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

Abstract

Lane detection seeks to accurately identify the position and geometry of lane markings in real-world driving environments. However, existing models often struggle with robustness and accuracy due to insufficient integration of high-level semantic understanding and low-level geometric features. To tackle these limitations, we propose SR-LMamba, a novel lane detection framework built upon the Sketch-and-Refine paradigm of SRLane. At the core of our approach lies LMamba, a lightweight three-stage backbone network that accelerates inference while effectively capturing both geometric structures and sequential patterns through a synergistic combination of curvelet transform and the Mamba architecture. In the Refine stage, we introduce the Criss-Cross Lane Association Module (CLAM), which employs a multi-lane criss-cross attention mechanism to enhance feature interactions and applies polynomial regression to refine lane curve fitting. To further boost performance, we design tailored loss functions—angle loss and criss-cross attention loss—aligned with the model architecture. Experimental results show that SR-LMamba achieves an F1 score of 80.04%, outperforming current state-of-the-art models with similar parameter sizes, and demonstrating superior robustness across four challenging driving scenarios. In addition, we publicly release our code and models at https://github.com/chenml1/SR-LMamba .