COVID-19 Social Distance Surveillance Using Deep Learning

Abstract

COVID-19 disease discovered from the novel corona virus can spread through close contact with a COVID-19 infected person. One of the measures advised to contain the spread of the virus is to maintain social distancing by minimizing contact between potentially infected individuals and healthy individuals or between population groups with high rates of transmission and population groups with no or low-levels of transmission. Motivated by this practice, we propose a deep learning framework for social distance detection and monitoring using surveillance video that can aid in reducing the impact of COVID-19 pandemic. This work utilizes YOLO, Detectron2 and DETR pre-trained models for detecting humans in a video frame to obtain bounding boxes and their coordinates. Bottom-centre points of the boxes were determined and were then transformed to top-down view for accurate measurement of distances between the detected humans. Based on the depth of each bottom-centre point estimated using monodepth2, dynamic distance between pairs of bounding boxes and corresponding distance threshold (safe distance) to prevent violation of social distancing norm were computed. Bounding boxes which violate the distance threshold were categorized as unsafe. All the experiments were conducted on publicly available Oxford Town Center, PETS2009 and VIRAT dataset. Results showed that Detectron2 with top-down view transformation and distance thresholding using pixel depth estimation outperformed other state-of-the-art models. The major contribution of this work is the estimation and integration of variable depth information in obtaining the distance threshold for evaluating social distances between humans in videos.