Research On Trajectory Prediction Algorithm Based On Graph Convolutional Neural Network Under The Influence Of Multiple Factors

Predicting the future trajectories of surrounding moving individuals is important for autonomous vehicles and other autonomous mobile platforms,and only by accurately and reasonably predicting the movements of neighboring individuals can reasonable path planning be made to maintain the efficiency and safety of the transportation system.Realistic traffic scenarios contain many factors that may affect the future trajectories of individuals,such as the motion states of surrounding individuals,social relationships between individuals,and scene information,which pose a great challenge to the trajectory prediction task.However,existing research works often choose to aggregate some trajectory hidden states or use the distance between individuals to calculate the influence between individuals,failing to comprehensively consider or reasonably model the influence factors,and the prediction effect is not satisfactory.To address the above problems,this paper explores the relative distance,speed,direction and scene information and other influencing factors contained in the trajectory data in the limited data semantics,and designs new trajectory prediction algorithms by combining graph convolutional neural network,the main research content includes:(1)Proposed SSAGCN,a trajectory prediction model based on social soft attention function,which uses social soft attention function to reconstruct the graph relations contained in the spatio-temporal graph of trajectory sequences,combining graph convolutional neural network to extract interaction features.The social soft attention function can comprehensively consider many factors that affect the future trajectory of individuals,and reasonably use these factors to calculate the interaction between individuals in social interactions according to the laws of practice.Scene information is also incorporated into the SSAGCN trajectory prediction model in this process,and its role is extended in time and space using a sequential scene sharing mechanism.In this thesis,experiments are conducted on publicly available datasets,and the experimental results demonstrate the positive effect of the social soft attention function,in addition to the comparison with the baseline approach demonstrating that the SSAGCN model can predict trajectories that conform to social rules and physical constraints.(2)Proposed Multi-StreamGCN,a trajectory prediction model based on multistream graph convolution,which introduces a multi-relational graph,models each factor affecting the movement intention of traffic individuals as a separate relational graph,and then combines the self-attention mechanism with multi-stream graph convolution for trajectory prediction.The influence between individuals is learned in a multi-relationship graph based on a self-attention mechanism,and the social interaction features of each relationship in space are extracted using multi-stream graph convolution.After feature fusion,Mufti-StreamGCN further calculates the time dependence in the sequence of interacting features based on the self-attention mechanism.In this thesis,we tested the Multi-StreamGCN model on publicly available datasets and demonstrated the better performance of the Multi-StreamGCN model through quantitative and qualitative analysis.