SR-LMamba: A lane detection model for complex scenes integrating curvelet transform with Mamba architecture

Area of Science:

• Computer Vision
• Machine Learning
• Autonomous Driving

Background:

• Existing lane detection models lack robustness and accuracy due to poor integration of semantic and geometric features.
• Real-world driving environments present complex challenges for reliable lane marking identification.

Purpose of the Study:

• To propose SR-LMamba, a novel lane detection framework addressing limitations in current models.
• To improve the accuracy, robustness, and efficiency of lane detection systems.

Main Methods:

• Developed SR-LMamba, a framework utilizing the Sketch-and-Refine paradigm.
• Introduced LMamba, a lightweight three-stage backbone combining curvelet transform and Mamba architecture.
• Incorporated the Criss-Cross Lane Association Module (CLAM) for enhanced feature interaction and polynomial regression for curve fitting.
• Designed specialized angle loss and criss-cross attention loss functions.

Main Results:

• SR-LMamba achieved a significant F1 score of 80.04%.
• The model demonstrated superior robustness across four challenging driving scenarios.
• Outperformed existing state-of-the-art models with comparable parameter sizes.

Conclusions:

• SR-LMamba offers a robust and accurate solution for lane detection.
• The proposed architecture effectively integrates high-level semantic understanding with low-level geometric features.
• Publicly released code and models facilitate further research and development.