A Novel AI-Powered Approach to Derive Minimum Dynamic Time-to-Collision Using Vehicles’ Key Conflicting Points: A Point-Level Framework for Traffic Conflict Analysis at Signalized Intersections

Abstract

Traffic conflict analysis at signalized intersections provides a proactive approach for evaluating roadway safety without relying solely on historical crash data. However, existing computer vision–based methods often rely on simplified vehicle representations and static conflict indicators, which limit their ability to capture continuously evolving vehicle interactions. This study proposes a traffic conflict identification framework that utilizes a deep neural network–based pose estimation algorithm to extract vehicle key points from surveillance camera footage and transform them into a plan-view representation. Using these reconstructed vehicle polygons, a new surrogate safety measure, minimum dynamic time-to-collision (mDTTC), is developed to continuously evaluate vehicle interactions at the point level. Unlike traditional time-to-collision (TTC), the proposed metric accounts for dynamic motion states, vehicle orientation, and evolving interaction geometry over time. The framework was applied to a set of conflict scenarios at signalized intersections using multicamera video data. The results demonstrated that traditional TTC frequently underestimated or misrepresented conflict severity, particularly in head-on and angle conflicts, leading to false severe conflict indications. In contrast, the proposed mDTTC metric provided more accurate conflict characterization by continuously updating vehicle dynamics and spatial relationships.

The proposed approach enhances the accuracy of traffic conflict detection and severity assessment while offering a scalable and cost-effective solution using existing surveillance infrastructure. The framework supports improved intersection safety evaluation and enables more reliable identification of safety-critical events.