| 摘要 |
PROBLEM TO BE SOLVED: To estimate a recognition posture parameter by removing an outlier from the comparison result of feature quantity. SOLUTION: A predetermined number of pairs are selected at random from a candidate corresponding feature point pair group to be obtained by comparing preliminarily stored feature quantity Vj<SP>[i]</SP>of all the feature points of all model moving images with feature quantity W<SB>k</SB>of the k-th feature point of an input moving image, and a three-dimensional affine transformation parameter is calculated, and projected to a parameter space, and the three-dimensional affine transformation parameter space is clustered by an NN method. Then, projection to the parameter space and clustering is repeated until predetermined conditions are satisfied, and a corresponding feature point pair group as the member of a cluster Zmax having the largest number of members is defined as a recognition corresponding feature point pair group for each model action, and the centroid Cmax of the cluster Zmax is defined as a recognition posture parameter. This invention may be applied to a recognition processor and a personal computer or the like. COPYRIGHT: (C)2009,JPO&INPIT
|
