CLRNetV2: A Faster and Stronger Lane Detector

Area of Science:

• Computer Vision
• Artificial Intelligence
• Intelligent Transportation Systems

Background:

• Accurate lane detection is crucial for intelligent vehicle navigation systems.
• Current methods struggle with complex lane geometries (e.g., Y-shapes) and underutilize multi-level features.
• Integrating semantic and local features remains an under-explored area for robust lane detection.

Purpose of the Study:

• To develop a lane detection method that effectively utilizes both high-level semantic and low-level features.
• To improve the accuracy of lane detection, especially for complex and dense lane scenarios.
• To enhance the localization precision of detected lanes.

Main Methods:

• Proposed the Cross Layer Refinement Network (CLRN) integrating semantic and local feature levels.
• Introduced Fast-ROIGather for enhanced global context gathering.
• Developed the Correlation Discrimination Module (CDM) for accurate dense lane prediction.
• Implemented LineIoU loss for whole-unit lane regression and improved localization.

Main Results:

• The CLRN significantly outperforms existing state-of-the-art lane detection methods.
• Demonstrated superior performance in detecting complex and dense lane structures.
• Achieved improved localization accuracy through the novel network design and loss function.

Conclusions:

• The proposed Cross Layer Refinement Network effectively leverages multi-level features for enhanced lane detection.
• The integration of semantic and local features, along with specialized modules, addresses limitations in current lane detection techniques.
• This approach offers a significant advancement for robust and accurate lane detection in intelligent vehicles.