Modeling And Research On Trajectory Prediction In Interactive Scenarios Based On Spatio-temporal Attention Mechanism

Autonomous driving technology has developed rapidly in recent years,and as a result of its potential to reduce traffic congestion,enhance travel efficiency,and boost user pleasure,it has gradually come into the focus of the public in society.The original intention of autonomous driving technology is to make travel easier,but more importantly,it also aims to increase safety by assessing and judging potential unanticipated threats in the traffic environment and generating corresponding plans for action.Therefore,autonomous driving intelligent vehicles must be precise in complex interacting circumstances.It can forecast the future driving paths of close-by cars,which helps with perception and comprehension of the traffic environment.In the actual road traffic environment,the driving behavior of vehicle entities will be influenced by past trajectories,interactions with other nearby vehicle entities,and other factors,leading to considerable uncertainty in their trajectory states in the future.How should time and space be handled in interactive scenarios? The problem with trajectory prediction is information with challenge.The serialization of historical trajectory encoding,the inadequate processing of implicit temporal dependencies,and the inadequate extraction of interaction information between vehicle entities are some of the issues with the current vehicle trajectory prediction research.The prediction accuracy,validity,and generalizability of existing trajectory prediction models are decreased in road situations where there are many multi-vehicle interactions.In order to address the challenges associated with the task of vehicle entity trajectory prediction in interactive scenarios in light of the shortcomings of the existing trajectory prediction models,this paper proposes a trajectory prediction model based on the spatio-temporal Attention mechanism,which realizes the parallel processing of vehicle trajectory sequences and various surrounding areas.vehicle entity interaction features are efficiently extracted.The research content of this paper is mainly carried out from the following three aspects:(1)The multi-head Graph Attention Network-based trajectory prediction model,GAT-LSTM,is constructed.The model first uses the LSTM encoder to time-encode the historical trajectory sequence to extract the kinematic features of the vehicle entity;the Graph Attention Network GAT composed of two multi-head graph attention layers is designed to extract the current interaction features of the surrounding vehicle entities and the target vehicle entity at each time step;finally,the Attention-LSTM decoder that combines Attention mechanism and LSTM is proposed to decode the kinematic features and interaction features to generate the future trajectory of the target vehicle entity.The prediction model is tested on the open-source trajectory dataset NGSIM.The results of quantitative analysis and qualitative analysis show that the GAT-LSTM model based on multi-head Graph Attention Network can achieve relatively accurate trajectory prediction.(2)The trajectory prediction model STI-Attention based on spatio-temporal interaction is proposed.In the kinematic feature extraction stage,the model uses the Self-Attention encoder composed of multiple Transformer encoder modules to perform time encoding of historical trajectories,which realizes the parallel processing of trajectory sequences,and it can pay attention to the correlations between different time steps;then the spatio-temporal interaction feature extraction layer is designed,which is composed of a multi-head Graph Attention Network GAT and the SelfAttention encoder based on a Transformer encoder module.It can not only extract the interaction information of surrounding vehicle entities to the target vehicle entity,but also process the time dependence of interaction information,which can be used to extract spatio-temporal interaction features.Finally,the Self-Attention decoder based on Encourage Diversification Error Function is designed to realize the generation of multiple futural trajectories and the selection of the most optimal trajectory.The experiments are carried out on the trajectory dataset NGSIM,and the experimental results demonstrate the effectiveness of the spatio-temporal interaction feature extraction in this model.(3)The trajectory prediction model SF-STI-Attention fused with scene information is proposed by introducing two scene information,vehicle physical characteristics and lane information.The model converts the physical features of the vehicle entity to the four vertices and the center point,adds the lane information of the vehicle entity at different time steps,and realizes the fusion of scene information by encoding.Each point of the vehicle entity corresponds to the Self-Attention encoder and a spatio-temporal interaction feature extractor.Finally,the kinematic features and interaction features of the five points are spliced together,and the future trajectory is generated through the Self-Attention decoder.On the data set collected from the real traffic environment,the prediction accuracy of the model is greatly improved,and it can predict results that are closer to the real trajectories in different scenarios,which further proves that the trajectory prediction model has strong generalization ability and rationality.