Effective lane detection on complex roads with convolutional attention mechanism in autonomous vehicles

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Autonomous vehicles (AVs) require robust perception systems for safe operation.
• Existing lane detection models struggle with adverse conditions like extreme weather and poor road quality.
• Accurate lane detection is crucial for lane holding and departure warnings in AVs.

Purpose of the Study:

• To develop an accurate lane detection approach for autonomous vehicles operating in challenging environments.
• To address limitations of current deep learning models in detecting lanes on poor, curvy, or unmarked roads and in extreme weather.
• To propose a novel model, Lane Detection with Convolutional Attention Mechanism (LD-CAM), for enhanced lane detection.

Main Methods:

• The proposed LD-CAM model utilizes an encoder-decoder architecture.
• An enhanced Convolutional Block Attention Module (E-CBAM) is integrated to refine feature map quality.
• The model was trained and validated on diverse datasets including Tusimple, Curve Lanes, and Cracks and Potholes.

Main Results:

• The LD-CAM model achieved high performance metrics: 97.90% accuracy, 98.92% precision, 97.90% F1-Score, 98.50% IoU, and 98.80% Dice Coefficient.
• Demonstrated superior performance on both structured and defective roads under extreme weather conditions.
• Outperformed existing models in challenging scenarios not well-handled by current literature.

Conclusions:

• The LD-CAM model offers a significant advancement in lane detection for autonomous vehicles.
• It provides reliable lane detection capabilities even in adverse environmental and road conditions.
• This research contributes to the development of safer and more dependable autonomous driving systems.