Automatic Acquisition of Hierarchical, Textured 3D Geometric Models of Urban Environments: Project Plan

This paper presents an overview of a planned system for automatic population of geospatial databases which represent urban exteriors as textured geometric model data. The salient features of the system are that 1) it employs \pose-imagery:" high-quality, digital still images, each timestamped, and annotated with a reliable estimate of the acquiring camera's absolute position and attitude; 2) it exploits both dense and sparse image features by sifting evidence for in-terpixel correlation across every image; 3) it addresses the combinatorial aspects of large-scale 3D reconstruction from thousands of arbitrary images; and 4) it has no \human in the loop." This work involves both engineering and research challenges. The engineering challenges include rapid acquisition of 6-DOF pose-instrumented, high-resolution imagery, and its insertion into a hierarchical three-dimensional data structure. The research challenges include dense reconstruction techniques using thousands of images, and the incremental construction of a textured three-dimensional model from selected spatially related image subsets. Also presented are a description of the synthetic , hybrid, and actual pose image datasets to be acquired and processed, as well as an evaluation plan for the project.