Pedestrian Trajectory Prediction Based On Spatio-Temporal Graph Convolutional Neural Network

As an important part of automatic driving vehicle perception technology,pedestrian trajectory prediction plays a key role in safe path planning.Pedestrian movement has great randomness and is affected by the complex interaction behavior of surrounding traffic participants.Therefore,it is still a challenge to accurately predict pedestrian trajectory.In this paper,the problem of pedestrian trajectory prediction is deeply studied,and a Spatio-Temporal Graph Convolutional Neural Networks based social interaction model enhanced with pre-prediction for pedestrian trajectory prediction(STGCNN-SIM)is proposed.The research contents are as follows:(1)Construction of multi social feature interaction model.By analyzing the shortcomings of the existing models,this paper proposes to extract four social interaction features from the trajectory to describe the interaction behavior between pedestrians and surrounding traffic participants:(1)the relative distance feature between pedestrians and nearby traffic participants;(2)the angle feature between two velocity vectors of pedestrians and nearby traffic participants;(3)the angle feature between the pedestrian velocity vector and the distance vector;(4)the angle feature between the velocity vector of the nearby traffic participants and the distance vector.A Spatio-Temporal Graph Convolutional Neural Networks for pedestrian trajectory prediction is established.Experiments show that some newly introduced social interaction features can improve the prediction accuracy.(2)Construction of pre-prediction model.The network framework of pedestrian trajectory prediction model(STGCNN-SIM)using pre-prediction module to enhance social interaction behavior is proposed.The model not only uses the interaction behavior in historical observation stage,but also pre-predicts the future trajectory based on the constant velocity(CV)dynamics model to establish the future interaction behavior.Four social interaction features are extracted from the trajectories of the historical observation stage and the pre-prediction stage,weighted into the adjacency matrix to describe the interaction behavior,and the Spatio-Temporal attention mechanism is designed to screen the important interaction features.The information fusion module fuses the interaction behaviors of the two stages to realize pedestrian trajectory prediction.(3)Dataset establishment.The HNU pedestrian trajectory dataset was established and published to enrich the experimental samples of social groups with different cultural backgrounds,religious beliefs and value systems.(4)Experimental verification.The future 4.8s pedestrian trajectory prediction experiment is carried out on ETH,UCY and HNU pedestrian dataset established in this paper.The results show that the ADE/ FDE of STGCNN-SIM model proposed on HNU dataset is 0.43m/0.74 m,which is better than the existing model,and the calculation time overhead is moderate,which is 62.29 ms.