A physics-informed risk force theory for estimating pedestrian crash risk by severity using artificial intelligence-based video analytics

Pedestrians are a vulnerable road user group, and assessing their crash risk at critical locations, such as signalized intersections, is crucial for developing targeted countermeasures. While conflict-based safety assessments using traffic conflict measures effectively estimate crash risk, they often overlook the heterogeneity of different motorized and non-motorized road users. Conversely, field-based theories account for road user heterogeneity, yet their application in crash risk assessment, specifically evaluating pedestrian crash risk, and particularly by severity level using real-world data, remains underexplored. This study introduces a novel application of physics-informed risk force theory for assessing pedestrian crash risk by injury severity, utilizing facility-based video data at signalized intersections. The study derives risk forces that encompass pedestrian and vehicle heterogeneity as a nearness-to-collision component and vehicle impact speed as a severity component. Stationary and non-stationary extreme value models, incorporating exogenous traffic parameters at the signal cycle level, were applied to 72 h of video data collected from three signalized intersections in Queensland, Australia. The non-stationary univariate extreme value model with risk force as a measure of nearness-to-collision reliably estimated total crash frequency compared to historical crash records. In addition, the bivariate extreme value model with risk force and impact speed reasonably predicted pedestrian crashes by severity levels. The results also indicate that an increased volume of interacting pedestrians and left-turning vehicles elevates the likelihood of total and severe crashes. The proposed pedestrian crash risk assessment framework offers a unified and efficient proactive approach that can enhance automated safety analysis of traffic facilities, thereby assisting road authorities in real-time safety management.