An End-to-End Lane Detection Model with Attention and Residual Block

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Lane detection is crucial for autonomous driving but struggles with complex scenarios like missing or multiple lane lines.
• Existing methods often lack the accuracy needed for real-world autonomous driving applications.

Purpose of the Study:

• To propose a novel end-to-end lane detection model that enhances accuracy in complex driving scenes.
• To develop a computationally efficient model suitable for real-time autonomous driving.

Main Methods:

• An end-to-end lane detection model integrating attention mechanisms and residual blocks.
• A U-shaped network architecture with three downsampling operations to preserve image resolution and lane information.
• Utilizing convolutional operations enhanced by attention for sensitive feature representation of lane lines.

Main Results:

• The proposed model achieves high accuracy in lane detection on public datasets, even in complex scenes.
• The model demonstrates computational efficiency with a small weight size of 2.25M.
• The residual block addresses potential gradient issues, while the attention block focuses feature learning.

Conclusions:

• The novel lane detection model effectively addresses challenges in complex driving environments.
• The integration of attention and residual blocks within a U-shaped structure offers a promising solution for accurate and efficient autonomous driving.
• Further discussion on experimental challenges provides insights for future research.