HDGT: Heterogeneous Driving Graph Transformer for Multi-Agent Trajectory Prediction via Scene Encoding.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Autonomous driving systems require accurate representation of complex driving scenes.
• Driving scenes contain heterogeneous elements (agents, lanes, signs) with rich semantic and spatial relationships.
• Existing methods struggle with the relative and ego-centric nature of spatial information.

Purpose of the Study:

• To develop a novel backbone model for encoding driving scenes in autonomous driving.
• To effectively represent heterogeneous elements and their diverse relationships within a driving scene.
• To improve the performance of downstream tasks like trajectory prediction.

Main Methods:

• Proposing the Heterogeneous Driving Graph Transformer (HDGT) model.
• Constructing a heterogeneous graph with diverse node and edge types representing driving scene elements.
• Employing a local node-centric coordinate system for encoding relative spatial information.
• Utilizing a hierarchical transformer structure within a graph neural network (GNN) for heterogeneous data aggregation.

Main Results:

• HDGT models the driving scene as a heterogeneous graph, capturing complex inter-object relationships.
• The model successfully encodes spatial relations in an ego-centric manner.
• Achieved state-of-the-art performance on trajectory prediction benchmarks, including the INTERACTION Prediction Challenge and Waymo Open Motion Challenge.

Conclusions:

• The Heterogeneous Driving Graph Transformer (HDGT) provides a powerful framework for autonomous driving scene representation.
• HDGT's approach to heterogeneous graph construction and spatial encoding significantly enhances trajectory prediction accuracy.
• This model offers a promising advancement for intelligent vehicle perception and motion forecasting.