Ultra Fast Deep Lane Detection With Hybrid Anchor Driven Ordinal Classification

Area of Science:

• Computer Vision
• Artificial Intelligence
• Autonomous Driving Systems

Background:

• Current lane detection methods often rely on pixel-wise segmentation, facing limitations in efficiency and robustness against occlusions and poor lighting.
• Human lane recognition effectively utilizes contextual and global information, especially in adverse conditions.

Purpose of the Study:

• To develop a novel, efficient, and robust lane detection method addressing limitations of existing approaches.
• To achieve ultra-fast speeds and improved performance in challenging scenarios like severe occlusions and extreme lighting.

Main Methods:

• Formulating lane detection as an anchor-driven ordinal classification problem using global features.
• Representing lanes with sparse coordinates on hybrid (row and column) anchors.
• Utilizing the large receptive field property inherent in the ordinal classification approach.

Main Results:

• The proposed method significantly reduces computational costs through anchor-driven representation.
• It demonstrates state-of-the-art performance in both speed and accuracy across four datasets.
• A lightweight version achieves over 300 frames per second (FPS).

Conclusions:

• The anchor-driven ordinal classification approach offers a simple yet effective solution for fast and accurate lane detection.
• This method shows strong potential for real-time applications in autonomous driving, particularly in challenging environmental conditions.