RTVD-Net:An real-time violence detection method based on pre-training of human skeleton images
Peng Zhang, Weimin Lei, Xinlei Zhao, Lijia Dong, Zhaonan Lin · 2023
Video violence detection has received increasing attention in recent years. Identifying violent behavior can effectively combat criminal behavior and prevent public safety in monitoring places. In this paper, we propose a real-time violence detection network, called RTVD-Net. Firstly, we introduce the lightweight pose estimator YOLO-Pose into violence detection tasks and proposes a pre-training method based on YOLO-Pose to extract human skeleton features. Secondly, The 2DCNN model(ACTION-Net) was used to classify violent behavior in videos. We propose a Weight selection module that takes the parameters generated from bone data as a weight coefficient, and a Keyframe Weight Assignment that connects bone information with RGB features to generate the final detection results. The method proposed in this article was comprehensively experimented and ablated on the Hockey Fights dataset and RWF2000 dataset. The experimental results showed that the accuracy of RTVD-Net was superior to the most advanced methods before, and the speed could reach 33 fps, enabling real-time detection of violent behavior. In addition, RTVD-Net can also perform well in violent behavior in multi person scenarios.