Efficient video-based traffic conflict prediction and interpretable risk analysis at signalized intersections via deep learning.
TL;DR
The research results provide an efficient and accurate traffic-conflict prediction framework at signalized intersections, while identifying key contributing factors influencing their occurrence, thereby providing a decision support for enhancing urban intersection safety and mitigating accident risks.
Abstract
Objectives
To achieve accurate and real-time prediction of traffic conflicts at signalized intersections and identify their key contributing factors, thereby supporting proactive safety management and reducing accident risks.
Methods
This study proposes a novel multi-stage traffic-conflict prediction framework that integrates a real-time video image processing system and an advanced conflict-prediction model. Specifically, a real-time video analysis system integrating the YOLOv8 object detection framework and the OC-SORT algorithm for multi-vehicle tracking is first developed to extract key vehicle trajectory data collections. This joint approach effectively overcomes environmental occlusions, enabling the automatic extraction of high-precision vehicle trajectory data. By incorporating a dynamic scaling Conflict Region of Interest (CROI) strategy, the system effectively reduces the overall data volume and suppresses disturbances from non-essential regions, which is beneficial to improving the training efficiency and prediction accuracy of the conflict-prediction model. Furthermore, a Spatio-Temporal Graph Attention Network (ST-GAT)-based conflict-prediction model is developed, in which the graph attention mechanism is adopted to capture fine-grained spatiotemporal dependencies across lanes and video frames, improving the potential conflict detection. Finally, a causal forest analysis is applied to the ST-GAT outputs to interpret the influence of critical traffic factors on conflict frequency, providing an intuitive and interpretable characterization of their impacts.
Results
A case study using field video data from a representative signalized intersection in Nanning, China, shows that the CROI strikes an effective balance between conflict-prediction model training efficiency and prediction accuracy. On both the training and testing sets, the ST-GAT model outperforms existing deep learning architectures in terms of both accuracy and robustness for conflict-risk identification. Furthermore, the interpretability analysis indicates that an increase in mainline traffic volume is strongly associated with amplified conflict risk, with sensitivity modulated by opposing mainline flows, mainline speeds, and merging maneuvers from the minor approach.
Conclusions
The research results provide an efficient and accurate traffic-conflict prediction framework at signalized intersections, while identifying key contributing factors influencing their occurrence, thereby providing a decision support for enhancing urban intersection safety and mitigating accident risks.