ROBUST PLACE RECOGNITION WITHIN MULTI-SENSOR VIEW SEQUENCES USING BERNOULLI MIXTURE MODELS

Abstract

This article reports on the use of Hidden Markov Models to improve the results of Localization within a sequence of Sensor Views. Local image features (SIFT) and multiple types of features from a 2D laser range scan are all converted into binary form and integrated into a single, binary, Feature Incidence Matrix (FIM). To reduce the large dimensionality of the binary data, it is modeled in terms of a Bernoulli Mixture providing good results that were reported in an earlier presentation. We have improved the good performance of the approach by incorporating the Bernoulli mixture model inside a Bayesian Network Model, an HMM, that accumulates evidence as the robot travels along the environment.