Geometric Graph Neural Network Modeling of Human Interactions in Crowded Environments

TL;DR

This paper proposes a geometric graph neural network (GNN) architecture that integrates domain knowledge from psychological studies to model pedestrian interactions and predict future trajectories and demonstrates improved prediction accuracy through reduced average and final displacement error metrics.

Abstract

Modeling human trajectories in crowded environments is challenging due to the complex nature of pedestrian behavior and interactions. This paper proposes a geometric graph neural network (GNN) architecture that integrates domain knowledge from psychological studies to model pedestrian interactions and predict future trajectories. Unlike prior studies using complete graphs, we define interaction neighborhoods using pedestrians' field of view, motion direction, and distance-based kernel functions to construct graph representations of crowds. Evaluations across multiple datasets demonstrate improved prediction accuracy through reduced average and final displacement error metrics. Our findings underscore the importance of integrating domain knowledge with data-driven approaches for effective modeling of human interactions in crowds.

Figures (3)

• Figure 1: A visual representation of the neighborhood of interaction for a pedestrian (shown in black), incorporating a 180-degree field of view and a 5-meter distance threshold, selected based on empirical observations from psychological studies as described in rio2018local. Pedestrians within the shaded region, which extends up to 90 degrees on either side of the walking direction, are considered part of the neighborhood. The thickness of the connecting lines decreases with increasing distance, indicating that closer neighbors have a stronger influence on the pedestrian's motion (corresponding to Graph 2 neighborhood definition with distance-based weighting according to our kernel definitions)
• Figure 2: Illustration of our geometric GNN model. From observed pedestrian trajectories over $T$ frames, we construct a kernel-based spatio-temporal graph by defining neighborhood relationships and weighting edges using kernel functions. This graph is processed by a spatio-temporal graph CNN, similar to the architecture proposed by mohamed2020social, to generate a spatio-temporal embedding. A temporal CNN then predicts the distribution of future trajectories. $V$ represents the set of nodes, $P=2$ is the dimension of pedestrian positions, and $\hat{P}$ denotes the dimensions of the embedding from the predicted Gaussian distribution. Our approach integrates domain knowledge from psychological studies, resulting in a more context-aware and interpretable adjacency matrix.
• Figure 3: Visualizing the predicted distribution of pedestrian trajectories for a representative scene (UNIV). Sampled trajectories from the predicted Gaussian distribution are depicted as dot points in the scatter plot, while observed and true trajectories are shown as dashed and dotted lines respectively.