Lane Attribute Classification Based on Fine-Grained Description

Area of Science:

• Computer Vision
• Artificial Intelligence
• Autonomous Driving Systems

Background:

• Road traffic marking detection is crucial for vehicle environment perception.
• Current lane detection methods lack fine-grained attribute detection, focusing only on location and overall attributes.
• Intelligent vehicles require dynamic attribute detection for enhanced road environment understanding.

Purpose of the Study:

• To develop a fine-grained attribute detection method for lane lines (Lane-FGA) to meet the needs of intelligent vehicles.
• To improve the dynamic attribute detection of lane lines and provide more comprehensive road environment information.
• To address the limitations of existing lane detection algorithms in urban environments.

Main Methods:

• Constructed a fine-grained attribute detection method using pixel-level attribute sequence points.
• Developed a lane dataset with both instance and fine-grained attribute information via manual and intelligent annotation.
• Designed a cyclic iterative attribute inference algorithm to handle occluded or damaged lane areas.

Main Results:

• The proposed Lane-FGA method achieves an average accuracy of 97% in various lane attribute detection tasks.
• The pixel-level approach effectively describes complete attribute distribution and matches lane detection results.
• The developed dataset and inference algorithm successfully addressed annotation challenges.

Conclusions:

• The Lane-FGA method provides a significant advancement in fine-grained lane attribute detection for intelligent vehicles.
• The approach enhances road environment perception by enabling dynamic attribute judgments at different segment positions.
• This work contributes a valuable dataset and robust algorithm for improving autonomous driving safety and performance.