Research On Attention-based Temporal Convolutional Network For Pedestrian Trajectory Prediction

Understanding human behavior is an important part of the realization of intelligent unmanned systems.Accurately perceiving and predicting the trajectory of a moving target provides a data foundation and security guarantee for the unmanned system to perform tasks in a complex realistic environment.As the most important feature in the visual world,trajectory is the key basic information for unmanned systems to understand the real world.Predicting the future trajectory of surrounding targets can more efficiently carry out path planning and decision-making.Accurately predicting the trajectory can enhance the safety of the intelligent system,and prevent itself from colliding with other moving objects or obstacles.Aiming at the problem that the current trajectory prediction model cannot effectively capture the temporal features in the trajectory sequence and the prediction accuracy is not high,the thesis proposes a trajectory prediction model based on the temporal attentional convolutional network.The causal convolution,temporal attention and enhanced residual in the model can effectively process temporal features of the trajectory and achieve good results.The temporal features and spatial interaction features in the trajectory sequence in the actual scene influence the development of each other,and most of the current works only consider some of the features.This thesis uses the graph attention network to extract the spatial interaction information in the trajectory,and combines the temporal attentional convolutional network extracts the temporal-spatial correlation information in the sequence.The thesis designs and implements a pedestrian trajectory prediction model based on temporal convolutional neural network based on the double attention mechanism,and considers the temporal features,spatial interaction features and temporal-spatial correlation features of the pedestrian trajectory sequence in the trajectory prediction task.The proposed model achieves a good trajectory prediction effect and good real-temporal performance.In order to verify the effectiveness of the proposed model,this paper compares the accuracy and running speed of the proposed model with mainstream models such as STGAT on the ETH and UCY datasets.Experimental results show that when the proposed model observes 3.2 seconds(8 frames)before the trajectory,the average error of predicting the trajectory in the future 4.8 seconds(12 frames)is 0.23 meters.Compared with STGAT,the average prediction accuracy is improved by 49.5%.Compared with the Trajectron,the running speed is increased by 25%.In addition,the proposed model also achieves high prediction accuracy on the self-made dataset,which proves that the trajectory prediction method has good generalization ability.