Research On Pedestrian Trajectory Prediction Algorihm Under Attention Mechanism

With the advent of the internet of things era,intelligent products gradually appear and integrate into people’s life.The pedestrian trajectory prediction with complex social interaction and physical interaction is of great significance to the planning of the motion path of autonomous vehicles,robots and other agents.At present,the deep learning is becoming more and more perfect.The traditional manual custom model is gradually eliminated because it can not effectively model a variety of complex environments.Instead,the pedestrian trajectory prediction algorithm based on neural network and attention mechanism has achieved remarkable results in improving the prediction accuracy.Based on the in-depth study of deep learning and pedestrian trajectory prediction algorithms,this paper focuses on the physical environment modeling and historical trajectory time-dependent learning in pedestrian trajectory prediction,analyzes the existing model architecture,and proposes the following pedestrian trajectory prediction algorithms:(1)Social and scene perceived pedestrian trajectory prediction based on GAN.In order to solve the problem that the state refinement for LSTM towards pedestrian trajectory prediction(SR-LSTM)does not consider the influence of physical environment on pedestrian trajectory prediction,a social and scene perceived pedestrian trajectory prediction model based on GAN is proposed.Firstly,the state refinement module of SRLSTM model is used to model the social interaction of pedestrians,and the state refinement module captures the current important intention of neighbors in order to select important information from adjacent pedestrians.Then,the semantic pooling mechanism is introduced to learn the interaction between pedestrians and physical environment.The semantic pooling defines the semantics of the physical scene and learns its correlation with pedestrian trajector.Finally,in order to generate more realistic and reasonable trajectory samples,Info GAN network is used for training.(2)Pedestrian trajectory prediction based on spatio-temporal attention mechanism.In order to solve the problem that the existing pedestrian trajectory prediction algorithms only model from the long-term dependence of pedestrian historical trajectory and ignore the local dependence of time series data,the model proposes a pedestrian trajectory prediction algorithm based on spatio-temporal attention mechanism from the point of view of spacetime map.Firstly,the graph attention network(GAT)is used to learn the social interaction characteristics between pedestrians and their neighbors.Then,considering the local dependence of pedestrian historical trajectory,R-Transformer that is a new attention model is introduced to learn the characteristics of historical trajectories.Finally,the spatiotemporal features fusion method of STAR model is used to fuse the temporal and spatial characteristics and predict the trajectory.In order to verify the effect of the model,ADE and FDE are used as evaluation indexes to carry out experiments on two common data sets of ETH and UCY,and the prediction results of the two models are compared with those of other models.The comparison results show that compared with the SR-LSTM model,the social and scene perceptual pedestrian trajectory prediction model based on GAN reduces ADE by 8.9% and FDE by 12.8%.At the same time,compared with other models,it can also reduce the prediction error.The pedestrian trajectory prediction model based on the spatial-temporal attention mechanism achieves higher prediction accuracy and can generate more realistic and reasonable samples than STGAT and other spatial-temporal modeling methods.