Enhancing Object RecognitionBased on Contextual InformationUsing Markov Logic Networks

Abstract: Perception is a crucial part of an autonomous robotic system, as it processes sensory input and extracts useful information for action planning and execution such as recognizing objects in the environment where the robot is to act. Although object recognition has been studied extensively and a lot of progress has been made, current systems often face difficulties in dealing with ambiguities and uncertainties in the raw sensory data. It has been shown that using contextual information can reduce these ambiguities, however many different types of context measures exist and it is not always clear which type is the most effective for classification purposes.In this thesis we study how and to what extentMarkov Logic Networks (MLN) can be used to increase robustness in object classification by making use of context. MLNs consist of a combination of first-order logic and Markov Random fields, allowing for a solid framework for defining soft and hard rules that can be used efficiently in classification. Structure learning methods for MLNs allow for automatic improvement of the structure as well as flexibility when expanding the classification space. Therefore, it was of particular interest to study how the learning of the structure of an MLN performed against a manually constructed counterpart. We propose spatial relations, e.g.,’isRightTo’, ’isLeftTo’ and ’isAbove’, as a measure of context in order to reduce classification errors of items in various household scenes. Our experimental evaluations start with a comparison with a commonly used probabilistic classifier, the NaiveBayes classifier. Furthermore, we use a publicly available dataset to compare structure learning with a state-of-the-art system which uses MLNs with a manually designed structure. In addition, we test our approach with and without spatial relations on this dataset. Overall, the results show that MLNs outperform conventional classification algorithms and that spatial relations and structure learning increase the classification accuracy.