Pedestrian Trajectory Prediction Based on Improved Social Spatio-Temporal Graph Convolution Neural Network

Abstract

Pedestrian trajectory prediction is a key technology in the field of autonomous driving. The trajectory of pedestrians is not only affected by the surrounding objects, but also by the social interaction between adjacent pedestrians. Aiming at the problem that pedestrian interaction visual blind area is easy to be ignored in pedestrian trajectory prediction, a convolution network algorithm based on spatio-temporal graph is proposed. Firstly, Pedestrian feature information is obtained by spatio-temporal graph. Then the connection weights between irrelevant points are screened out according to the visual blind area. Finally, the time extrapolation convolution neural network (TXP-CNN) is used to predict the future trajectory of pedestrians. Through experiments on two public datasets (ETH and UCY), average displacement error (ADE) and the final displacement error (FDE) of the proposed model on the dataset are 0.42 and 0.73, respectively.