一种基于时空注意力机制的行人轨迹预测模型OA

With the rapid progress of urbanization,the ways in which people interact with cities are constantly changing.Precisely predicting the future trajectory of pedestrians can provide a decision-making basis and technical support for areas such as transportation and urban planning.To better establish a social model for predicting pedestrian trajectories,most researchers have introduced attention mechanisms to model the spatial relationships between pedestrians in the model,but less attention has been paid to the temporal correlation of mutual influence between pedestrians.In response to the above problems,this paper proposes a pedestrian trajectory prediction model based on Spatio-Temporal Multi-head Self-attention(STMS).STMS constructs an improved Graph Attention structure(N-GAT)to efficiently capture interactions and spatial interactions between pedestrians.Secondly,it innovatively introduces the S-GRU and TC-GRU modules to model the temporal dependencies of pedestrian spatial interactions.This approach not only extracts key features in the temporal dimension but also captures the relationships between interactions at different time steps.Finally,the model adopts an encoder-decoder(Seq2seq)framework to integrate the modeled pedestrian information and employs D-GRU as the decoder to predict trajectories.Through comparative experiments and ablation experiments on two public datasets ETH and UCY,we demonstrate that STMS exhibits superior modeling capabilities in capturing the interactive relationships among pedestrians,significantly enhancing the overall accuracy of pedestrian trajectory prediction.