Space-Time Social Relationship Pooling Pedestrian Trajectory Prediction Model

Socially acceptable trajectories with generative adversarial networks(SGAN)does not consider the long-term social relationships of pedestrians,a spatial-temporal social relationship pooling pedestrian trajectory prediction model is proposed.The spatial-temporal social collecting mechanism is used to learn all the social interactions of the pedestrian observation sequence and obtain the social collecting vector of spatial-temporal mapping.And then the social relationship pooling method is applied to pooling the spatial-temporal social collecting vectors into social vectors of“gravity-repulsion”relationships,as a part of the hidden state input of the RNN decoder.The model can not only maintain the short-term social sensitivity of pedestrians,but also enhance the memory of long-term social relationships,and improves the model’s prediction accuracy for pedestrians with complex social relationships.In order to verify the reliability of the proposed model,the performance was tested on the public standard data sets ETH and UCY.Experiments show that the average offset accuracy error of our model is increased by 20%compared with the SGAN model,and the final offset accuracy error is increased by 13.9%.