TopoMLP++: Towards Strong and Scalable Lane Topology Reasoning in Autonomous Driving

Area of Science:

• Computer Vision
• Artificial Intelligence
• Autonomous Driving Systems

Background:

• Topology reasoning in autonomous driving is crucial for lane and traffic element detection.
• High-performance pipelines and cost-effective data annotation are key challenges.
• Existing methods struggle with the complexity of real-world driving scenarios.

Purpose of the Study:

• To develop a robust "first-detect-then-reason" framework for autonomous driving topology reasoning.
• To address challenges in pipeline design and topological data annotation.
• To improve the accuracy and scalability of lane topology prediction.

Main Methods:

• Introduced TopoMLP++, a framework featuring a 3D lane detector with global attention and geometry-aware enhancements.
• Integrated 2D traffic detectors augmented by YOLOv8 results.
• Developed an agent-based data engine using a large language model (LLM) for pseudo-label generation and iterative refinement.

Main Results:

• TopoMLP++ achieved state-of-the-art results on the OpenLane-V2 dataset.
• The framework was the top solution in the IEEE CVPR Autonomous Driving Challenge.
• Using 50% labeled data, the augmented TopoMLP++ reached 96% of full-data training performance.

Conclusions:

• TopoMLP++ offers a high-performance and cost-effective solution for autonomous driving topology reasoning.
• The agent-based data engine significantly reduces annotation effort while maintaining high performance.
• This approach demonstrates the potential for scalable and accurate perception systems in autonomous vehicles.