StaBle-MambaNet: structure-aware and blur-guided lane detection with Mamba

Area of Science:

• Computer Vision
• Autonomous Driving Systems
• Deep Learning

Background:

• Autonomous driving perception systems struggle with blurred images from high-speed motion and complex lighting, degrading lane detection accuracy and robustness.
• Traditional two-stage methods (enhancement then recognition) are inefficient and cannot reliably distinguish blurring from genuine lane disappearance.

Purpose of the Study:

• To develop a robust lane detection method that addresses image blurring without full-image restoration.
• To improve the accuracy and stability of lane detection in challenging autonomous driving scenarios.

Main Methods:

• Proposed Inter-frame Stability-Aware Blur-enhanced Mamba Network (StaBle-MambaNet) identifies blurred regions and assesses lane structures locally.
• Employs a Structure-Aware Restoration Module for directional extrapolation and a Blur-Guided Consistency Reasoning Module for stability evaluation.
• Utilizes a lightweight Mamba model to process enhanced features, capturing dynamic variations and preserving structural evolution.

Main Results:

• StaBle-MambaNet significantly outperforms existing methods on public datasets like CULane and CurveLanes.
• Demonstrates superior performance under challenging conditions including nighttime, occlusion, and curved lanes.
• Achieves enhanced detection accuracy and structural stability compared to current mainstream approaches.

Conclusions:

• StaBle-MambaNet offers a more efficient and accurate solution for lane detection in blurred autonomous driving images.
• The proposed method enhances the robustness of perception systems, crucial for safe autonomous navigation.