Omni-Refinement Attention Network for Lane Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Lane detection is crucial for autonomous driving perception.
• Existing methods struggle with complex conditions like adverse weather, occlusions, and curves.
• Advanced lane detection requires integrating global semantic context and local visual features.

Purpose of the Study:

• To present ORANet, an enhanced lane detection framework.
• To improve lane detection performance and stability in challenging autonomous driving scenarios.
• To introduce novel attention modules for refined feature extraction and fusion.

Main Methods:

• ORANet builds upon the CLRNet baseline.
• Introduces Enhanced Coordinate Attention (EnCA) for long-range structures and global context.
• Incorporates Channel-Spatial Shuffle Attention (CSSA) for precise local feature extraction.
• Hierarchical synergy between EnCA and CSSA for feature refinement and fusion.

Main Results:

• ORANet demonstrates superior performance stability compared to CLRNet in complex roadway scenarios.
• Achieved nearly a 3% F1 score improvement under shadow conditions.
• Validated the effectiveness of the EnCA and CSSA modules in enhancing lane detection.

Conclusions:

• ORANet offers a robust solution for reliable lane detection in autonomous driving.
• The proposed attention mechanisms effectively address limitations of existing methods.
• ORANet shows significant potential for real-world deployment in autonomous vehicles.