GroupNet: Detecting the Social Distancing Violation using Object Tracking in Crowdscene
Anthony Boyko, Mohamed H. Abdelpakey, Mohamed Shehata · 2021
COVID-19 affects everyone on a daily-basis causing adjustments in which society functions. One of these major adjustments is the need to measure how well people distance from each other, that is referred to as social distancing. Previous work to automate social distancing violations does not take into consideration the exceptions to minimum distance guidelines. In this paper, we propose GroupNet, a novel multi-object tracking social distancing violation detector through the addition of group detection to reduce the number of false positives that are currently missed in existing literature. We define the social distancing violation occurs when two individuals are within a specified Euclidean distance of two meters. GroupNet leverages the contextual information learned by group detection. Moreover, GroupNet uses a Joint Detection and Embedding (JDE) multi-object tracker as a backbone network for group detection. To map from pixel-wise coordinates to the real-world equivalent coordinate, a pre-processed affine matrix is used for the transformation. GroupNet determines if two individuals are a group through leveraging a re-identification component from a multi-object tracker. Location of bounding boxes are tracked over time to obtain individuals relative distance between each other. Group-Net uses regression analysis to determine if the relative distance between two individuals changes over time. The likelihood of the change in relative distance is non-zero determines group existence between individuals (i.e. group detection). Performance of GroupNet is evaluated by manual inspection of output images due to the lack of labeled ground truth data. Moreover, GroupNet through the provided experiments shows an improvement in the reduction in the number of false-positives due to group detection analysis alongside the addition of minimal false-negatives.