Automated road pavement condition assessment - a deep learning approach

The assessment of road pavement condition is crucial for ensuring usability and public safety. Damage from pavement defects, such as potholes, impacts both road users and the state. A significant portion of Irish citizens have experienced vehicle damage due to potholes. Pavement inspection typically involves three main steps: data collection, defect identification, and defect assessment. While data collection is automated, defect identification and assessment are manual processes carried out by inspectors using the Pavement Surface Condition Index (PSCI) rating system in Ireland.

This project aimed to enable automatic road pavement assessment using image processing, deep learning, and machine learning techniques. It was conducted in collaboration with Pavement Management Services Ltd (PMS), a high-tech civil engineering consultancy. PMS collected extensive data using vehicles equipped with advanced imaging and sensing capabilities.

Key objectives of the project were:

1. **Automatic Defect Detection:** Image processing and object detection techniques were used. Image processing methods like cropping, filtering, feature extraction, edge detection, and segmentation optimized images for object detection. A deep learning model, Faster-RCNN, improved the accuracy of detecting various pavement defects.

2. **Testing on Imagery:** The approach was tested on diverse imagery, including Laser Crack Measuring System (LCMS) and LiDAR images. High-definition video data from drones were also evaluated.

3. **Optimal Image Combination:** The project determined optimal image combinations and views for automated defect detection, considering factors like lighting and camera angles.

4. **Automatic Defect Rating:** A machine learning algorithm was developed to automatically generate PSCI ratings from human-generated data.

5. **Extension to US Imagery:** The project extended the approach to US-generated pavement imagery, adapting models for differences between US and Irish pavement distress types.

This project advanced innovative software solutions for automatic pavement assessment, enhancing research impact and identifying intellectual property. Collaboration with CeADAR strengthened the project through expertise in data analysis and computer vision.