A People Counting System for Use in CCTV Cameras in Retail

This paper focuses on the feasibility of implementing a vision-based people counting system using footage from an existing surveillance camera in a restaurant establishment. The main challenge is to do so given the unique fixed viewpoint of the camera, which is optimized for security instead of data analytics. A three-step approach, namely people detection, tracking, and then people counting, is employed in creating the system. Neural networks such as YOLOv3 and Deep SORT are used. The proponents then partnered with a retail establishment in a high-traffic business district, to test the system. The results show that it is possible to achieve an accuracy of 82.76% for days when the restaurant waiting area is not crowded. The system also achieved an overall accuracy of 66.17% over five days of extensive testing, which includes extreme conditions wherein people in the video are densely packed and occluded. However, the system performance and accuracy can still be improved through downsizing the frames, retraining the models, and exploring other models.