Fighting Detection Based on Pedestrian Pose Estimation

Huadong Pan, Jun Yin, Haohua Ku, Cheng Liu, Fubiao Feng, Jia Zheng, Shixian Luo

Abstract: Recently, fighting detection in videos is almost based on motion optical flow. Motion optical flow can quickly locate the target of violent movement, but it will bring errors due to the rapid movement of pedestrians or the rapid change of attitude. In this paper, we add pedestrian key nodes to determine the behavior posture based on the optical detection method. Firstly, in the convolution structure, the response graph of the image is obtained, and the key nodes of the pedestrian body are obtained through the extreme of the response graph. Secondly, the limb of a pedestrian is represented by connecting the key nodes. Finally, the optical flow of the key nodes is utilized to determine the posture of the pedestrian and abnormal behavior. Some experiments can demonstrate that our method is more effective to detect the fighting behavior in videos.