Social Distancing Detection and Analysis through Computer Vision

The widespread COronaVIrus Disease 2019 or COVID-19 infectious disease has caused global deaths of 1.5+ million with more than 66.5 million worldwide cases as of 5 December 2020. As per WHO and various studies based on mathematical models concluded that social distancing is the most potent preventive measure against covid-19 which can help bring down the rising cases and mortality rate. With more than 50 candidate vaccines under development but none being 100% effective to cure covid-19, social distancing is the most effective approach to fight this global pandemic. Inspired by this we propose a computer vision-based social distancing detector based on the Detection Transformer object detection model and applied camera calibration and perspective transformation to make the complete system independent of the camera view-point and distortion which attains balanced mean average precision (mAP) and fps score suitable for real-time environments. Extensive and meticulous experiments were done using SOTA models; EfficientDet with EfficentNet B0 and B5 backbone and DETR with resnet-50 backbone. DETR outperformed both variants of EfficientDet and was used for real-time social distancing monitoring and analysis.