Pavement crack detection and characterization is a fundamental part of road intelligent maintenance systems. Due to the high non-uniformity of cracks, topological complexity, and similar noise from crack texture, the challenge arises in this domain with automated crack detection and classification in a complex environment. In this work, an overarching framework for a universal and robust automatic method that simultaneously characterizes the type of crack and its severity level was developed. For crack detection, we propose a novel and efficient crack detection network that captures the crack context information by establishing a multiscale dilated convolution module. On this foundation, an attention mechanism is introduced to further refine the high-level features. Moreover, the rich features at different levels are fused in an upsampling module to generate more detailed crack detection results. For crack classification, a novel characterization algorithm is developed to classify the type of crack after detection. The crack segment branches are then merged and classified into four types: transversal, longitudinal, block, and alligator; the severity levels of cracks are assessed by calculating the average width and distance between the crack branches. The proposed crack detection method effectively detects crack information in a complex environment, and achieves the current state-of-the-art accuracy. Compared to manual classification results, the classification accuracy of transversal and longitudinal cracks is higher than 95%, and the classification accuracy of block and alligator is above 86%.

Introduction

Automatic detection and classification of pavement cracks is an important part of intelligent transportation systems and acts as a primary rapid analysis of pavement distresses. The implementation of a fast and accurate automatic pavement crack detection system is essential for maintaining and monitoring complex transportation networks, and is an effective way to improve the road service quality [1]. Pavement crack automatic detection and characterization systems perform three primary tasks: data acquisition, crack detection, and crack classification. With the development of mobile mapping technology and hardware storage devices, fast acquisition devices are becoming more widely used in pavement distress screening [2] as they can quickly obtain road distress data. Fig. 1(a) shows a road surface image acquisition device installed on a roof, whereas Fig. 1(b) is a pavement image taken vertically, which can be used to measure the crack location and for qualitative analysis. In recent years, a numerous experts and scholars have devoted themselves to researching automatic detection of pavement cracks, and have obtained promising research results [3], [4]. At present, the research on automatic detection of pavement cracks is roughly divided into three methods: traditional image processing methods, machine learning methods, and deep learning methods.