Pedestrian Detection and Tracking with Deep Mutual Learning

Abstract

In the last decade, the application of pedestrian detection in computer vision has gradually increased, such as social distance detection in the epidemic era. In this paper, we improve the newly proposed YOLOv5 model, use the idea of deep mutual learning for training, compare the performance and accuracy of different parameters, and select a relatively good model. As for the application, after detecting an abnormal pedestrian or a designated pedestrian, we use the Deep SORT method to track the pedestrian via the pedestrian's ID. Experimental analysis shows that our model performs well in terms of mean average precision (mAP), total loss (TL), and frames per second (FPS).