TA-Unet: Integrating Triplet Attention Module for Drivable Road Region Segmentation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Road segmentation is critical for autonomous driving safety and efficiency.
• Current deep learning models often yield unsatisfactory results for practical implementation.
• Accurate segmentation of drivable regions is vital for reliable autonomous vehicle navigation.

Purpose of the Study:

• To propose a novel deep learning model, TA-Unet, for high-quality drivable road region segmentation.
• To improve the accuracy and reliability of road segmentation in autonomous driving systems.
• To address the class-imbalance problem in road segmentation tasks.

Main Methods:

• Developed TA-Unet, incorporating a triplet attention module into the U-Net architecture.
• Employed a triplet branch structure within the attention module to compute attention weights.
• Experimented with various loss functions, identifying a mixed loss function for performance enhancement.
• Validated the model on the publicly available UAS dataset, comparing against baseline and state-of-the-art methods.

Main Results:

• TA-Unet achieved superior performance compared to baseline methods.
• Achieved 98.74% pixel accuracy and 97.41% mean Intersection over Union (mIoU).
• Generated clearer and more precise segmentation maps than existing models.

Conclusions:

• The proposed TA-Unet model demonstrates significant improvements in road segmentation accuracy for autonomous driving.
• The triplet attention module and mixed loss function effectively enhance segmentation quality and address class imbalance.
• TA-Unet shows strong potential for practical application in self-driving vehicle systems.