A novel local invariant descriptor adapted to mobile robot vision

In the past years, recognition algorithms based on local features have received much attention; and their advantages over traditional recognition methods in dealing with similarity transformation, partial occlusion and extraneous noise, have been verified. But due to the difficulty of most current local-feature-based methods in real-time implementation, their application in mobile robot vision has been deferred. Motivated by it, this paper is expected to provide a vision system especially adapted to mobile robots based on local features. Compared with previous works, A novel local invariant descriptor combined with gradient information is provided; and accordingly a new matching method of interest points is introduced. The novel local invariant, which is characterized by Gaussian derivatives, also consists of descriptions of some non-interest points localized according to the average gradient at interest points, so that the novel local invariant descriptor achieves a high discrimination under various similarity transformation. And the new matching method greatly increases the robustness of the algorithm by employing a segmentation correlation and eliminating some possibly wrong matching results. Experimental results demonstrate our algorithm has a good recognition ability and robustness in case of rotation, partial occlusion, various similarity transformations, extraneous noise, etc. And it can be implemented in real time, which makes it very appropriate for mobile robot use.