Spatiotemporal Attention-Based Pedestrian Trajectory Prediction Considering Traffic-Actor Interaction

Abstract

Accurate pedestrian trajectory prediction is significant and challenging for traffic-actor protection and intelligent driving. However, most of the existing methods focus on pedestrian-pedestrian interaction ignoring the pedestrian-vehicle interaction and containing much ineffective interaction. To tackle the problem, this paper proposes a spatiotemporal LSTM attention network (STLAN) for pedestrian trajectory prediction based on a spatiotemporal dual attention module (STDAM). First, a pedestrian social force model is established to analyze the effective pedestrian-pedestrian interaction and pedestrian-vehicle interaction in mixed traffic scene. Second, a pedestrian trajectory prediction network architecture is constructed to integrate temporal and spatial information. By temporal attention STDAM captures underlying dependencies between time steps in individual pedestrian, assigning different weights to the input sequence of it. By spatial attention STDAM extracts latent interaction between traffic actors, assigning different weights to every traffic actor. In addition, random noise is introduced to simulate the uncertainty of pedestrian trajectory. Experimental results on a real-world traffic actors dataset show that STLAN outperforms several mainstream methods. Specifically, STLAN achieves 0.17/0.11 ADE/FDE and 3.79/0.38 AMD/AMV in the representative pedestrian-pedestrian interaction scene, with 0.15/0.17 ADE/FDE and 2.36/0.21 AMD/AMV in pedestrian-vehicle interaction scene. Furthermore, ablation experiments validate the effectiveness of STDAM in improving trajectory prediction accuracy.