Fine-grained damage detection of cement concrete pavement based on UAV remote sensing image segmentation and stitching

Abstract

A hybrid model for pavement damage segmentation, stitching, and detection based on unmanned aerial vehicle (UAV) remote-sensing pavement images was proposed as high-resolution panoramic images are required to analyze pavement damage feature parameters during the inspection process. Based on target recognition, active stereo vision was used to determine the autonomy awareness of the UAV system in the 3D structure. Then, incremental search strategy was employed to triangulate scale-invariant feature transform (SIFT) point features based on geometric constraints. Perspective-n-point and random sample consensus (PNP-RANSAC) algorithm were combined to find the best matching surface for mis-matching rejection, and the trigonometric function theory was used to fuse images. Subsequently, the semantic segmentation framework based on deep learning context encoder network (CE-Net) was employed to train suitable models to simultaneously detect and segment pavement damages. Finally, the stitching technique and hue, saturation, value (HSV) segmentation technique were combined to obtain the feature parameters to be detected. This method achieved fast and accurate stitching, feature parameter segmentation, and pavement health condition assessment of long-threaded pavement disease image datasets. The splicing accuracy rate and the global feature segmentation overlap rate were greater than 84% and 80%, respectively. The experimental data show that the method is able to detect full-width pavement breakage parameters better than existing optical sensor-based automotive or drone inspection methods.