| 摘要 |
PROBLEM TO BE SOLVED: To improve accuracy and speed of feature point matching between an observation target and a camera image thereof without allowing a calculation cost or a usage memory volume to increase.SOLUTION: A feature point detection part 31 detects a number of feature points from a camera image Ica and outputs them with local feature information. A marching part 32 applies an individual feature point and its local feature information to a correspondent point sorter 32a and outputs an individual correspondent point and its likelihood of an observation image corresponding to the respective feature point. A viewpoint estimation part 33 applies the corresponding point and its likelihood to a viewpoint sorter 33b and estimates a viewpoint of the camera image. A probability distribution database 33c of the viewpoint sorter 33b is built by applying a corresponding point and its likelihood of an observation target or its 3D model corresponding to an individual feature point detected from an individual projection image produced by projecting the observation target or its 3D model in different viewpoints, to the viewpoint sorter 33b with a viewpoint of the projection image handled as a teacher label, so that the database 33c provides inputs of plural corresponding points and their degrees of likelihood with probabilities of viewpoints. |
