DPDRF: Dynamic Predictive Driving Risk Field Based on Multi-Agent Trajectory Prediction and Digital Twins System
Jianhang Liu, Xizhao Sheng, Lizhuang Tan, Wei Zhang, Peiying Zhang, Kai Liu
Abstract
Assessing risk in traffic scenarios is crucial for advancing autonomous vehicle safety. As the number of vehicles and their interactions increase, there is a heightened need for prompt and precise risk evaluation. Traditional collision warning systems, which rely on conflict points, are limited to specific traffic scenarios and are inadequate for the dynamic demands of complex and diverse traffic environments. Additionally, current field-based risk assessment methods evaluate collision risks based only on instantaneous vehicle motion states, lacking in both timeliness and accuracy. To address these challenges, this paper proposed a novel risk assessment method. It involves constructing a vehicle-road cooperative environment using a distributed digital twins system and employs a multi-agent multi-modal trajectory prediction algorithm to forecast future vehicle motion states. By leveraging predictive multi-dimensional kinematic features, the proposed method dynamically generates future driving risk fields, effectively assesses collision risks in complex traffic scenarios by modeling multi-source collision risks within future traffic scenes. Extensive simulation results and analyses reveal that DPDRF significantly improves the average Pre-Collision Warning Time (PCWT) by 2.53 times compared to DDRPF and by 1.81 times compared to DSF. Furthermore, DPDRF reduces the average Pre-Collision Warning Error (PCWE) by 79% relative to PRF.