Deep Learning based Customer Count/Flow Monitoring System for Social Distancing

Despite the COVID-19 vaccination drives, use of preventative measures such as masks and social distancing are still deemed essential. This paper presents an application that will allow businesses/enterprises to monitor the flow of customers by detecting people as objects, counting the number of people, tracking the safe distance between them to maintain the two-meter distance norm. The proposed solution is set up to generate an alarm when the customers reach the allowed limit as per shop dimensions or overcrowding is detected. For the implementation, YOLOv4 and YOLOv3-Tiny were used for the task of object detection and transfer learning is used to set up weights. The models were evaluated using MSCOCO API with 100 image instances per class. The results of the YOLOv4 model are also compared with YOLOv3-Tiny in terms of calculating mean, average precision (AP), frames per second (FPS), and identification of groups (crowd). Experimental results (on several video clips from a shopping center CCTV) show that the YOLOv3-Tiny maintains real-time performance even on modest hardware. It is further demonstrated that if a high-end GPU is available, the overall detection of objects and cluster identification is much more accurate and clearer using YOLOv4.