On the Comparison of Social Distancing Violation Detectors with Graphical Processing Unit Support

Social distancing or sometimes referred as physical distancing is claimed as the best spread stopper in the present COVID-19 pandemic. Social distancing monitoring by using computer vision becomes an important technological aspect in the current pandemic. This type of technology ensures automatic human object detection followed by physical distance measurement. The actual distances are measured as the number of pixels separating two centroids. The social distancing violations are known based on the measured distances. In this works, we compare three deep learning methods used for social distancing monitoring i.e YOLOv3, YOLOv3-Tiny, and MobileNetSSD. Those methods are executed with and without GPU support, and we assess the their performances in terms of speed and detection accuracies. The results show that the use of GPU significantly increases the speed of both YOLOv3 and YOLOv3-Tiny, but not for MobilenetSSD. GPU support increases about 300 % the Frame-per-Second (FPS) rate of YOLOv3 and the highest FPS rate is achieved for YOLOv3-Tiny. The results also indicate that YOLOv3 offers the best detection accuracies compared to YOLOv3-Tiny and MobilenetSSD, but in the exchange of heavy computational process.