Around the world, citizens are practising physical distancing to help flatten the curve on the COVID-19 pandemic. But with many countries cautiously restarting parts of their economies, concerns have emerged regarding workplace safety in this new environment.
Last week, Silicon Valley based Landing AI introduced a new AI-enabled social distancing detection tool designed to help monitor and enforce physical distancing protocols in workplaces. With many manufacturers hastily refitting factories to produce surgical masks, hand sanitizer and other COVID-19 related materials for the general public and medical staff, it’s hoped the system can protect the workers in these and other such facilities from the threat of SARS-CoV-2 infection.
The idea is to integrate this new software into existing on-site closed-circuit television systems, where it can measure spaces between any two persons and flag pairs that are under minimum distancing requirements.
Since CCTV cameras usually shoot from one angle and the picture is two dimensional, researchers first needed to use a calibration process to transform the frame perspectives into a bird’s-eye (top-down) view for correct estimation. The system assumes everyone in frame is standing on the same flat ground, chooses four filmed ground points and automatically matches them to the corners of a top-down view. The Landing AI researchers have also included a lightweight tool that can help non-technical users with real-time mapping tasks.
The next step of the process is detection, which involves drawing a boundary box around each detected individual in the frame. For this task, the researchers applied an open-source pedestrian detection network based on Faster R-CNN architecture, along with non-max suppression (NMS) to eliminate redundant boxes and find the best fit.
The last step is measurement, which takes the location estimates for each person in the bird’s-eye view, scales the distance between each pair of boundary boxes, and depending on preset physical distancing minimums, indicates any non-compliant pair on the system display in real time with a red frame and red line.
The complete Landing AI blog post is here. Although the company stresses this not an identify detection system (“our current system does not recognize individuals, and we urge anyone using such a system to do so with transparency and only with informed consent”), many comments under company founder Andrew Ng’s Twitter post have raised privacy issues and questioned whether people really need intelligent oversight to practice physical distancing.
Author: Reina Qi Wan | Editor: Michael Sarazen