REGION: Relevant Entropy Graph spatIO-temporal convolutional Network for Pedestrian Trajectory Prediction

Modeling pedestrian interaction is an essential building block in pedestrian trajectory prediction, which raises various challenges such as the complexity of social behavior and the randomness of motion. In this paper, a new relevant entropy spatio-temporal graph convolutional network is proposed to model pedestrian interaction for pedestrian trajectory prediction, which contains regional spatiotemporal graph convolutional neural network and gated dilation causal convolutional neural network. The regional spatio-temporal graph convolutional neural network creates a matching graph structure for each time step, and calculates the weighted adjacency matrix of each graph structure through relevant entropy to obtain the sequence embedding representation of the pedestrian interaction relationship. The gated dilation causal convolutional neural network reduces the linear superposition of the hidden layer through the setting of the dilated factor, and uses the gating mechanism to filter the features. Experiments are carried out on the standard data sets ETH and UCY, higher accuracy and efficiency verify that the proposed method is effective in pedestrian interaction modeling.