Self-moving multi-sensor AI-based robotic technology for road crack inspection

Area of Science:

• Robotics and Artificial Intelligence
• Civil Engineering
• Computer Vision

Background:

• Human road inspections are physically demanding and subjective, impacting accuracy.
• Existing automated systems are often costly and lack integration.
• There is a need for efficient and affordable road assessment technologies.

Purpose of the Study:

• To introduce Mobi-A4Net, an affordable robotic-AI system for automatic road crack detection and assessment.
• To address limitations of current road inspection methods and automated systems.
• To develop a system suitable for real-time operation on unmanned ground vehicles (UGVs).

Main Methods:

• Developed Mobi-A4Net, an AI system utilizing a Mobile Adaptive Attention Aggregation Network on a UGV platform.
• Employed a compact deep-learning model with multi-scale attention for crack identification.
• Integrated advanced image processing, including Medial Axis Transform (MAT) skeletonization, for crack measurement.

Main Results:

• Achieved 99.7% detection accuracy for road cracks.
• Obtained a recall rate of 98.8% and a mean Intersection over Union (mIoU) of 95.4%.
• Demonstrated real-time performance with an inference speed of 9.6 milliseconds per image.

Conclusions:

• Mobi-A4Net offers an accurate and efficient solution for road crack detection and assessment.
• The system's affordability and real-time capabilities make it suitable for practical UGV applications.
• The developed AI system overcomes limitations of manual inspections and costly automated alternatives.