Decision fusion of global and local image features for Markov localization

This paper addresses a major problem in the context of visual robot localization. Vision-based localization easily leads to ambiguities in large-scale environments. A probabilistic method is proposed for mobile robots to recognize scenes for topological localization. Appearance-based scene classes are automatically learned from composite features which combine global and local image features extracted from sets of training images. A modified Scale Invariant Feature Transform (SIFT) feature descriptor, which integrates color with local structure, is used as local features to disambiguate the identification of features easily confused. The environment is defined as a topological graph where each node corresponds to a place and edges are paths connecting one node with another. In the course of traveling, each detected interest point vote for the most likely location, and the correct location is the one getting the largest number of votes. In the case of perceptual aliasing, a Hidden Markov Model (HMM) is used to increase the robustness of location recognition. Experimental results show that application of the proposed feature and decision fusion can largely reduce wrong matches and the proposed method is effective.