| This dissertation addresses the problem of visual place categorization, which aims at augmenting different locations in the environment visited by an autonomous robot with information that relates them to human-understandable concepts. In visual place categorization, there are two constraints that should be taken into consideration: (i) the categorical place label assigned to each location should be consistent with the observation (i.e., image) gathered at that location; and (ii) the overall labelling should be temporally coherent (i.e., the place labels assigned to consecutively visited locations should be consistent). To address both constraints together, this dissertation formulates the problem of visual place categorization in terms of energy minimization. A method based on graph cuts is used to minimize energy for a function of a data term and a temporal term. While the data term aims at assigning visual observations to a set of pre-specified place categories, the temporal term incorporates contextual evidence from neighbours to ensure that the labels vary smoothly almost everywhere while preserving discontinuities at the borders between adjacent places in the environment. For the data term, a novel context-based scene categorization method is presented that is invariant to common changes in dynamic environments (e.g., lighting condition, partial occlusion, etc.) and robust against intra-class variations. For the temporal term, a general solution is presented that incorporates statistic cues, without being restricted by constant and small neighbourhood radii, or being dependent on the actual path followed by the robot. An extensive set of experiments on several publicly available databases validates the robustness of the proposed approach in reliably labelling visual observations with place categories and efficiently incorporating contextual cues. |
