Statistical Selection and Interpretation of Imagery Features for Computer Vision-Based Pavement Crack–Detection Systems

Abstract

This paper aims to explore the statistics of pavement cracks using computer-vision techniques. The knowledge discovered by mining the crack data can be used to avoid subjective crack feature selection for vision-based pavement evaluation systems. Moreover, the statistical evaluation of crack features can be used as fundamental data to justify pavement rehabilitation policies. For this purpose, surface images of flexible pavements in different deterioration stages were analyzed using a novel image-processing technique. Seven imagery features of the detected objects including area, length, width, orientation, intensity, texture roughness, and wheel-path position, which are commonly used in pavement applications, were extracted and analyzed. A comprehensive statistical analysis was performed using filter feature subset selection (FSS) methods to rank crack features based on their significance (relevance and redundancy) for the pavement crack–detection problem. Based on the results, length, intensity, and wheel-path position were identified as the optimal feature-set for the vision-based system. Statistical characteristics of crack features were also analyzed to extract accurate quantitative measures for pavement conditions assessment. The statistical characterization identified longitudinal cracks within the wheel path as the dominant defect of the validation data set. Such information can help management agencies make informed pavement maintenance policies.