Real-Time Traffic Accident Risk Assessment Based on Multimodal Sensor Data Fusion

Abstract

For intelligent transportation systems, real-time traffic accident risk assessment is crucial, particularly in complicated urban road circumstances where vehicle trajectories, roadside perception, weather, speed, and traffic flow interact dynamically. A multimodal sensor data fusion framework for real-time accident risk assessment is proposed in this research. The technique creates a single spatiotemporal representation by combining camera-based visual features, radar motion data, vehicle trajectory data, traffic flow statistics, and environmental sensing factors. Short-term motion conflict, local traffic instability, and cross-modal risk consistency are captured by a lightweight fusion network. According to experimental results, the suggested approach outperforms single-modal and early-fusion baselines with an accuracy of 94.1%, an F1-score of 92.8%, an AUC of 96.2%, and an average inference latency of 31.6 ms. The findings show that while preserving real-time applicability for intelligent traffic monitoring systems, multimodal fusion can enhance accident risk discrimination.