Multiscale structure from multiviews by ▽2G filtered three‐dimensional voting

Abstract

AbstractA hierarchical matching is the most efficient way to recognize a 3‐D object. In most conventional methods of 3‐D object recognition, a hierarchical structure is formed after the detailed 3D structure of an object has been extracted by dividing it coarsely. This approach is very inefficient, since it is difficult to obtain a detailed 3D structure initially in addition to a large amount of computation time.This paper proposes a method which extracts a multiscale 3D structure directly from coarse to fine. In this method, a space containing an object is divided into coarse voxels, and a “3D voting” is applied. This voting adds a certain value to a voxel through which a back‐projected line connects the center of the camera lens and a feature point on an image. A voxel having a high value is regarded as a region containing a 3D feature point, and this is extracted. Extraction of a hierarchical structure is carried out by repeating the dividing process from coarse voxels to fine voxels so that the operation becomes efficient.This paper also analyzes the effects of the deviation of a feature point in an image and the deviation of the camera center on the 3D voting. The paper shows also that an application of a ▽2G filter to the voxel space (after 3D voting) is effective in suppressing the errors due to the deviations. The method has been confirmed successfully by processing six complex test images.