Research On Trajectory Prediction Algorithm Based On Hypergraph Convolution

Human trajectory prediction refers to predicting future trajectories of pedestrians in a given scene based on their trajectories in a historical time period.It is important to applications such as autonomous driving,robot navigation,and intelligent monitoring systems.The challenges faced by the pedestrian trajectory prediction task mainly include three aspects.Firstly,there are complex social interactions between pedestrians,and the trajectory of the predicted target will be affected by surrounding pedestrians.This interaction does not only exist between two pedestrians,and different pedestrians have different influences on the predicted target,so it is difficult to use a certain fixed rule to model the interactions between pedestrians.Secondly,surrounding environments also affect trajectory of pedestrians.They must avoid collisions with obstacles such as street lights and trees.Therefore,it is of great importance to use the scene information reasonably.Finally,predictions of future trajectories relies on the spatiotemporal features of pedestrian movements extracted by the model from the trajectory data.As time goes by,the interactions between the predicted target and surrounding pedestrians change continuously and have a continuous impact on the target’s future trajectory,so the spatial interaction features of pedestrians at each moment are also temporally dependent.This makes it difficult to extract discriminative spatiotemporal features from trajectory data.To solve the above problems,this thesis proposes two new trajectory prediction methods:(1)Traj ectory prediction method based on spatiotemporal hypergraph convolution(ST-HGCNN).This method introduces hypergraph convolution into the task of pedestrian trajectory prediction innovatively,which can effectively model complex interactions between pedestrians.Compared with methods that use traditional graph structures to model pedestrian trajectories,ST-HGCNN can not only encode high-order correlations between pedestrians,but also be more convenient in modeling different types of interaction relationships.The two kernel functions designed in ST-HGCNN consider the speed,direction and relative distance between pedestrians,which can effectively simulate the peer relationship and collision avoidance relationship between pedestrians.Experimental results on the pedestrian trajectory dataset ETH-UCY show that prediction results of ST-HGCNN are improved by 2.2%on the ADE and 8%on the FDE compared with Social-STGCNN.In addition,experimental results on the vehicle trajectory dataset VTP-TL show that this method is also suitable for vehicle trajectory prediction tasks and the prediction results are superior to other comparison methods on ADE/FDE.(2)Trajectory prediction method based on HGA-LSTM.To extract more discriminative spatiotemporal features from trajectory data,this thesis fuses recurrent neural networks and hypergraph convolutional neural networks into a unified structure.The input,hidden state,and memory cells of HGA-LSTM are all hypergraph structured data,and all three gating units use hypergraph convolution to process input information and hidden states,so the hidden states and memory cells can not only contain temporal dynamic features,but also carry spatial interaction features.In addition,HGA-LSTM adds an attention mechanism to the hypergraph convolution to calculate the influence degree of different pedestrians on the predicted target.Feature augmentation module in HGA-LSTM uses convolutional neural network and attention mechanism to extract scene features,fuse scene features with original pedestrian features,and uses LSTM to eliminate the scale difference between the two features.Experimental results on the pedestrian trajectory dataset ETH-UCY show that compared with Social-STGCNN,prediction results of HGA-LSTM are improved by 9%on the ADE and 13.3%on the FDE.Experimental results on the vehicle trajectory dataset VTP-TL show that the prediction results of this method are better than other comparison methods on ADE/FDE.