An Integrated Approach to Mask Wearing Classification and Crowd Counting in Public Spaces Using YOLOv5s and Deep SORT

Abstract

For safety and health reasons, we often need to monitor the flow of people in some public places. Object occlusion is a long-term, challenging problem for counting people based on the whole body of a pedestrian. In addition, counting approaches based on face object detection may perform poorly because facemasks obscure some important facial features. Thus, we propose to integrate face object detection based on mask wearing classification and people flow estimation. Through our research, we have identified YOLOv5s and Deep SORT as the optimal combination for this integration among various alternatives, and our system has been demonstrated to be effective across diverse population flow densities. Furthermore, we found out that using the face part as the tracking target performs better than using the whole body of a pedestrian for people flow estimation, especially in dense crowd cases. These findings make our approach highly feasible for real-world crowd monitoring applications, ensuring effective and reliable crowd control while considering safety and health measures.