Visual Detection of Road Cracks for Autonomous Vehicles Based on Deep Learning.

Area of Science:

• Civil Engineering
• Computer Science
• Artificial Intelligence

Background:

• Road crack detection is crucial for concrete pavement integrity assessment.
• Traditional methods struggle with noisy surfaces and real-world conditions, impacting autonomous vehicle safety.
• Developing robust, automated crack detection is essential for infrastructure maintenance.

Purpose of the Study:

• To introduce an advanced image-based road crack detection method.
• To combine Random Forest with deep convolutional neural networks (CNNs) for improved accuracy.
• To evaluate the performance of state-of-the-art CNN models in identifying concrete pavement cracks.

Main Methods:

• Utilized three deep CNN models: MobileNet, InceptionV3, and Xception.
• Trained models on a dataset of 30,000 images to develop an effective crack detection system.
• Optimized model performance by systematically comparing validation accuracy across different base learning rates, identifying 0.001 as optimal.

Main Results:

• Achieved a maximum validation accuracy of 99.97% with an optimal base learning rate of 0.001.
• Evaluated trained models on 6,000 unseen test images (224x224 pixels).
• Demonstrated outstanding test performance with 99.95% accuracy, 99.95% precision, 99.94% recall, and 99.94% F1 Score.

Conclusions:

• The proposed hybrid approach effectively detects road cracks on real concrete surfaces.
• The deep CNN models, particularly when optimized, offer a robust and flexible solution for pavement inspection.
• This technique holds significant promise for enhancing road safety and maintenance through automated crack identification.