The algorithm uses information inherent in a single line correspondence (position, orientation, and scale) to reduce the number of correspondences that must be examined in order to find an approximately correct pose. For m model lines and n image lines, the number of poses hypothesized is O(mn) [where “O(x)” signifies “of the order of x”]. This number stands in contrast to the generally much larger numbers [O(m³n³) or O(n³)] generated in some prior algorithms. In this algorithm, starting with an approximate pose instead of a precise pose makes it possible to greatly reduce the number of poses that must be examined to ultimately find a correct precise pose.

Most of the hypothesized poses are expected to be inaccurate because most of the feature correspondences used to generate them are incorrect. Accordingly, in the second stage of the process, each hypothesized pose is ranked on the basis of similarity of the corresponding local neighborhoods of lines in the model and image. The similarity measure is largely unaffected by image clutter, partial occlusion, and fragmentation of lines. The ranking of the hypothesized poses is invariant under image translation, scaling, and rotation, and is partially invariant under affine distortion of the image. The combination of (1) the generation of hypothesized poses from image lines that are assumed to be unfragmented with (2) the neighborhood similarity measure makes it possible to quickly generate a ranked list of approximate model poses that is likely to include a number of highly ranked poses that are close to the correct model pose.

In the third stage of the process, a robust pose-estimation subalgorithm effects a subprocess of refinement and verification, starting from the few approximate poses ranked most highly in the second stage. The subalgorithm used in this stage is a modified version of a previously developed algorithm that was selected because it is efficient, is tolerant of clutter and occlusion, and does not make binary correspondence decisions until an optimal pose is found.

This work was done by Philip David of the Army Research Laboratory and Daniel DeMenthon of the University of Maryland. For more information, download the Technical Support Package (free white paper) at www.defensetechbriefs.com/tsp under the Information Sciences category. ARL-0009

This Brief includes a Technical Support Package (TSP).

Shape-Based Recognition of 3D Objects in 2D Images (reference ARL-0009) is currently available for download from the TSP library.

Login first to download.

Username

Password

Remember me


Lost Password? Register