基于CrossFormer的自动驾驶车辆周边行人轨迹预测OA北大核心

To solve the problems of long-term pedestrian trajectory prediction around autonomous vehicles with insufficient performance and poor adaptability to complex scenarios by the existing methods,the novel method was proposed.The pedestrian trajectory prediction problem was modeled,and the CrossFormer-based pedestrian trajectory prediction method was developed.The dimension-segment-wise(DSW)embedding technology was introduced to explicitly learn the correlations between adjacent time frames,and the two-stage self-attention mechanism (TSA) was combined to comprehensively capture the long-term dependencies of pedestrian trajectories.The hierarchical encoder-decoder structure was employed to adaptively capture pedestrian trajectory dependencies at different time scales for enhancing the model scalability in long-term prediction.The multi-modal information fusion,the self-attention mechanisms and the scalability optimization were innovatively integrated to achieve efficient solution for pedestrian trajectory prediction tasks.The experiments were conducted on the two datasets of ETH and Jiangsu University campus pedestrian trajectory data(JDD).The time series segmentation analysis and quantitative and qualitative evaluations were performed.The results show that by the proposed method,the values of average displacement error (ADE) and final displacement error (FDE) are respective 0.627 and 1.32 on the ETH dataset,which are significantly better than those by the traditional methods of LSTM with 0.895 and 1.74 and SR-LSTM with 0.728 and 1.66.On the JDD dataset,the values of ADE and FDE are 0.281 and 0.53,respectively,which are far superior to those of GAN with 0.562 and 1.01 and STGAT with 0.673 and 1.43.The robustness and generalization ability of the proposed method in complex scenarios are verified.