Efficient monocular 3D lane detection via Mamba-enhanced CM-3DLane framework.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Autonomous Driving Systems

Background:

• Monocular 3D lane detection offers superior spatial information for intelligent driving compared to 2D methods.
• Current 3D lane detection methods face challenges in accurately localizing slender lane structures and integrating multi-level features efficiently.

Purpose of the Study:

• To develop an efficient and accurate monocular 3D lane detection framework (CM-3DLane) for intelligent driving scenarios.
• To address limitations in feature integration and computational efficiency in existing 3D lane detection techniques.

Main Methods:

• Utilized a Convolutional Neural Network (CNN) backbone for local feature extraction.
• Introduced the Lane-Aware Mamba (LAMamba) block with a 2D selective scan (SS2D) for efficient long-range dependency modeling.
• Implemented a Cross-Scale Attention Fusion (CSAF) module for multi-scale feature fusion and a Refined Anchor Dynamic Ranking (RADR) module for anchor selection.

Main Results:

• Achieved state-of-the-art performance with 58.3 F1 score on OpenLane and 96.5 F1 score on ApolloSim.
• Demonstrated high computational efficiency suitable for real-time deployment.
• Significantly improved feature extraction and global lane context modeling.

Conclusions:

• The CM-3DLane framework effectively enhances 3D lane detection accuracy and efficiency.
• The proposed LAMamba block and CSAF module are key to capturing global spatial dependencies and fusing multi-scale features.
• CM-3DLane offers a promising solution for robust vehicle perception in complex driving environments.