Relaxation labeling in stereo image matching

Citation
G. Pajares et al., Relaxation labeling in stereo image matching, PATT RECOG, 33(1), 2000, pp. 53-68
Citations number
72
Categorie Soggetti
AI Robotics and Automatic Control
Journal title
PATTERN RECOGNITION
ISSN journal
00313203 → ACNP
Volume
33
Issue
1
Year of publication
2000
Pages
53 - 68
Database
ISI
SICI code
0031-3203(200001)33:1<53:RLISIM>2.0.ZU;2-F
Abstract
This paper outlines a method for solving the global stereovision matching p roblem using edge segments as the primitives. A relaxation scheme is the te chnique commonly used by existing methods to solve this problem. These tech niques generally impose the following competing constraints: similarity, sm oothness, ordering and uniqueness, and assume a bound on the disparity rang e. The smoothness constraint is basic in the relaxation process. We have ve rified that the smoothness and ordering constraints can be violated by obje cts close to the cameras and that the setting of the disparity limit is a s erious problem. This problem also arises when repetitive structures appear in the scene (i.e. complex images), where the existing methods produce a hi gh number of failures. We develop our approach from a relaxation labeling m ethod ([1] W.J. Christmas, J. Kittler, M. Petrou, structural matching in co mputer vision using probabilistic relaxation, IEEE Trans. Pattern Anal. Mac h. Intell. 17(8)(1995) 749-764), which allows us to map the above constrain ts. The main contribution is made, (1) by applying a learning strategy in t he similarity constraint and (2) by introducing specific conditions to over come the violation of the smoothness constraint and to avoid the serious pr oblem produced by the required fixation of a disparity limit. Consequently, we improve the stereovision matching process. A better performance of the proposed method is illustrated by comparative analysis against some recent global matching methods. (C) 1999 Pattern Recognition Society. Published by Elsevier Science Ltd. All rights reserved.