A Conflict Warning Approach Based on Roadside Perception-Driven Trajectory Prediction

Abstract

Highway traffic crashes primarily originate from delayed responses to dynamic traffic conflicts. Implementing real-time proactive conflict prediction mechanisms is crucial for collision warning systems and safety applications. While vehicle-centric sensing is constrained by limited fields of view and occlusion vulnerabilities, and approaches based on unmanned aerial vehicle (UAV) trajectories face operational reliability challenges in severe weather and lighting conditions. Roadside multisensor fusion systems overcome these limitations through persistent spatiotemporal observation. To this end, this paper proposes a conflict warning approach based on roadside perception-driven trajectory prediction. First, we propose a novel trajectory prediction architecture that processes variable-length roadside perception trajectories while jointly encoding road topology constraints and multiagent interaction dynamics. Subsequently, we introduce an adaptive conflict detection framework leveraging vehicle proximity distribution to model probability of near-collision event occurrence, learning context-dependent proximity thresholds from trajectory data. These contributions are then integrated into a conflict warning system that demonstrates robustness across diverse highway interaction scenarios, validated against unified trajectory dataset and co-simulated environments. Our warning method achieved a true positive rate of 95.52% and a false alarm rate of 3.03%, and demonstrates early-warning temporal stability.