Efficient codebook generation for appearance-based localization

Mobile robot localization relies on efficient environment modeling, which in turns relies on robust feature set representation. Local point of interest descriptors provide distinguishable visual features, however suffer from their huge size. In this paper, the problem of reducing the size of such descriptors for the sake of robot localization is addressed. A two-phase solution is proposed. The first uses an entropy measure for features evaluation and selection. The second generates a codebook from the entropy-based features, which manages to reduce the size of the features significantly, while still preserving similar performance like that of the non reduced descriptor. The localization precision within an indoor environment is 96%, with 90% reduction in the number of features.