Improving Activity Recognition via Satellite Imagery and Commonsense Knowledge

Activity recognition gained relevance in recent years because of its numerous applications. Despite relevant improvements, current classifiers are still inaccurate in several usage conditions or require time-consuming training. In this paper we show how localisation data and common sense knowledge could be used to improve activity recognition. More specifically, given the GPS position of the user, we both gather (i) a list of neighbouring commercial activities using a reverse geo-coding service and (ii) classify the satellite image of the area with state-of-the-art techniques. The approach maps classification labels produced by the three classifiers (i.e., activity, reverse geocoding localisation, satellite imagery localisation) to concepts within the ConceptNet network for the sake of improving activity recognition accuracy.