YOLOv3 based Real Time Social Distance Violation Detection in Public Places

Abstract

The prevalent COVID 19 pandemic is incessantly taking toll on the lives of people throughout the world. Moreover, the dearth of effectual remedies has caused an expeditious rise in the total COVID 19 cases. Though vaccines have been developed, the enormous task of vaccinating a large population is still challenging. Also, as new variants emanate, the resilience from infections conceivably decreases. Hence, it’s most unlikely that we’ll achieve herd immunity globally so soon. Thus, since the transmission of COVID causing coronavirus roots mainly to social proximity between people, it is necessary to stringently comply to the non pharmaceutical preventive measures of wearing masks and maintaining physical distancing. Howbeit, it has evidently been found that people are being lethargically ignorant to the social distancing norms with passing time. Hence, an autonomous mechanism intended at social distancing violation detection through monitoring of people is needed to be introduced at an authority level. In this paper, the implementation of YOLO Object detection transfer learning process has been used for accomplishing this aim of real time detection of social distancing violation. Our social distance prediction approach uses a pre-trained YOLOv3 object tracking algorithm for identifying people in an input video stream. A Distance estimation algorithm is further used, that works by computing euclidean distance between the centroids of each pair of detected people. This approach highlights the people violating the social distancing criteria as well as calculates the number of times social distancing gets violated as any two people get closer than a set threshold value of minimum permissible distance. A number of experiments on various pre-recorded video streams has been conducted in order to estimate the viability of this method. Through experimental outcomes, it has been found that this YOLO based object detection method with the proposed social distance prediction algorithm produces favourable results for tracking social distancing in public spaces.