[Paper Review] Social-WaGDAT: Interaction-aware Trajectory Prediction via Wasserstein Graph Double-Attention Network

Jiachen Li, Hengbo Ma+2 more|arXiv (Cornell University)|Feb 14, 2020

Autonomous Vehicle Technology and SafetyEngineering50 references52 citations

TL;DR

Social-WaGDAT presents an interaction-aware multi-agent trajectory predictor using a Wasserstein-based generative model and a graph double-attention network to capture spatio-temporal relations, validated on pedestrians and vehicles with a kinematic constraint layer.

ABSTRACT

Effective understanding of the environment and accurate trajectory prediction of surrounding dynamic obstacles are indispensable for intelligent mobile systems (like autonomous vehicles and social robots) to achieve safe and high-quality planning when they navigate in highly interactive and crowded scenarios. Due to the existence of frequent interactions and uncertainty in the scene evolution, it is desired for the prediction system to enable relational reasoning on different entities and provide a distribution of future trajectories for each agent. In this paper, we propose a generic generative neural system (called Social-WaGDAT) for multi-agent trajectory prediction, which makes a step forward to explicit interaction modeling by incorporating relational inductive biases with a dynamic graph representation and leverages both trajectory and scene context information. We also employ an efficient kinematic constraint layer applied to vehicle trajectory prediction which not only ensures physical feasibility but also enhances model performance. The proposed system is evaluated on three public benchmark datasets for trajectory prediction, where the agents cover pedestrians, cyclists and on-road vehicles. The experimental results demonstrate that our model achieves better performance than various baseline approaches in terms of prediction accuracy.

Motivation & Objective

• Motivate safe and high-quality planning for autonomous systems in crowded, interactive environments.
• Develop a generic, generativeTrajectory Prediction framework that explicitly reasons about inter-agent relations.
• Incorporate a dynamic graph representation with topological and temporal attention to model high-order interactions.
• Enforce physical feasibility through a kinematic constraint layer to improve realism and performance.
• Demonstrate state-of-the-art predictive accuracy on multiple pedestrian and vehicle trajectory benchmarks.

Proposed method

• Use a three-part pipeline: deep feature extractor (state, relation, and context features), graph-based encoder with a graph double-attention network (GDAT), and a decoder with a kinematic constraint layer.
• Construct spatio-temporal graphs for history and future, where node attributes combine agent states and context while edge attributes encode spatial relations.
• Apply a topological attention layer to update node features based on graph structure, followed by a temporal attention layer to summarize information across time.
• Encode historical and future node attributes into a latent representation via an encoding function, enabling sampling of future trajectory hypotheses.
• Decode by a recurrent-based mechanism (GRU) with a kinematic bicycle model-inspired constraint cell to ensure physically feasible trajectories.
• Train end-to-end with a Wasserstein-based objective that blends reconstruction, KL divergence, and maximum mean discrepancy terms.

Experimental results

Research questions

• RQ1 How can explicit relational reasoning between multiple agents improve trajectory prediction in interactive scenarios?
• RQ2 Can a graph-based, double-attention mechanism capture high-order inter-agent influences and temporal evolution effectively?
• RQ3 Does incorporating scene context and kinematic constraints yield measurable gains in prediction accuracy for pedestrians and vehicles?
• RQ4 How does a Wasserstein-based generative framework compare to other generative approaches (GAN/VAE variants) for multi-agent trajectory forecasting?
• RQ5 What is the impact of joint modeling of history and context on long-horizon forecasting performance?

Key findings

• Social-WaGDAT achieves state-of-the-art or competitive ADE/FDE on pedestrian benchmarks (ETH/UCY, SDD) and vehicle benchmarks (ID) across multiple scenarios.
• The graph double-attention mechanism enables effective modeling of both spatial relationships and temporal evolution, yielding notable gains over baselines.
• Incorporating context information (occupancy maps and velocity fields) significantly improves long-horizon predictions, especially in road-rule constrained scenarios.
• The kinematic constraint layer consistently improves prediction quality by enforcing feasible trajectories and smoothing outputs, with larger gains in turning-rich scenarios (roundabouts/intersections).
• The method demonstrates strong performance across pedestrians, cyclists, and on-road vehicles, demonstrating generalization to heterogeneous agents.