Pose Estimation of Three-Dimensional Objects from Single Camera Images

Eliminating False Matches for the Projective Registration of Free-Form Surfaces With Small Translational Motions

“…The reason why the FLS algorithm is chosen is that most of the existing registration algorithms [9], [24] are iterative, involving a number of parameters to be set by the user or require human intervention [27] or feature extraction [14], [21]- [23], [40], [42]. If the parameters are set properly or the structural features are accurately extracted, then the algorithms are accurate.…”

“…First, we have applied the rigid motion constraints to define the qualities of the traditional ICP criterion established possible point matches ICP criterion, as long as at least two of them satisfy the rigid motion constraints exactly, then any one of them that satisfies the rigid motion constraints exactly represents a real 3-D-2-D correspondence. This is in contrast with the feature matching based methods [14], [40], [42] that often take for granted that the finally established correspondences are real. Our study has explicitly revealed the conditions under which accurate 3-D-2- D TABLE VII RELATIVE CALIBRATION ERROR e IN PERCENTAGE OF THE CAMERA POSITIONt FOR THE ICPF ALGORITHM USING THE REAL WALL AND TOWER IMAGES WITH DIFFERENT PERCENTAGES OF APPEARING AND DISAPPEARING POINTS.…”

“…Digital Object Identifier 10.1109/TSMCB.2005.843978 [4], feature matching [14], [40], [42], stereo simplex [36], softassign [5], [6], and many others. The intrinsic natural and practical idea of the ICP algorithm has attracted much attention from the machine vision and image processing community since it was proposed in 1992 independently by several researchers [1], [3], [44].…”

“…In the E (expectation) step, the camera pose is fixed and the 3-D-2-D correspondences are established by Step 1 and Step 2 above. In the M (maximization) step, the 3-D-2-D correspondences are fixed and the camera pose is updated by Step 3 above; (2) the novel approach has an advantage of easy implementation, since it does not need to extract any structural features from images [14], [21]- [23], [40], [42] or estimate the motion parameters of interest [30], [34] from the possible point matches before they are evaluated; and finally (3) the novel approach can produce encouraging projective registration results, as demonstrated in this paper.…”

“…While the existing registration algorithms [14], [40], [42] mainly use feature matching to establish and/or evaluate the possible point correspondences, the extraction of structural features is often sensitive to noise, occlusion, and appearance and disappearance of points. The rigid motion constraints based registration algorithm, however, takes a different strategy for registration.…”