Abstract
In order to find dangerous targets in the crowd in time and prevent dangerous events from happening, this paper proposes a crowd abnormal behavior detection method based on YOLOV4 improved algorithm. By deepening the depth of the backbone feature network, the network parameters are updated with the exponential moving average; then the Mosica is replaced with Mixup data enhancement function; finally improve YOLOV4's extraction feature pyramid (PAN) structure to realize the recognition. The average detection accuracy (MAP) result obtained by the original YOLOV4 algorithm training model is 76.43%, and MAP value obtained by the improved YOLOV4 model is 81.41%. Compared with the original algorithm, the detection accuracy rates of the three detection categories fire, smoke, and knife are increased by 4.64%, 7.61%, and 2.68%, respectively. The experimental results show that the crowd abnormal behavior detection system based on YOLOV4 improved algorithm accuracy.