Event-based human behavior detection

In recent years, school bullying has occurred frequently. In order to discover campus security risks in time, most colleges and universities have deployed campus video surveillance systems. The existing methods are mainly based on campus monitoring platform for behavior detection, but it is not conducive to privacy protection. Because the event camera only outputs the part of the brightness information that changes. Therefore, this paper uses event camera to capture video images. A key point detection algorithm based on event image is proposed. Image feature extraction and boundary regression in the network are optimized. Then, the optimized network is combined with the behavior recognition algorithm for detection. Specifically, event cameras are used to capture event images of human behavior and then optimize the YOLOv8 network. To achieve better image fusion, we embedded SA-Net into the backbone of YOLOv8. In order to accelerate the convergence of boundary box, IF-MPDIoU loss function based on Inner IoU and Focal IoU is used to replace the original loss function. After the key points are extracted from the YOLOv8 network, ST-GCN network is used for action recognition. The experimental results show that the detection accuracy of the key points of human bones by YOLOv8 network reaches 82.60%. Compared with the original network, the detection accuracy of the optimized model is increased by ${1 5. 1 0 \%}$. Combined with ST-GCN network, the action recognition accuracy is 59.21%. Compared with the original network, the detection accuracy of the behavior recognition model is improved by 3.65%. In addition, the behavior detection method in this paper is not only suitable for campus environment, but also can be extended to other environments.