Abstract: In order to solve the problems of low detection speed and large amount of model parameters of YOLOv4 in the target detection task, an improved target detection algorithm of YOLOv4 is proposed. CSPDarknet53 of the YOLOv4 backbone is replaced by Mobilenet to improve the feature extraction network of YOLOv4, and the original standard 3×3 convolution of PANet is replaced by a depth-division convolution to reduce the computational burden, so as to improve the detection speed and reduce the model parameters. The K-means+〖KG-*3〗+ algorithm is then used to perform anchor dimension clustering on a dataset consisting of 8565 images to improve the accuracy of the algorithm. At the same time, a system for recording pedestrian wearing of masks and measuring people’s temperature is built to perform epidemic control tasks in crowded places. The FPS has been improved by 200% and the model parameters have been reduced by 82% compared with the original algorithm, while maintaining the accuracy of the YOLOv4-Mobilenet. The improved model can detect an average of 67 frames per second, which can detect mask wearing in real applications, and the results show that the model is efficient and reliable.

Key words: YOLOV4, Mobilenet, depth seperable convolution, K-means+〖KG-*3〗+, mask wearing test