V2V-Enhanced Collision Risk Prediction Beyond the Sensing Horizon Using Prefix-Based Temporal Modeling

This paper presents a dynamic collision risk prediction framework based on a variable-length, prefix-based temporal modeling strategy that enables continuous, frame-by-frame risk estimation under partial observation histories. Unlike conventional fixed-length approaches, the proposed method progressively refines predictions as new observations become available, allowing early risk estimation from limited temporal context. The framework fuses monocular vision and 2D LiDAR to extract spatial and kinematic features, which are processed by a recurrent temporal model to infer a probabilistic collision risk at each time step. To overcome the line-of-sight limitation of ego-centric sensing, a Vehicle-to-Vehicle (V2V) cooperative extension is introduced, where neighboring vehicles broadcast raw kinematic observations that are incorporated as synthetic feature inputs into the temporal model, enabling risk estimation for occluded threats beyond the sensing horizon. The system is evaluated in a Digital Twin simulation environment across safe driving, near-collision, collision, and occluded intersection scenarios. Results show that the ego-centric model achieves stable predictions within 0.7s, with a precision of 0.94, recall of 0.90, and collision warnings up to 1.75s in advance. The cooperativeextension further improves early warning in occluded scenarios, recovering 1.35s of warning time lost (WTL) and reaching a collision warning up to approximately 2.3s in advance, while maintaining consistent probability evolution during transitions between cooperative and onboard sensing.