| With the improvement of living standards,people have more requirements for trip that has advantages of safe,comfortable and efficient.It then gives birth to intelligent driving,assisted driving,intelligent navigation and other application systems,which aim to use advanced information technology,sensor technology and intelligent processing technology to analyze the current situation and intention of surrounding road participants,so as to avoid risks and choose efficient travel routes in advance.The high-precision trajectory prediction of moving targets is the key technology to promote the development of these advanced application systems,and it has become a research hotspot.Among the current research methods of trajectory prediction,one is to predict the target trajectory by collecting the information of traffic participants,such as speed and position.It connects the broken or unobserved individual trajectories in the crowd to obtain the relationship between participants for the purpose of trajectory prediction.However,this method doesn’t consider the impact of traffic scene context on target movement.Another method is to jointly model scene information（such as roads,crosswalks,etc.）and participant information（such as cars,pedestrians,etc.）through neural networks,and finally obtain the result of target trajectory prediction.However,it only focuses on the analysis of the scene content,ignoring that not only the traffic participation elements affect the target movement,but also their location will affect the target movement.Based on the existing research,in order to improve the trajectory prediction accuracy,this paper proposes a trajectory prediction model based on scene spatiotemporal analysis.Specifically,the first step is to grid the perceptual scene,extract the deep semantic features of each grid by using convolutional neural network,and then form each channel grid into a graph network to mine the spatial features of the perceptual scene through graph convolutional neural network.This method can not only analyze the spatial semantic features of the scene,but also mine the correlation between the spatial features,which provides a more accurate analysis for the scene where the target is located.In addition,considering that the time information of the scene will also have an impact on trajectory prediction,this paper also propose a dynamic scene modeling method on the basis of scene spatial analysis.That is,the spatial semantic features of continuous multi frame perceptual scene are fused to obtain the spatiotemporal analysis features of the current frame.Finally,the target trajectory prediction result is obtained by establishing an end-to-end model of scene parsing and predicting the trajectory.The results of a large number of experiments on Stanford Drone Dataset（SDD）show that the model proposed in this paper has improved in ADE and FDE compared with the current advanced trajectory prediction models such as Sophie and Desire. |
