Improving Situation Recognition via Commonsense Sensor Fusion

Pervasive services often rely on multi-modal classification to implement situation-recognition capabilities. However, current classifiers are still inaccurate and unreliable. In this paper we present preliminary results obtained with a novel approach that combines well established classifiers using a commonsense knowledge base. The approach maps classification labels produced by independent classifiers to concepts organized within the Concept Net network. Then it verifies their semantic proximity by implementing a greedy approximate sub-graph search algorithm. Specifically, different classifiers are fused together on a commonsense basis for both: (i) improve classification accuracy and (ii) deal with missing labels. Experimental results are discussed through a real-world case study in which two classifiers are fused to recognize both user's activities and visited locations.