Unified Lightweight Traffic Police Action Recognition Network for Edge Device

This paper introduces a unified lightweight network for recognizing traffic police gestures, designed specifically for use in autonomous vehicles and optimized for edge devices. Autonomous vehicles need to understand traffic police signals, particularly in situations where traffic lights aren't sufficient. Our method combines visual appearance data with skeleton keypoint information to improve recognition accuracy. We evaluated the Korean Traffic Police Action Recognition dataset, our model outperforms leading methods like SlowFast and ResNet with LSTM, achieving 97.52% accuracy and a 97.53 F1 score. It also offers real-time performance on both GPUs and edge devices.