Feature Matching for Omnidirectional Image based on Singular Value Decomposition

The omnidirectional image sensor has many advantages of the branches of map-building, navigation, and localization for mobile robot and the concern with the sensor has been growing. When the omnidirectional image sensor is used to various application, the feature matching between two images is the essential problem that always occurs at the first step of an image processing. Although several studies have been made on the problem, there are still some drawbacks which limit in the motion of a mobile robot or which are only valid under certain assumption, that a the mobile robot should move slowly. This paper discusses a issue related to find the correspondence of features in two omnidirectional images regardless of the motion of the mobile robot. We propose an feature matching method for omnidirectional image that modifies the previous method based on the singular value decomposition. Instead of the zero-mean normalized correlation used in the traditional approach, the sum of the squared-difference with a Gaussian weight is introduced. It describes the similarity of the matched pairs in omnidirectional images. To show the feasibility of the proposed method an experiment is applied to 100 indoor omnidirectional images. The experimental results show that the proposed method works well even when some features in one image do not have corresponding features in the other image or when the start position of the matched features is not known. Despite the well-known combinatorial complexity of the problem, this method shows that a good solution can be obtained directly by the singular value decomposition of an appropriate proximity matrix. The processing time is suitable for the real mobile robot.