Semantic Place Labeling Using a Probabilistic Decision List of AdaBoost Classifiers

The success of mobile robots relies on the ability to extract from the environment additional information beyond simple spatial relations. In particular, mobile robots need to have semantic information about the entities in the environment such as the type or the name of places or objects. This work addresses the problem of classifying places (room, corridor or doorway) using mobile robots equipped with a laser range scanner. This paper compares the results of several AdaBoost algorithms (Viola-Jones AdaBoost, Gentle AdaBoost, Modest AdaBoost and Generalized AdaBoost for the place categorization) to train a set of classifiers and discusses these solutions. Since the problem is multi-class and these AdaBoosts provide only binary outputs, the AdaBoosts are arranged into Probabilistic Decision Lists (PDL), where each AdaBoost of the list gives a confidence value of each class. Then, Probabilistic Relaxation Labeling (PRL) is performed to smooth the classification results. Moreover, heuristics for removing incorrect regions are proposed to reduce the classification error. Experimental results suggest that PDL can be extended to several binary classifiers and show that PRL improves significantly the classification rates of the classifiers.