Deep neural network-based physical distancing monitoring system with tensorRT optimization

During the COVID-19 pandemic, physical distancing (PD) is highly recommended to stop the transmission of the virus. PD practices are challenging due to humans' nature as social creatures and the difficulty in estimating the distance from other people. Therefore, some technological aspects are required to monitor PD practices, where one of them is computer vision-based approach. Hence, deep learning-based computer vision is utilized to automatically detect human objects in the video surveillance. In this work, we focus on the performance study of deep learning-based object detector with Tensor RT optimization for the application of physical distancing monitoring system. Deep learning-based object detection is employed to discover people in the crowd. Once the objects have been detected, then the distances between objects can be calculated to determine whether those objects violate physical distancing or not. This work presents the physical distancing monitoring system using a deep neural network. The optimization process is based on TensorRT executed on Graphical Processing Unit (GPU) and Computer Unified Device Architecture (CUDA) platform. This research evaluates the inferencing speed of the well-known object detection model You-Only-Look-Once (YOLO) run on two different Artificial Intelligence (AI) machines. Two different systems-based on Jetson platform are developed as portable devices functioning as PD monitoring stations. The results show that the inferencing speed in regard to Frame-Per-Second (FPS) increases up to 9 times of the non-optimized ones, while maintaining the detection accuracies.