Query-Informed Multi-Agent Motion Prediction.

Area of Science:

• Computer Science
• Robotics
• Artificial Intelligence

Background:

• Autonomous vehicles require accurate motion prediction in dynamic environments.
• Vectorization methods are prevalent but often neglect crucial physical vehicle attributes like speed and heading.
• Predicting diverse future trajectories (e.g., turns) necessitates multimodal outputs.

Purpose of the Study:

• To propose QINET, a novel method for accurate multimodal trajectory prediction for all agents in a scene.
• To improve scene encoding by enhancing agent feature attributes with physical information.
• To generate diverse future trajectories using a novel decoder architecture.

Main Methods:

• Enhanced scene encoding incorporating agent speed and heading, ensuring rotational and spatial invariance.
• A decoder utilizing cross-attention and a self-learned query matrix for multimodal trajectory generation.
• Application to the Argoverse motion prediction benchmark.

Main Results:

• QINET achieves state-of-the-art performance on the Argoverse motion prediction benchmark.
• The method demonstrates fast and accurate multimodal trajectory prediction for multiple agents.
• Enhanced physical attributes improve the representation of agents in the scene.

Conclusions:

• QINET effectively addresses limitations in current motion prediction methods.
• The proposed approach enables more robust and versatile trajectory forecasting for autonomous driving.
• This work advances the capability of autonomous systems to navigate complex traffic scenarios.