DACR-AMTP: Adaptive Multi-Modal Vehicle Trajectory Prediction for Dynamic Drivable Areas Based on Collision Risk
Peichao Cong, Yixuan Xiao, Xianquan Wan, Murong Deng, Jiaxing Li, Xin Zhang
Abstract
Accurate vehicle trajectory prediction in autonomous driving technology poses significant challenges due to the varying driving states of different vehicles, their motion patterns, and multi-modal driving characteristics. To address these challenges, an innovative adaptive multi-modal vehicle trajectory prediction model, termed DACR-AMTP, is introduced in this paper. DACR-AMTP investigates the dynamic interrelationships among multiple vehicles using a dynamic drivable area determination strategy in a graph-like structure. This strategy aids in guiding the vehicle trajectory prediction trend based on risky collision probabilities. Additionally, a multi-headed attention mechanism is implemented, integrating regionally fused trajectory information to efficiently capture the long-term spatio-temporal dependencies present in the vehicle's historical data. This, in turn, enhances the rationality and accuracy of vehicle trajectory prediction. An adaptive multi-modal trajectory prediction generator, constructed based on the full-probability theorem, incorporates dynamic drivable area occupancy states and collision probabilities to adaptively output multi-modal vehicle drivable trajectories. Experimental results on three publicly available datasets demonstrate that DACR-AMTP can achieve real-time multi-vehicle trajectory prediction, outperforming current state-of-the-art algorithms in terms of prediction accuracy. Furthermore, ablation experiments underline the crucial role of the dynamic drivable area determination strategy in long-term vehicle trajectory prediction.