EGAT: Extended Graph Attention Network for Pedestrian Trajectory Prediction

Wei Kong,Yun Liu,Hui Li,Chuanxu Wang

DOI: https://doi.org/10.1155/2021/9985401

IF: 3.12

2021-10-19

Computational Intelligence and Neuroscience

Abstract:To improve foresight and make correct judgment in advance, pedestrian trajectory prediction has a wide range of application values in autonomous driving, robot interaction, and safety monitoring. However, most of the existing methods only focus on the interaction of local pedestrians according to distance, ignoring the influence of far pedestrians; the range of network input (receptive field) is small. In this paper, an extended graph attention network (EGAT) is proposed to increase receptive field, which focuses not only on local pedestrians, but also on those who are far away, to further strengthen pedestrian interaction. In the temporal domain, TSG-LSTM (TS-LSTM and TG-LSTM) and P-LSTM are proposed based on LSTM to enhance information transmission by residual connection. Compared with state-of-the-art methods, the model EGAT achieves excellent performance on both ETH and UCY public datasets and generates more reliable trajectories.

mathematical & computational biology,neurosciences

What problem does this paper attempt to address?

This paper proposes a new method called Extended Graph Attention Network (EGAT) to solve the problem of pedestrian trajectory prediction. In pedestrian interaction prediction, existing methods mainly focus on the interaction of nearby pedestrians, ignoring the influence of distant pedestrians, leading to a small input range (receptive field) of the network. The paper proposes a Feature Updating Mechanism (FUM) to explore global influence and increase the receptive field to focus on distant pedestrians. In addition, the paper introduces Time-Series Graph LSTM (TSG-LSTM), which includes TS-LSTM and TG-LSTM, and Prediction LSTM (P-LSTM) to enhance information transmission at the current time.

In the temporal domain, TS-LSTM and TG-LSTM use LSTM's residual connections to enhance information transmission. TS-LSTM encodes the information of individual pedestrians, while TG-LSTM encodes the interaction between pedestrians. P-LSTM predicts future trajectories based on observed trajectories and mitigates the decrease in prediction accuracy as the prediction length increases.

Compared to existing methods, EGAT performs excellently on the ETH and UCY public datasets, generating more reliable trajectories. The study indicates that pedestrian trajectory prediction is influenced not only by spatial proximity but also by environmental complexity and distant pedestrians. Therefore, EGAT improves the accuracy and comprehensiveness of predictions by expanding the range of attention networks to consider all non-local pedestrians.