Learning-based road crack detection using gradient boost decision tree

Abstract

Crack detection is an essential part of road maintenance systems and has attracted growing attentions in recent years. This paper proposes a learning-based road crack detection approach, which takes the fixed size image patches as training sample, and computes a number of gradient, color and texture features as an input feature vector. At the same time, the paper maps binary matrix of the correspondent ground truth crack to discrete score. Then the paper adopts the gradient boosting decision tree method to train the model, hence the diverse ground truth image patch will be stored in the leaves of each tree. While applying to detection, the trained model calculates the likelihood independently of each pixel in the tested image whether it is a part of crack and combines the result into a global reasoning. Experimental results demonstrate the state-of-the-art accuracy and high efficiency of the proposed approach compared with competing methods.