MHFS-FORMER: Multiple-Scale Hybrid Features Transformer for Lane Detection

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Lane detection is crucial for autonomous driving but remains difficult in complex scenarios like poor lighting or damaged markings.
• Existing methods often require complex post-processing and strong prior knowledge, limiting their real-world applicability.

Purpose of the Study:

• To develop an end-to-end Transformer-based model for accurate and efficient lane detection.
• To address limitations of current algorithms in handling complex driving environments and reduce reliance on post-processing.

Main Methods:

• Proposed MHFS-FORMER, an end-to-end Transformer model inspired by DETR.
• Introduced MHFNet for fusing multi-scale features with a Transformer Encoder.
• Developed a multi-reference deformable attention module for improved representation and global context awareness.
• Designed ShuffleLaneNet to explore channel and spatial information for enhanced target recognition.

Main Results:

• Achieved 96.88% accuracy on the TuSimple dataset.
• Reached a real-time processing speed of 87 frames per second (fps) on the TuSimple dataset.
• Obtained an F1 score of 77.38% on the CULane dataset.
• Demonstrated superior performance compared to existing CNN-based and Transformer-based methods.

Conclusions:

• MHFS-FORMER effectively handles complex lane detection scenarios.
• The model achieves high accuracy and real-time performance, outperforming prior approaches.
• The proposed architecture offers a robust and efficient solution for autonomous driving systems.