aceEfficient video-based traffic conflict prediction and interpretable risk analysis at signalized intersections via deep learning
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 th…
"Motorway Traffic - geograph.org.uk - 1464292" by Maxwell Hamilton is licensed under CC BY-SA 2.0. To view a copy of this license, visit https://creativecommons.org/licenses/by-sa/2.0/.
Researchers built a multi-stage video-based framework for signalized intersections that combines YOLOv8 detection with OC-SORT tracking to extract vehicle trajectories, uses a dynamic scaling Conflict Region of Interest to reduce data volume, and predicts conflicts with a Spatio-Temporal Graph Attention Network followed by causal forest interpretation. Tested on field video from an intersection in Nanning, China, the ST-GAT model outperformed existing deep learning architectures on training and testing sets.
The work matters because it moves intersection safety from reactive crash analysis to proactive, real-time conflict prediction that can inform urban traffic management and accident mitigation. What remains uncertain is how well the approach transfers beyond the single study site, under different weather, occlusion, traffic composition, and signal timing conditions, and how it would perform in live deployment.
- Developed real-time video analysis system integrating YOLOv8 object detection and OC-SORT multi-vehicle tracking to extract high-precision vehicle trajectory data.
- Introduced dynamic scaling Conflict Region of Interest (CROI) strategy to reduce data volume and suppress non-essential regions, improving training efficiency and prediction accuracy.
- Built Spatio-Temporal Graph Attention Network (ST-GAT) that captures fine-grained spatiotemporal dependencies across lanes and video frames.
- Applied causal forest analysis to ST-GAT outputs to interpret influence of traffic factors on conflict frequency.
A multi-stage framework integrating YOLOv8 detection, OC-SORT tracking, dynamic CROI filtering, and ST-GAT prediction enables accurate real-time traffic-conflict prediction at signalized intersections to enhance safety and mitigate accident risks.
The rundown
The system first extracts trajectories using YOLOv8 and OC-SORT, explicitly designed to overcome environmental occlusions, then applies a dynamic scaling CROI to filter irrelevant areas before training the predictor.
In the Nanning case study, CROI balanced training efficiency and accuracy, and ST-GAT outperformed existing deep learning architectures on both training and testing sets for conflict-risk identification.
Interpretability analysis found 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.
Sources
- Peer-reviewedTraffic Injury Prevention2026-08-13
How should this claim be treated?
The debate