A Lightweight Traffic Police Action Recognition Deep Learning Network for Edge Device

The traffic police gesture recognition is necessary for the perception task of autonomous vehicles. Most of the previous SOTA algorithms are considering accuracy instead of network capability to run edge devices of autonomous vehicles. Since the hardware resources of autonomous vehicles are limited, we propose a lightweight model that can recognize the action of traffic police. We compared our proposed network with SOTA methods and experimented with an open dataset. The results show that the proposed method archives an accuracy rate of 100% same as the SOTA method. And inference speed is faster than the SOTA method and higher than real-time requirements at RTX3090(150 fps) and 33.3 fps at the NVIDIA Jetson Xavior NX board.