[Paper Review] SGCN:Sparse Graph Convolution Network for Pedestrian Trajectory Prediction

Liushuai Shi, Le Wang+5 more|arXiv (Cornell University)|Apr 4, 2021

Autonomous Vehicle Technology and Safety48 references17 citations

TL;DR

This paper proposes SGCN, a Sparse Graph Convolution Network that models sparse directed interactions and motion tendencies in pedestrian trajectories using two separate sparse directed graphs—spatial and temporal. By fusing these graphs to estimate a bi-Gaussian distribution, SGCN achieves 9% lower Average Displacement Error (ADE) and 13% lower Final Displacement Error (FDE) than state-of-the-art methods on ETH and UCY datasets, with improved generalization in complex interaction scenarios.

ABSTRACT

Pedestrian trajectory prediction is a key technology in autopilot, which remains to be very challenging due to complex interactions between pedestrians. However, previous works based on dense undirected interaction suffer from modeling superfluous interactions and neglect of trajectory motion tendency, and thus inevitably result in a considerable deviance from the reality. To cope with these issues, we present a Sparse Graph Convolution Network~(SGCN) for pedestrian trajectory prediction. Specifically, the SGCN explicitly models the sparse directed interaction with a sparse directed spatial graph to capture adaptive interaction pedestrians. Meanwhile, we use a sparse directed temporal graph to model the motion tendency, thus to facilitate the prediction based on the observed direction. Finally, parameters of a bi-Gaussian distribution for trajectory prediction are estimated by fusing the above two sparse graphs. We evaluate our proposed method on the ETH and UCY datasets, and the experimental results show our method outperforms comparative state-of-the-art methods by 9% in Average Displacement Error(ADE) and 13% in Final Displacement Error(FDE). Notably, visualizations indicate that our method can capture adaptive interactions between pedestrians and their effective motion tendencies.

Motivation & Objective

• To address the limitations of dense, undirected interaction modeling in pedestrian trajectory prediction, which introduces superfluous interactions and fails to capture realistic motion tendencies.
• To model adaptive, directed interactions between pedestrians that reflect real-world influence dynamics, rather than assuming symmetric or universal interactions.
• To incorporate motion tendency—short-term trajectory direction—into prediction to improve accuracy, especially in non-colliding or group-walking scenarios.
• To develop a graph-based deep learning framework that jointly learns spatial interaction and temporal motion patterns using sparse, directed graphs.
• To outperform existing state-of-the-art methods in trajectory prediction accuracy on benchmark datasets like ETH and UCY.

Proposed method

• The method constructs a Sparse Directed Spatial Graph to model asymmetric, adaptive interactions between pedestrians, where only relevant agents influence a target pedestrian’s trajectory.
• A Sparse Directed Temporal Graph is learned to capture motion tendencies from short-term trajectory segments, modeling the direction and stability of movement.
• Self-attention mechanisms are used in the Sparse Graph Learning module to dynamically assign attention weights to potential interaction partners, enabling adaptive graph construction.
• The outputs from the spatial and temporal graphs are fused to predict a bi-Gaussian distribution over future trajectories, capturing both mean and uncertainty.
• The model is trained end-to-end using a differentiable loss function that minimizes ADE and FDE on observed trajectory sequences.
• The framework is evaluated on ETH and UCY datasets, with ablation studies confirming the contribution of each component.

Experimental results

Research questions

• RQ1 Can sparse directed interaction modeling reduce the impact of superfluous interactions compared to dense undirected models in pedestrian trajectory prediction?
• RQ2 Does incorporating motion tendency improve prediction accuracy, especially in scenarios with minimal or no collision?
• RQ3 Can a graph-based approach that jointly models spatial interactions and temporal motion tendencies outperform existing state-of-the-art methods?
• RQ4 How does the model perform in complex scenarios such as parallel walking, group movement, or crossing paths?
• RQ5 To what extent do the learned sparse directed interactions align with ground-truth interaction patterns in real-world pedestrian dynamics?

Key findings

• SGCN achieves a 9% relative improvement in Average Displacement Error (ADE) and a 13% relative improvement in Final Displacement Error (FDE) compared to state-of-the-art methods on the ETH and UCY datasets.
• Visualizations show that SGCN produces less overlapping predicted trajectory distributions than Social-STGCNN and SGAN, indicating reduced false collision predictions.
• The model accurately captures static pedestrians with low variance in predictions, reflecting that they are not influenced by others, as seen in scenarios where one pedestrian remains still.
• Sparse Directed Interaction visualizations confirm that SGCN dynamically identifies relevant interaction partners, such as only the red pedestrian influencing the blue pedestrian in a detour scenario.
• Ablation studies show that SGCN-V2 and SGCN-V3, which use sparse interaction, outperform SGCN-V4, which uses dense interaction, proving the benefit of sparsity.
• The method generalizes well in complex scenes like parallel walking and multi-pedestrian meetings, where predictions closely follow ground-truth trajectories.