on edge hardware. Their framework trains a compact YOLOv8-S student model (11.2 million parameters) to mimic a YOLOv8-L teacher model (43.7 million parameters), achieving a 3.9x compression while preserving quantization robustness. Evaluated on the full-scale BDD100K dataset (70,000 training images) with Post-Training Quantization to INT8, the teacher model suffers catastrophic degradation (-23% mAP), while the KD student retains accuracy with only -5.6% mAP loss.

Further analysis reveals that KD transfers precision calibration rather than raw detection capacity. At INT8, the KD student achieves 0.748 precision versus 0.653 for direct training, a 14.5% gain at equivalent recall, and reduces false alarms by 44% compared to the collapsed teacher. Remarkably, the INT8 KD student exceeds the teacher's FP32 precision (0.748 vs. 0.718) in a model 3.9x smaller. These findings establish knowledge distillation as a requirement for deploying accurate, safety-critical VRU detection on edge hardware, enabling real-time pedestrian and cyclist detection in autonomous vehicles and driver-assistance systems without sacrificing performance.