A Real-Time Mobile Robotic System for Crack Detection in Construction Using Two-Stage Deep Learning.

Area of Science:

• Civil Engineering
• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Civil infrastructure deterioration presents significant public safety risks.
• Manual inspections are subjective, labor-intensive, and limited by accessibility.
• Automated inspection systems are needed to overcome these limitations.

Purpose of the Study:

• To develop and evaluate a real-time robotic inspection system integrating deep learning and autonomous navigation.
• To enable simultaneous automated crack detection and collision-free navigation for infrastructure assessment.
• To improve the efficiency, accuracy, and safety of civil infrastructure inspections.

Main Methods:

• A two-stage deep learning neural network: U-Net for segmentation and Pix2Pix conditional generative adversarial network (GAN) for refinement.
• Adversarial residual learning to enhance boundary accuracy and reduce false positives.
• Deployment on an Unmanned Ground Vehicle (UGV) with RGB-D camera and LiDAR for perception and navigation.

Main Results:

• The two-stage model achieved a mean Intersection over Union (mIoU) of 73.9 ± 0.6% and an F1-score of 76.4 ± 0.3% on the CrackSeg9k dataset.
• The robotic system successfully detected micro-cracks as small as 0.3 mm in various validation tests.
• Demonstrated robust performance in simulation, laboratory experiments, and real-world campus hallway tests.

Conclusions:

• The proposed robotic system offers a robust, autonomous solution for field-deployable infrastructure inspection.
• Deep learning integration significantly enhances automated crack detection capabilities.
• The system has the potential to revolutionize infrastructure monitoring and maintenance practices.