[Paper Review] DAG-Net: Double Attentive Graph Neural Network for Trajectory Forecasting

Alessio Monti, Alessia Bertugli+2 more|arXiv (Cornell University)|May 26, 2020

Autonomous Vehicle Technology and SafetyEngineering35 references58 citations

TL;DR

DAG-Net introduces a double attentive graph neural network and a recurrent VAE conditioned on future goals to produce multi-modal trajectory predictions in urban and sports scenarios, achieving state-of-the-art results.

ABSTRACT

Understanding human motion behaviour is a critical task for several possible applications like self-driving cars or social robots, and in general for all those settings where an autonomous agent has to navigate inside a human-centric environment. This is non-trivial because human motion is inherently multi-modal: given a history of human motion paths, there are many plausible ways by which people could move in the future. Additionally, people activities are often driven by goals, e.g. reaching particular locations or interacting with the environment. We address the aforementioned aspects by proposing a new recurrent generative model that considers both single agents' future goals and interactions between different agents. The model exploits a double attention-based graph neural network to collect information about the mutual influences among different agents and to integrate it with data about agents' possible future objectives. Our proposal is general enough to be applied to different scenarios: the model achieves state-of-the-art results in both urban environments and also in sports applications.

Motivation & Objective

• Address multi-modal trajectory forecasting by modeling diverse future paths conditioned on agents' goals.
• Incorporate social interactions among agents through graph neural networks with attentive pooling.
• Leverage a recurrent variational autoencoder conditioned on future objectives to generate plausible future displacements.
• Capture both future goals and mutual influences to improve long-term prediction accuracy.

Proposed method

• Use a VRNN core to model temporal dynamics and generate displacements as latent variables.
• Introduce a Goals network that encodes and refines predicted future objectives via a graph attention mechanism.
• Apply a second Attentive GNN to refine agents' hidden states with neighborhood information.
• Condition encoder, prior, and decoder on the predicted goals to guide trajectory generation.
• Train with an extended ELBO objective that includes a cross-entropy term for goal alignment.
• During inference, employ a goal generation network that shares information with other agents’ dispositions to produce coherent futures.

Experimental results

Research questions

• RQ1 How can future goals be represented and integrated to guide multi-agent trajectory prediction?
• RQ2 Can attentive graph neural networks effectively model both inter-agent interactions and goal relationships to improve accuracy?
• RQ3 Does conditioning a VRNN on structured goals yield better multi-modal predictions in urban and sports settings?

Key findings

• DAG-Net achieves stronger ADE and FDE performance than baselines on Stanford Drone Dataset (0.53 m ADE, 1.04 m FDE).
• In basketball, DAG-Net improves offensive and defensive predictions, outperforming STGAT, Social-Ways, and Weak-Supervision in ADE/FDE.
• Ablation studies show benefits from both agent interactions and future goal conditioning over vanilla VRNN and Attentive-VRNN baselines.
• Longer-horizon predictions benefit from the integrated goals-and-interactions approach, especially in highly dynamic team sports contexts.
• The dual-attention mechanism stabilizes predictions while preserving multi-modality in future trajectories.