A framework for proactive safety evaluation of intersection using surrogate safety measures and non-compliance behavior.

In recent years, identifying road users' behavior and conflicts at intersections have become an essential data source for evaluating traffic safety. According to the Federal Highway Administration (FHWA), in 2020, more than 50% of fatal and injury crashes occurred at or near intersections, necessitating further investigation. This study developed an innovative artificial intelligence (AI)-based video analytic tool to assess intersection safety using surrogate safety measures and non-compliance behavior. To extract the trajectory data, a real-time AI detection model - YOLO-v5 with a tracking framework based on the DeepSORT algorithm was deployed. 54 h of high-resolution video data were collected at six signalized intersections (including three 3-leg and three 4-leg intersections) in Glassboro, New Jersey. Non-compliance behaviors, such as redlight running and pedestrian crossing outside the crosswalk, are captured to better understand the risky behaviors at these locations. The proposed approach achieved an accuracy of 92% to 98% for detecting and tracking road users' trajectories. Additionally, the developed tool also provided directional traffic volumes, pedestrian volumes, vehicles running a red light, pedestrian crossing outside the crosswalk events, and PET and TTC for crossing conflicts between vehicles. Furthermore, an extreme value theory (EVT) was used to estimate the number of crashes at each intersection utilizing the frequency of PETs and TTCs. Finally, the intersections were ranked based on the calculated score considering the severity of crashes. Overall, the developed tool and the crash estimation, as well as the model and ranking method, can provide valuable information for engineers and policymakers to assess the safety of intersections and implement effective countermeasures to mitigate intersection-involved crashes.