In image registration or image matching, the feature extracted by using the traditional methods does not include the depth information which may lead to a mismatch of keypoints. In this paper, we prove that when the camera moves, the ratio of the depth difference of a keypoint and its neighbor pixel before and after the camera movement approximates a constant. That means the depth difference of a keypoint and its neighbor pixel after normalization is invariant to the camera movement. Based on this property, all the depth differences of a keypoint and its neighbor pixels constitute a local depth-based feature, which can be used as a supplement of the traditional feature. We combine the local depth-based feature with the SIFT feature descriptor to form a new feature descriptor, and the experimental results show the feasibility and effectiveness of the new feature descriptor.

No abstract