Knowledge Distillation Boosted Ultra-light Weight Road Object Detection on Edge Device

Road object detection is required for automatic driving and intelligent traffic control and detecting illegal behavior. However, many edge computing devices do not have high computing power. Therefore, it is necessary to deploy a lightweight object detection model while maintaining its performance. Thus, we developed a scenario on the 24-hour automatic patrol camcorder for the corners and alleys. The device instantly identified and detected suspicious vehicles or pedestrians and sent notifications. Hence, the model had low complexity and efficiency to meet the requirement of real-time detection. We improved the model's precision within the limited inference time. In the experiment, YOLOv5 was selected as the base model for lightweight model construction on the Snapdragon 660 Mobile Platform on Xiaomi Mi A2. The designed model YOLOv5p with parameter size 0.7 M boosted knowledge distillation from mean average precision (mAP) of 18 to 19.3% on the BDD100K dataset.