Autonomous image localization for visual inspection of civil infrastructure

Abstract

Low-cost, high-performance vision sensors in conjunction with aerial sensing platforms are providing new possibilities for achieving autonomous visual inspection in civil engineering structures. A large volume of images of a given structure can readily be collected for use in visual inspection, overcoming spatial and temporal limitations associated with human-based inspection. Although researchers have explored several algorithms and techniques for vision-based inspection in recent decades, a major challenge in past implementations lies in dealing with a high volume of images while only a small fraction of them are important for actual inspection. Because processing irrelevant images can generate a significant number of false-positives, automated visual inspection techniques should be used in coordination with methods to localize relevant regions on the images. When combined, automated visual inspection will be able to meet the objectives and quality of human visual inspection. To enable this technology, we develop and validate a novel automated image localization technique to extract regions of interest (ROIs) on each of the images before utilizing vision-based damage detection techniques. ROIs are the portions of an image that contain the physical region of the structure that is targeted for visual interrogation, denoted as the targeted region of interest (TRI). ROIs are computed based on the geometric relationship between the collected images and the TRIs. Analysis of such highly relevant and localized images would enable efficient and reliable visual inspection. We successfully demonstrate the capability of the technique to extract the ROIs using a full-scale highway sign structure in the case where weld connections serve as the TRIs.