Deep Learning for Visual Analytics of the Spread of COVID-19 Infection in Crowded Urban Environments

Abstract

The novel SARS-CoV-2 coronavirus caused a global pandemic in 2020 with millions of diagnosed cases and a staggering number of deaths. As a preventive measure, many governments issued social distancing and shelter-in-place mandates to limit human contact and slow the rate of infection. The large extent and duration of the crisis is poised to transform the health sector and alter current practices in retail, business, manufacturing, and construction. While medical researchers are working on antidote and vaccine solutions, contact tracing and self-isolation are deemed effective methods to control community spread. This paper presents a visual analysis approach that uses convolutional neural networks (CNNs) to generate quantifiable metrics of contact tracing. In particular, the YOLO-v3 architecture was trained on an annotated video dataset containing pedestrians. Network pruning and non-maximum suppression were applied to optimize model performance, resulting in 69.41% average precision. The fully trained model was then tested on sample crosswalk video data from Xiamen, China, recorded during the COVID-19 pandemic, followed by projecting detected pedestrians onto an orthogonal map for contact tracing by tracking movement trajectories and simulating the spread of droplets among the healthy population. Results demonstrate that the proposed technique is capable of tracing and documenting infection sources, times, and locations.