DEVELOPMENT OF AN AUTONOMOUS ROAD SURFACE DAMAGE INSPECTION PROGRAM USING DEEP CONVOLUTIONAL NEURAL NETWORK

Abstract

This study presents the development and application of an autonomous program for road surface inspection using Laser Crack Measurement System (LCMS) images acquired by a road surface property measurement vehicle. The program automatically detects road lane line marks and makes grid mesh of analyzed parts. It applies a robust color space transformation and Hough transformation algorithms for the detection of analysis area between two lane lines. The program utilizes a deep convolutional neural network (CNN) to automatically detect crack, non-crack, and patch parts on a high-resolution road surface image. Transfer learning and fine-tuning parameters were employed during CNN model training for the image classification. The model accuracy for the classification of three classes was 97.9%. The precision for each class, crack, non-crack and patch classes were 99.7%, 97.9%, and 95.5%, respectively. The study reveals the novelty of transfer learning application, and the importance of data preparation in CNN model training for road surface image inspection. The program is automated for image importation, detection of the analysis area, analysis visualization, and summary of the crack and patch parts. It significantly improves road inspection efficiency. Compared to the inspection time of a professional engineer in 1-km road surface images with a crack ratio of 20%, the program shortens inspection time by 98% for lane line detection and marking grid mesh, and road damage judgment by 82%. Total inspection time of 1-km road is shortened by 50%. This study opens the potential application of AI on road inspection.