A Mobile Robot for Autonomous Scene Capture and Rendering

In recent years, projects such as Google Street View have contributed to widespread metrology of open human environments. Despite the ubiquity of such projects, imagecapturing remains a labor-intensive process. In this work, we apply a new image-based heuristic to capture indoor scenes with an autonomous mobile robot. We task the robot to take a set of photographs that is sufficient to build a relatively complete 3D model of its environment. Our key insight is that the completeness of the the map, at any given viewpoint, is closely correlated with the quality of a synthesized image. Our heuristic exploits this relationship to choose a strategic set of new locations for the robot to visit. Because our heuristic is based on a realtime rendering algorithm, it is easily accelerated on the GPU, which allows us to evaluate a large set of candidate locations quickly. Our algorithm starts with an initial set of photos obtained through random exploration, which may not show a large part of the scene. We then compute a sparse 3D point cloud, and randomly pick a number of candidate viewpoints in the region bounded by the 3D points. Finally, we select the best viewpoints according to our metric, and send the robot to those locations to gather more photos. The process can be repeated, until no unexplored regions remain. When the exploration is finished, we construct a complete 3D model of the scene, which we use to create a virtual tour.