Estimating distance between pedestrians from street view images using geometric properties

Abstract

A National Science Foundation (NSF)-supported Rapid Response Research (RAPID) project was granted in the early months of the COVID-19 lockdown for the purposes of tracking Seattle, Washington through the progression of the pandemic. Street view data, reminiscent of Google street view, is generated via a RAPID vehicle with a mounted 360-degree camera as it tours a predetermined route through several locations in Seattle. One major aspect that many locations enacted during the lockdown was the recommendation for individuals to practice social distancing. While there have been studies that look at metrics to quantify social distance compliance, these findings are indirectly measured, focus on indoor environments, and are generally inconclusive to characterize social distance adherence on an individual basis. This unresolved research question, along with the data generated by the RAPID vehicle, presents an opportunity to develop a routine that directly estimates the social distance between pedestrians in outdoor settings. Some approaches to estimating distance between objects from images rely on specialized camera equipment, computationally complex frameworks like Structure from Motion, or deep learning networks that require three dimensional training metadata. These methods either exceed equipment limitations of the RAPID vehicle or are not scalable to the large image dataset being generated. Therefore, I propose an approach that processes 360-degree video data into distortion-free images, utilizes a state-of-the-art pedestrian detection algorithm, and develops a geometric social distance estimation method. A physical experiment is designed to generate ground-truth data in order to optimize the social distance estimation method and evaluate its performance. This yield a test root mean square error (RMSE) of 1.13 ft, which represents the error when estimating distances between pedestrians. A 95% confidence interval, constructed via bootstrapping, for the true RMSE is determined to be (1:07; 1:41). Furthermore, because of the computational efficiency of the proposed method for estimating distances using geometric properties, the number of required arithmetic operations scale in O(m2), which is the lower bound and thus optimal for the number of pedestrians m found in any given image. These results summarize an estimation that can be applied to research regarding the estimation of distance between pedestrians when the reported tolerances are acceptable. Such research could include tracking social distance compliance over time, determining size of social groups, or even non-COVID related work like evaluating how bike friendly a city is.