Active Vision and Visual Attention for Indoor Environment Classification

The purpose of this thesis is to give a contribution to facing the indoor classification problem using an active visual attention approach. In particular, our efforts are focused on the development of an active vision system able to extract useful information from a set of indoor ambient images and infer the environment classes. Our goal consists in building a symbolic house ambient map, i.e. one in which each ambient can be suitably labelled by hypotheses of the kind “bedroom, with probability p”.The basic ideas exploit a suitable combination of methodologies leading to a reasoning process that elaborates over the hypotheses emerging from the features analysis, in so establishing probabilistic causal relations between observations and current states. The whole classification system, is based on a combination of context free and context dependent analyses. The first one uses image features independent from the context, like color and intensity, to obtain a set of homogeneous regions, through a clustering procedure. The second one is a two phases context dependent classification process. In the first stage, the system gives a probabilistic classification of the textures extracted from the operating environment images. In the second stage, on the basis of a probabilistic region growing approach, a classification of textured area is provided. The output of the previous steps is used to find a probabilistic hypothesis concerning the indoor environments. The results of the analysis of both cases are combined using suitable weights depending on three factors: the entropy of the features, the correlation among the resulting areas obtained from the context free and context dependent image analysis and the results drawn by an off line learning procedure. The upshot of the weighted combination is a data structure which we define to be the system observation model. This model is the perceptual input to the reasoning activity of the system, structuring the agent current representation state. The system reasoning is based on a hidden Markov model: given a sequence of observation states it updates