Robust 3D lane detection in complex traffic scenes using Att-Gen-LaneNet.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Robust 3D lane detection is essential for advanced autonomous driving systems.
• Complex traffic scenes, including adverse weather and uneven terrain, pose significant challenges to current lane detection algorithms.
• Existing methods often struggle with accuracy and reliability in diverse environmental conditions.

Purpose of the Study:

• To propose a generalized two-stage network, Att-Gen-LaneNet, for robust 3D lane detection.
• To enhance the performance of lane detection in complex and varied traffic scenarios.
• To improve the accuracy and reliability of autonomous driving systems through better lane recognition.

Main Methods:

• Developed Att-Gen-LaneNet, a two-stage network integrating Efficient Channel Attention (ECA) and Convolutional Block Attention Module (CBAM).
• Improved the ENet semantic segmentation network with a weighted cross-entropy loss function for ambiguous distant lane segmentation.
• Introduced an interpolation loss function in the second stage for precise lane fitting.

Main Results:

• Achieved 99.7% Pixel Accuracy and 89.5% MIoU in the first stage segmentation.
• Outperformed existing methods by 6% in F-score and Average Precision on the Apollo Synthetic dataset.
• Demonstrated superior overall performance in 3D lane detection across complex traffic scenes.

Conclusions:

• Att-Gen-LaneNet provides a robust solution for 3D lane detection in challenging environments.
• The proposed network significantly improves accuracy and reliability compared to existing methods.
• The approach is applicable to a wider range of complex traffic scenes, advancing autonomous driving capabilities.