STIRNet: A Spatial-Temporal Interaction-Aware Recursive Network for Human Trajectory Prediction

Abstract

Pedestrian trajectory prediction is one of the important research topics in the field of computer vision and a key technology of autonomous driving system. However, it’s full of challenges due to the uncertainties of crowd motions and complex interactions among pedestrians. We pro-pose a Spatiotemporal Interaction-aware Recursive Net-work (STIRNet) to predict multiply socially acceptable trajectories of pedestrians. In this paper, a recursive structure is used to capture spatio-temporal interactions by spatial modeling and temporal modeling alternately. At each time-step, the spatial interactions are modeled by a graph attention network, in which the nodes feature are represented by temporal motion features. The learned spatial interaction context is used to capture temporal motion features through an LSTM model. The temporal motion features are used to infer future positions and update nodes features. Experimental results on two public pedestrian trajectory datasets (ETH and UCY) demonstrate that our proposed model achieves superior performances compared with state-of-the-art methods on ADE and FDE metrics.