Channel Spatio-temporal Convolutional Network for Pedestrian Trajectory Prediction

Abstract Accurate and timely prediction of the future path of agents in the vicinity of an agent is the core of avoiding conflict in automated applications. The traditional method based on the RNN model requires high computational cost in the prediction process, especially for long-series prediction. In order to obtain a more efficient and accurate prediction trajectory, a channel spatio-temporal convolutional network framework, CSTCN, is proposed in this paper. The framework models the spatial environment as a block of data input to the CSTCN and captures spatio-temporal interactions using an improved temporal convolutional network. Compared with the traditional model, the spatial and temporal modeling of the proposed model is calculated in each local time window so that it can be executed in parallel to obtain higher computational efficiency. Experimental results on five trajectory prediction benchmark datasets demonstrate that the proposed model is superior to the other seven state-of-the-art models in efficiency and accuracy.