LG-STSGCN: Long-Term Gated Pedestrian Trajectory Prediction Based on Spatial–Temporal Synchronous Graph Convolutional Network

Abstract

Pedestrian trajectory prediction is fundamental research in many practical applications, such as video surveillance, autonomous vehicles, and robotic systems. However, the existing methods do not capture the spatial–temporal correlation of pedestrians well and simultaneously, as well as do not learn the temporal global interaction features of pedestrians effectively. To address these issues, we propose a long-term gated pedestrian trajectory prediction model based on spatial–temporal synchronous graph convolutional network. The proposed method consists of three components. First, we construct a localized spatial–temporal graph to characterize the temporal information, spatial information and spatial–temporal correlation information among pedestrians in the pedestrian trajectory prediction fully. Then, we introduce a gated mechanism into the temporal convolutional network, in parallel with the gated spatial–temporal synchronous graph convolutional network, in order to improve the model's ability to capture the global correlation of spatial–temporal data. Finally, we add random noise and use a diversity loss function to train and predict trajectories. We conduct experiments on ETH and UCY datasets and the proposed method is proved to outperform previous approaches.