A Trajectory Prediction Approach Incorporating Spatial-Temporal Safety Features for Autonomous Vehicles

Abstract

Trajectory prediction is essential for autonomous driving in complex traffic environments. However, existing methods mainly focus on accuracy while paying limited attention to safety considerations. This paper proposes a multimodal trajectory prediction approach that integrates safety features in both spatial and temporal perspectives. By introducing safety indicators from these two perspectives, the proposed approach utilizes historical trajectories, interactive scenarios, and safety features jointly to predict future trajectories with improved safety and accuracy. A multimodal trajectory generation and corresponding probability assignment mechanism is designed to generate multiple candidate future trajectories and assign a probability to each predicted mode, which is further optimized through the probability assignment loss during training. In addition, a high-risk scenario identification strategy is developed based on the quantiles of safety metrics, providing a unified evaluation paradigm for assessing trajectory prediction performance under extremely risky conditions. Experimental results on the Argoverse 1 Motion Forecasting Data Set showed that our method achieves a superior minimum average displacement error (minADE) metric across diverse high-risk scenarios (minADE is reduced by 9.8% at maximum), demonstrating its strong safety, robustness, and cross-scenario generalization.