Photogrammetry Data for VTTI Bridge (CIAMTIS)

The study completed sought to develop a framework by which discrete images, such as those collected during standard bridge inspections, could be localized and registered against a 3D photogrammetric model of the bridge infrastructure. By doing so, more detailed spatial documentation could be achieved, avoiding common issues related to ambiguity between inspection notes and their corresponding location on a bridge.For this study, a bridge located within the Virginia Tech Transportation Institute’s (VTTI) Virginia Smart Roads (VSR) complex was selected to demonstrate the technologies and methods required to localize and register inspection images on a three dimensional (3D) model. The bridge, located on VSR’s highway section, spans a rail right-of-way owned and operated by CSX Transportation and is of a multi-girder composite construction. The bridge deck, which carries two lanes of VSR highway section traffic, as well as the bridge piers and pier caps, are of a concrete construction, while the supporting girders are made of steel. The bridge is composed of three spans: a central main span of 150 feet in length flanked at either end by side spans, each of 90 feet in length, connecting to the roadway.Photos of the bridge were collected using a DJI Mavic Enterprise Pro unmanned aerial vehicle (UAV) which flew several paths around the bridge, as illustrated in Figure X. The images were grouped into three categories, corresponding with the three phases of inspection described in the objectives section:Reconstruction Nadir Images: These were obtained from a fast initial overhead (nadir) drone scan of the bridge. The done path was a manually defined as a linear grid, which was simple to program. The images were used to generate a coarse PC model of the bridge that would take roughly 10 to 20 minutes to reconstruct and could be used to compute an optimize path for a more refined scan.Reconstruction Detailed Images: These were obtained from an algorithmically defined path, and generally covered the side of the bridge at a closer range. The path was computed by a custom algorithm to fill in gaps in the coarse model and create what we refer to as the refined model.Query Images: These images simulated an arm’s length and snooper truck inspection which yields close-up photos of bridge details or defects. These were obtained either by a cell phone on a boom or from close up drone images. These images are not shown in Figure X as they did not contribute to the reconstruction. They were instead localized on the refined model, with results shown in the results section.From the reconstruction images, COLMAP, an open-source software used to create photogrammetric processing pipelines, was used to generate two subsets of 3D models of the bridge: the first using the original full-resolution images, and the second using reduced resolution images. The reduced resolution images were downsampled to 1/4th of the original size, with an additional contrast enhancement filter. This created a more light-weight low-resolution model that took less time to generate, but was less detailed. Within each subset (full and reduced resolution), two further distinct models were explored, one generated using an OpenCV feature set and one generated using a Radial feature set, resulting in a total of four distinct models against which query images were localized and registered.Once the models were generated, COLMAP was used to localize and register both drone and cell phone query images against each of the four models. The results of this process included a determination as to whether each individual query passed was localized and registered, as well as the estimated location and orientation of localized and registered queries within the point cloud. The significant indicator of performance was whether an image was successfully localized or not, rather that the positioning error of the localization itself, as this tended to be accurate enough for the practical purpose of anchoring inspection data.