Deep spatial attention networks for vision-based pavement distress perception in autonomous driving

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Autonomous driving safety relies on accurate road surface perception.
• Existing research prioritizes traffic participant detection over road quality assessment.
• A gap exists in robust road defect monitoring for autonomous systems.

Purpose of the Study:

• To develop a high-performance semantic segmentation method for real-time road surface defect detection.
• To enhance road defect detection accuracy using a novel multi-scale spatial attention module.
• To improve the safety and comfort of autonomous driving through better road condition awareness.

Main Methods:

• Utilized onboard camera images for real-time road defect monitoring.
• Implemented a semantic segmentation framework enhanced with a multi-scale spatial attention module.
• Trained and validated the model on a custom dataset of 2,400 annotated images.

Main Results:

• Achieved superior detection precision and computational efficiency compared to existing methods.
• Demonstrated improved performance in mean Intersection over Union (mIoU) for road defect detection.
• Maintained low computational cost and high inference speed, crucial for real-time applications.

Conclusions:

• The proposed method offers a significant advancement in vision-based road surface quality perception for autonomous driving.
• The approach provides a balance between accuracy and efficiency, suitable for onboard systems.
• Seamless integration with control strategies can enhance passenger experience by ensuring a smooth and reliable ride.