Recognizing Articulated Objects in Range Images Using Invariants
Abstract
Articulated targets such as tanks can have many degrees of freedom, in addition to the unknown variables of viewpoint. Recognizing such a target in an image can involve a search in a high-dimensional space that involves all these unknown variables. In this project we use invariance to reduce this search space to a manageable size. Our method avoids feature detection for improved robustness. This is done by dealing with large parts of the visible object as wholes rather than with individual features. More specifically, we define a grid on the image, and draw a sphere around each grid point. We fit a 3D quadric to the object part contained within the sphere. We then find the Euclidean invariants of the quadric and assign them to the grid point, thus obtaining an invariant representation of the image. This is done at various scales by changing the sphere radii and grid spacing. The invariant representation is than matched against ones obtained from known models at various articulations. We apply the method to range images of objects such as backhoes.
Document Details
- Document Type
- Technical Report
- Publication Date
- Feb 01, 2002
- Accession Number
- ADA408100
Entities
People
- Isaac Weiss
- Manjit Ray
Organizations
- University of Maryland