Fast and Accurate Lane Detection via Graph Structure and Disentangled Representation Learning

Area of Science:

• Computer Vision
• Machine Learning
• Autonomous Driving Systems

Background:

• Lane detection is crucial for autonomous driving but computationally intensive.
• Existing methods struggle to balance accuracy and runtime efficiency due to complex feature extraction.
• Information loss during feature compression hinders performance.

Purpose of the Study:

• To develop an efficient and accurate lane detection method.
• To address the computational challenges in feature extraction for lane detection.
• To mitigate information loss in feature compression.

Main Methods:

• A two-phase feature extraction process: local feature extraction using anchor lines and global feature aggregation via graph networks.
• A novel feature compression module employing decoupling representation learning to preserve critical information.
• Adaptive learning of node distances for weighted summing in global feature aggregation.

Main Results:

• Achieved high accuracy with F1 scores of 96.81% on Tusimple and 75.49% on CULane benchmarks.
• Demonstrated a fast running speed of 248 FPS, indicating significant runtime efficiency.
• The proposed feature compression module effectively retained statistical and spatial feature information.

Conclusions:

• The proposed lane detection method offers a superior balance of speed and accuracy.
• The novel feature extraction and compression techniques are effective for real-time lane detection.
• This method has strong potential for enhancing the safety and performance of autonomous vehicles.