Devising countermeasures for improving pedestrian safety, especially targeted measures for severe and non-severe crashes, is crucial for road authorities. However, such efforts predominantly rely on police-reported crash data, facing obvious and ethical issues and hindering proactive safety management.

While computer vision techniques offer high-resolution trajectory data of road users, the fundamental research question is, "Could we estimate pedestrian crashes and their severity without actually using crash data?" To answer this question, researchers at Queensland University of Technology, Australia, collected large video data of pedestrian movements at signalized intersections in Brisbane, Queensland, Australia.

They published their study in Communications in Transportation Research.

"We developed a hybrid model for estimating pedestrian crash frequency by severity levels to investigate the determinants of pedestrian crashes. Using machine learning, extreme vehicle-pedestrian interactions are identified and modeled through extreme value theory considering the severe and non-severe nature of a crash," says Fizza Hussain, a researcher at the School of Civil and Environmental Engineering, Queensland University of Technology.

Estimating pedestrian crashes with severity

In this study, the research team observed exceptional performance of the developed model in estimating pedestrian crash frequency by severity levels. For instance, the five-year observed mean severe and non-crashes were two and 29, respectively, and the corresponding predictions by the best-fitted model were 2.91 and 30.91, respectively.

"In the past, we needed to rely on crash statistics from three to five years to understand the crash risk level of a transport facility. The finding of this study provides us with evidence that we can now accurately predict crash risks of transport facilities just by observing the traffic movement for a week or so," Prof Shimul (Md Mazharul) Haque, a Professor of Transport Safety, says.

Increasingly, road authorities are interested in predicting crash frequency by severity levels to devise tailored countermeasures. For instance, at signalized intersections, the proposed modeling results will provide insights into crash occurrences along with severity, facilitating road authorities to prioritize their actions according to severity level.

Role of machine learning in estimating pedestrian crash frequency by severity

Machine learning has been gaining prominence, and its usage in estimating crash frequencies from traffic conflicts is rather scant. This research demonstrates that when using machine learning to identify risky pedestrian interactions, the performance of crash risk prediction models increases by about three times compared to using conventional (non-machine learning methods).

"These results suggest the superiority of applying machine learning in estimating pedestrian crash frequency by severity levels. We hope this research could lay a strong foundation for future applications of machine learning in such vital scenarios of pedestrian safety and developing countermeasures," says Yuefeng Li, a Professor of Computer Science.

More information: Fizza Hussain et al, Integrating machine learning and extreme value theory for estimating crash frequency-by-severity via AI-based video analytics, Communications in Transportation Research (2024). DOI: 10.1016/j.commtr.2024.100147

Provided by Tsinghua University Press