Method and program for registration of three-dimensional shape

• US 8,233,742 B2
• Filed: 05/10/2006
• Issued: 07/31/2012
• Est. Priority Date: 06/07/2005
• Status: Active Grant

• Current Assignee: National Institute of Advanced Industrial Science and Technology (Government of Japan)
• Original Assignee: National Institute of Advanced Industrial Science and Technology (Government of Japan)
• Inventors: Masuda, Takeshi
• Primary Examiner(s): Allison, Andrae S
• Application Number:
  US11/921,572
• Publication Number:
  US 20090128546A1
• Time in Patent Office: 2,274 Days
• Field of Search: 382/294, 382/295, 382/298, 382/300, 382/264, 382/254, 382/293, 382/266, 382/276, 382/277, 382/278, 382/280
• US Class Current: 382/294
• CPC Class Codes: G06T 7/35 using statistical methods
• CPC Class Codes: G06V 10/431 Frequency domain transformation; Autocorrelation

Abstract

As a feature, a local range image described in a log-polar coordinate system with a tangential plane set as an image plane is used. In a created image, ambiguity concerning an angular axis around the normal is normalized as a power spectrum using Fourier series expansion and changed to an amount invariable with respect to rotation. The power spectrum is dimensionally compressed by expanding the power spectrum in a peculiar space using a peculiar vector. Corresponding points are searched by nearest neighbor in a dimensionally compressed space to calculate a correspondence relation among the points. Wrong correspondence is removed by verification to determine a positional relation among range images. A reliable correspondence relation is narrowed down by verification by cross-correlation and a RANSAC to create a tree structure representing a link relation among the range images. A shape mode is created by applying a simultaneous registration method to plural range images of the tree structure using a result of this registration as an initial value.

Claims

1. A method for registration of a three-dimensional shape, the method comprising:

creating, based on an input range image sampled by an image sensor, a local range image described in log-polar coordinate system with a tangential plane of a point on a curved surface set as an image plane, where the log-polar coordinate system is spread around the point on the curved surface;

expanding the local range image by Fourier series expansion to calculate a power spectrum with respect to an angular axis around normal at the point on the curved surface and setting the calculated power spectrum as a feature;

searching for corresponding points having nearest features between different input range image to calculate a correspondence relation among the points; and

determining a positional relation between the different input range images on the basis of the calculated corresponding points, and,

wherein, when the correspondence relation of the feature is calculated, the power spectrum is dimensionally compressed by being expanded in a peculiar space using a peculiar vector and a search for corresponding points having nearest features is performed in a dimensionally compressed space.

2. The method for registration of a three-dimensional shape according to claim 1, wherein input range images are coarsely represented by a signed distance field to create the local range image at high speed.

3. The method for registration of a three-dimensional shape according to claim 1, wherein registration between three-dimensional shapes described in a signed distance field is enabled by creating the local range image from the signed distance field.

4. A non-transitory computer-usable storage medium having a program comprising computer readable instructions stored thereon for execution by a processor to perform a method for registration of a three-dimensional shape, the method comprising:

creating, based on an input range image sampled by an image sensor, a local range image described in log-polar coordinate system with a tangential plane of a point on a curved surface set as an image plane, where the log-polar coordinate system is spread around the point on the curved surface;

expanding the local range image by Fourier series expansion to calculate a power spectrum with respect to an angular axis around normal at the point on the curved surface and setting the calculated power spectrum as a feature;

searching for corresponding points having nearest features between different input range images to calculate a correspondence relation among the points;

determining a positional relation between the different input range images on the basis of the calculated corresponding points; and

creating a shape model by registering plural input range images using the determined positional relation among all the range images as an initial value, and

wherein, when the correspondence relation of the feature is calculated, the power spectrum is dimensionally compressed by being expanded in a peculiar space using a peculiar vector and a search for corresponding points having nearest features is performed in a dimensionally compressed space.