| Abstract | In this paper, we propose a holistic classification scheme for different room types, like office or meeting room, based on 3D features. Such a categorization of scenes provides a rich source of information
about potential objects, object locations, and activities typically found in them. Scene categorization is a challenging task. While outdoor scenes can be sufficiently characterized by color and texture features, indoor scenes consist of human-made structures that vary in terms of color and texture across different individual rooms of the same category. Nevertheless, humans tend to have an immediate impression in which room type they are. We suggest that such a decision could be based on the coarse spatial layout of a scene. Therefore, we present a system that categorizes different room types based on 3D sensor data extracted by a Time-of-Flight (ToF) camera. We extract planar structures combining region growing and RANSAC approaches. Then, feature vectors are defined on statistics over the relative sizes of the planar patches, the angles between pairs of (close) patches, and the ratios between sizes of pairs of patches to train classifiers. Experiments in a mobile robot scenario study the performance in classifying a room based on a single percept. |
