A big part of COVID-19 response centers around the safe return of employees and customers in public spaces such as shops, plants, airports, cafes, etc. This requires ongoing needs to monitor compliance with physical distancing and crowding guidelines. Since a lot of public spaces already employ security cameras, we can use computer vision to monitor physical distancing in a noninvasive, anonymized manner targeted toward well-defined spaces. We have built a robust, end-to-end pipeline by combining object detection and tracking against camera feeds with simple configuration. This pipeline can support real-time alerts for urgent situations and post-facto analysis for assessing overall effectiveness based on location/day of week/time of day. This solution uses novel features to map physical distance to pixel distance and attach “persistent” conditions to alert condition to avoid false positives.
