Automatic Bridge Deck Damage Using Low Cost UAV-based Images

Abstract

Bridge structures are subjected to deterioration due to excessive usage, overloading, aging, and environmental impacts. Use of visual inspection by live, on-site inspectors predominates the requisite inspection of these structures, despite the known disadvantages of subjective results, high costs, and traffic disruptions due lane closures needed for close-range inspection access. Over the last two decades, significant advancements have occurred in the field of a remote sensing for bridge inspection. Prominent amongst these are use of a point cloud based inspection derived from images collected with an unmanned aerial vehicle (UAV). The approach can rapidly acquire surface details and overcome many of the shortcomings of live, visual inspection but further processing has been required. This paper automates that method for damage inspection of bridge decks. To achieve that, this paper first proposes a robust and efficient method to automatically extract a point cloud of a bridge deck through a cell-based region growing segmentation. Next, locations and areas of the patch deterioration are automatically determined by comparing elevations of the point clouds to the surface of the undamaged bridge deck. Finally, a deep learning method, using a one-class autoencoder, is employed to classify the point cloud of the bridge deck into cracking area and undamaged one.