A Study of Feature Matching in Trinocular Vision Systems

• Main Authors: Rong-Nan Chiou, 邱榮楠
• Other Authors: Chin-Hsing Chen
• Format: Others
• Language: en_US
• Published: 1993
• Online Access: http://ndltd.ncl.edu.tw/handle/08209531212763275807

博士 === 國立成功大學 === 電機工程研究所 === 81 === Base on three-view analysis, this disseration presents object recognition and stereo analysis for a trinocular sys- tem. In object recognition, the 3D surface data, represented by corners and line drawing, and the outer boundary of image projection are used to find the exact corresponding model of an unknown input object. In the stereo analysis, feature pr- ojection location constraint is performed by the (feature- deviation)- to-(epipolar distance) transformation and the optimal epipolar constraint. The inherent matching problems associated with the spurious features produced by overlapping among objects, the periodic structure of figures and the occlusion features were solved. Because the optimal epipolar constraint is enforced before the system start to work, com- putation time is greatly saved and the ambiguous correspond- ences are profusely reduced. Stereo errors caused by the deviation from the exact projection location of a feature were estimated for various camera configurations. Based on the estimation of the stereo error, the thresholding of the least square error required for matching two sets of 3D data was then derived. The derivation result has been sucessfully applied in the polyhedron recognition for 3D feature matching. Since the stereo error and the deviation of the feature are utilized, the threshold values for 2D and 3D feature matching can be automatically calculated for different camera configuration and adaptively set for various matched feature sets. This increases the precision of the matching results and solves the troublesome problems in the threshold setting process.