Research On Pedestrian Trajectory Prediction Of Topological High-Order Neighborhood Graph Neural Network

With the development of automatic driving technology,it is very important to accurately predict the pedestrian trajectory in dynamic scenes.Graph Neural Network(GNN)has been widely used to predict individual relationships in social networks.On the one hand,fully mining the network topology information in graph data can improve GNN’s reasoning ability on graph data and generate accurate node feature representation for downstream tasks;On the other hand,using GNN to process the social network generated by pedestrian interaction can efficiently model the social consciousness and improve the accuracy of trajectory prediction.Therefore,how to improve the utilization of graph structure information by GNN and how to extract the social interaction between pedestrians in pedestrian trajectory prediction task are urgent technical problems to be solved.Aiming at the problem that the existing GNN ignores the graph structure information and the homogeneous extraction of social interaction by pedestrian trajectory prediction model,this paper studies the node aggregation method of graph structure information and the modeling of pedestrian social consciousness information.The main work is as follows:(1)Research on Reconstruction of Neighborhood in Graph Neural Networks with Attention-Based Topological Patterns(ATPGNN).The factors affecting the performance of existing GNN are analyzed,the improvement scheme is designed,and the ATPGNN model is proposed.Aiming at the problem that the existing GNN ignores high-order similar nodes,firstly,the graph data is mapped to the potential space through the node embedding method and the high-order topology graph is created,and then the node aggregation operation is performed on the high-order topology graph to obtain the feature representation of high-order similar nodes.Aiming at the problem that the existing GNN cannot make full use of the topological information of the graph,the node embedding vector of the graph structure information is applied to the edge weight calculation between nodes.Aiming at the problem that the existing GNN cannot change the model architecture to adapt to different datasets or downstream tasks,this paper proposes to modularize the three components of the model.Finally,the attention mechanism of graph is applied to aggregate long-distance similar node information,node graph structure information and node features to update each node to obtain the final stable feature representation.(2)Research on multimodal pedestrian trajectory prediction of heterogeneous social interaction.Based on the ATPGNN model,this paper analyzes the establishment relationship of social topology information between pedestrians in the prediction scene,and proposes the pedestrian trajectory prediction model Social-ATPGNN.Aiming at the problem of extracting the homogeneity of pedestrian social interaction at the same time in the spatial domain by the existing pedestrian trajectory prediction model,a non-fully connected directed social topology is constructed.Aiming at the problem that the traditional GNN ignores the structural information in the social topology,ATPGNN is used to replace the traditional GNN to fully mine and make use of the structural information in the social topology.Aiming at the problem of serial operation and large parameter scale of LSTM module for extracting pedestrian trajectory correlation in time domain,the Temporal Convolutional Network with residual connection is used to replace LSTM module,reduce the model parameters and realize the parallel operation of pedestrian trajectory in different time steps.