Change Detection in Unmanned Aerial Vehicle Images for Industrial Infrastructure Rooftop Monitoring

Abstract

This work proposes a change detection algorithm based on aerial images from an UAV (Unmanned Aerial Vehicle) for infrastructure rooftop monitoring. Taking pixel-wise differences simply between different temporal aerial images for the same scene may cause many false positives because of (a) the residual misalignments even after an image alignment preprocess; and (b) shadow and lightness differences in the images. To address such inherent characteristics of UAV images that are unfavourable in image comparing, we propose to use (a) a colour distance metric based on a weighted LAB space, which can not only mitigate the shadow effects but also enable users to be flexibly involved in detecting specific-coloured objects; (b) a Gaussian-blur filtering to emphasize major changes, while neutralising subtle changes; and (c) long-edge removal and local refinement process to reduce major false positives caused by the residual misalignments. We also conduct a mock-up experiment at a real infrastructure plant to evaluate the proposed method in terms of detecting smoke, liquid leakage, and cracking ducts, which are major phonomena of industrial malfunctions and accidents. The results show that the algorithm is able to spot the smoke and liquid leakage. On the contrary, detecting cracks is found to be not straightforward as they are viewed relatively small to be detected at the drone-filming height. We also discuss how to overcome this limitation and suggest potential approaches to improve the proposed algorithm further.