One stage detector (RetinaNet)-based crack detection for asphalt pavements considering pavement distresses and surface objects

• Published in: Journal of civil structural health monitoring Vol. 11; no. 1; pp. 205 - 222
• Main Authors: Tran, Van Phuc, Tran, Thai Son, Lee, Hyun Jong, Kim, Ki Deok, Baek, Jongeun, Nguyen, Thanh Tu
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01-02-2021; Springer Nature B.V
• Subjects: Asphalt pavements; Categories; Civil Engineering; Control; Cracks; Digital cameras; Digital imaging; Dynamical Systems; Engineering; Fatigue cracks; Fatigue failure; Flaw detection; Machine learning; Measurement Science and Instrumentation; Model accuracy; Object recognition; Original Paper; Pixels; Training; Vibration; Deep learning; Pavement distress; Automated distress detection; Crack detection; Faster R-CNN; RetinaNet
• ISSN: 2190-5452; 2190-5479
• DOI: 10.1007/s13349-020-00447-8

In this study, a supervised machine learning network model is proposed to detect and classify various types of cracks developed in asphalt pavements, including lane markers. Crack images captured from a digital camera are classified into nine categories following the pavement distress identification manual proposed by the Federal Highways Administration (FHWA). These categories are three different types of cracks, such as fatigue, longitudinal, and transverse cracks with three severity levels of the low, medium, and high for each crack type. To establish a training dataset for crack detection, 1000 images with the original size of 3704 × 10,000 pixels are divided into 20,000 smaller images of 1852 × 1000 pixels image size. The training images are labeled based on the nine categories and trained using an updated version of faster R-CNN called RetinaNet. The trained network model is validated using pavement surface images obtained from 2400 m of two road sections. It is observed from the validation study that the detection and classification accuracy of the trained network model is 84.9% considering both the crack type and severity level. When considering the crack type only, the detection accuracy of the network model is 89.1%.