A Fast and Accurate Lane Detection Method Based on Row Anchor and Transformer Structure

Area of Science:

• Computer Vision
• Artificial Intelligence
• Automotive Engineering

Background:

• Lane detection is crucial for Advanced Driver Assistance Systems (ADASs), enabling vehicles to identify lane markings and determine their position.
• Existing deep learning lane detection methods face challenges with real-time processing on limited hardware and accuracy in complex environments.

Purpose of the Study:

• To enhance a row-anchor-based lane detection method for improved speed and accuracy in complex scenarios.
• To address the limitations of current deep learning approaches in autonomous driving systems.

Main Methods:

• Leveraged a Transformer encoder-decoder structure for enhanced global feature extraction and lane line detection in intricate environments.
• Incorporated the Feature-aligned Pyramid Network (FaPN) as an auxiliary branch.
• Introduced a novel structural loss with expectation loss to refine detection accuracy.

Main Results:

• Achieved a rapid prediction speed of 129 Frames Per Second (FPS), with a single prediction time of 15.72 ms on an RTX3080.
• Attained 96.16% accuracy on the Tusimple dataset, a 3.32% improvement over the baseline method.
• Demonstrated commendable accuracy and real-time performance suitable for on-board hardware.

Conclusions:

• The enhanced lane detection method effectively addresses the speed and accuracy challenges in complex driving scenarios.
• The integration of Transformer and FaPN structures, along with novel loss functions, significantly improves lane detection capabilities for ADASs.
• The proposed approach offers a viable solution for real-time, accurate lane detection critical for safe autonomous driving.