Entropy-based features for robust place recognition

In this paper, an appearance-based modeling of the environment is presented for the sake of mobile robot localization. The model allows perception and recognition within a topological context. Highly descriptive SIFT is used to extract local features from visual data acquired from an indoor environment. A method is developed to select those features, which are best for localization using a probabilistic modeling and an entropy measure. The impact of feature selection on the localization performance is more than 60% reduction in the storage and recognition time overhead. The methodology insures the recognition of different places with 96% precision, in spite of perceptual aliasing and image variability.