Title: From Cues to Horizons: Dynamic Risk Horizon Profiling for Trajectory Prediction

Abstract: Ensuring the safety of autonomous driving hinges on the ability to accurately and reliably predict vehicle trajectories. While recent research has begun integrating safety risks into trajectory prediction models to quantify threats from nearby agents, most existing risk-aware methods rely on historical risk data as a secondary guidance signal. These approaches often neglect the future evolution and inherent uncertainty of risk. To address this gap, we introduce a Risk Horizon Profiling (RHP) module that utilizes a continuous, learnable potential field model for risk-aware trajectory prediction. The RHP module assesses the spatial-temporal proximity of surrounding objects to map risk distributions across future time horizons. This process enhances trajectory prediction by adaptively pinpointing moments that human drivers consider critical. We validated our approach using two distinct datasets: highD, representing highway corridors, and SHRP2, representing urban streets, both of which encompass a wide range of risk scenarios, including safe driving, near-crashes, and actual crashes. Our framework demonstrated superior performance compared to baseline methods, achieving a 25.0% decrease in 5-second RMSE on the highD dataset and a 29.1% reduction in 5-second minFDE on SHRP2. These findings highlight the method’s strong efficacy for both short- and long-horizon predictions, as well as its robust generalization across highway and urban environments. By facilitating more realistic autonomous vehicle path planning and strategic decision-making, this work contributes to safer autonomous driving and the advancement of driver-assistance systems. The source code for this research is accessible at: https://github.com/bilab-nyu/RHP

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC