Indoor place classification for intelligent mobile systems

Publication Type:
Issue Date:
Full metadata record
Place classification is an emerging theme in the study of human-robot interaction which requires common understanding of human-defined concepts between the humans and machines. The requirement posts a significant challenge to the current intelligent mobile systems which are more likely to be operating in absolute coordinate systems, and hence unaware of the semantic labels. Aimed at filling this gap, the objective of the research is to develop an approach for intelligent mobile systems to understand and label the indoor environments in a holistic way based on the sensory observations. Focusing on commonly available sensors and machine learning based solutions which play a significant role in the research of place classification, solutions to train a machine to assign unknown instances with concepts understandable to human beings, like room, office and corridor, in both independent and structured prediction ways, have been proposed in this research. The solution modelling dependencies between random variables, which takes the spatial relationship between observations into consideration, is further extended by integrating the logical coexistence of the objects and the places to provide the machine with the additional object detection ability. The main techniques involve logistic regression, support vector machine, and conditional random field, in both supervised and semi-supervised learning frameworks. Experiments in a variety of environments show convincing place classification results through machine learning based approaches on data collected with either single or multiple sensory modalities; modelling spatial dependencies and introducing semi-supervised learning paradigm further improve the accuracy of the prediction and the generalisation ability of the system; and vision-based object detection can be seamlessly integrated into the learning framework to enhance the discrimination ability and the flexibility of the system. The contributions of this research lie in the in-depth studies on the place classification solutions with independent predictions, the improvements on the generalisation ability of the system through semi-supervised learning paradigm, the formulation of training a conditional random field with partially labelled data, and the integration of multiple cues in two sensory modalities to improve the system's functionality. It is anticipated that the findings of this research will significantly enhance the current capabilities of the human robot interaction and robot-environment interaction.
Please use this identifier to cite or link to this item: