On the improvement of three-dimensional reconstruction from large datasets

Abstract

The advent of digital cameras heralded many possibilities of structure and shape recovery from imagery that are quickly and inexpensively acquired by such devices. Throughout the years numerous techniques have emerged, and state-of-art algorithms are now able to deliver 3D structure acquisition results from low cost sensors with quality and resolution comparable to industry standard systems such as LIDAR and expensive photogrammetric equipments. DEMs, which are intensely used in geophysics and geography subject studies, have been largely benefited from such progress. Current imaging devices capable to produce high-definition images are compact, lightweight, and can be easily attached to unmanned aerial vehicles (UAV), in contrast to other means of 3D data acquisition such as LiDAR and dedicated photogrammetric equipments, which are associated to high financial and logistical costs. However, the processing time of the collected imagery to produce a DEM quickly becomes prohibitive as the number of input images increases, demanding powerful hardware and days of processing time to generate full DEMs of large datasets containing thousands of images. In this work we propose an efficient approach based on Structure-from-Motion (SfM) and multi-view stereo reconstruction techniques to automatically generate DEM -- Digital Elevation Models -- from aerial images and also 3D models in general. Our approach, which is image-based only, uses the increasingly meta-data information such as GPS in EXIF tags to initialize our graph structure, a keypoint filtering technique to maintain high repeatability of matches across pairs and reduce the matching effort, a vocabulary tree score to reduce the space search of matching and multiple local bundle adjustment refinement instead of the global optimization in a novel scheme to speed up the incremental SfM process. The results from six large aerial datasets obtained by UAVs with minimal cost and four terrestrial datasets show that our approach outperforms current strategies in processing time, and is also able to provide better or at least equivalent results in accuracy compared to three state-of-the-art methods.