Quantitative Performance Evaluation Algorithms for Pavement Distress Segmentation

Abstract

Algorithms for pavement distress image segmentation are crucial to developing an automatic pavement distress detection and classification system. Many algorithms for pavement distress segmentation have been developed in the past decade; however, the lack of good methods to evaluate their performance quantitatively hinders the focused development of better segmentation algorithms. In this paper, a novel method is developed to quantitatively evaluate the performance of different pavement distress segmentation algorithms. This method uses the buffered Hausdorff distance to estimate the deviation of the cracks in the automatically segmented image from the ground truth cracks. The proposed method captures the local effectiveness of segmentation methods around the crack region without compromising its robustness to isolated pixel deviations caused by noise. Besides real pavement images, synthetic images simulating extreme pavement distress conditions are used to evaluate the capability of the proposed method and show its merits. The proposed method outperforms four other possible quantification methods and demonstrates its superior capability in providing a better score separation to distinguish the performance of different segmentation algorithms.