Traffic Agents Trajectory Prediction Based on Spatial–Temporal Interaction Attention

Abstract

Trajectory prediction aims to predict the movement intention of traffic participants in the future based on the historical observation trajectories. For traffic scenarios, pedestrians, vehicles and other traffic participants have social interaction of surrounding traffic participants in both time and spatial dimensions. Most previous studies only use pooling methods to simulate the interaction process between participants and cannot fully capture the spatio-temporal dependence, possibly accumulating errors with the increase in prediction time. To overcome these problems, we propose the Spatial–Temporal Interaction Attention-based Trajectory Prediction Network (STIA-TPNet), which can effectively model the spatial–temporal interaction information. Based on trajectory feature extraction, the novel Spatial–Temporal Interaction Attention Module (STIA Module) is proposed to extract the interaction relationships between traffic participants, including temporal interaction attention, spatial interaction attention, and spatio-temporal attention fusion. By adaptive allocation of attention weights, temporal interaction attention is a temporal attention mechanism used to capture the movement pattern of each traffic participant in the scene, which can learn the importance of historical trajectories at different moments to future behaviors. Since the participants number in recent traffic scenes dynamically changes, the spatial interaction attention is designed to abstract the traffic participants in the scene into graph nodes, and abstract the social interaction between participants into graph edges. Coupling the temporal and spatial interaction attentions can adaptively model the temporal–spatial information and achieve accurate trajectory prediction. By performing experiments on the INTERACTION dataset and the UTP (Unmanned Aerial Vehicle-based Trajectory Prediction) dataset, the experimental results show that the proposed method significantly improves the accuracy of trajectory prediction and outperforms the representative methods in comparison.