Augmenting Activity Recognition with Commonsense Knowledge and Satellite Imagery

Activity recognition gained relevance because of its applications in a variety of fields. Despite relevant improvements, classifiers are still inaccurate in several real-world circumstances or require excessively time-consuming training routines. In this paper we show how satellite imagery and common sense knowledge can be used for improving users' activity recognition performed on a mobile device. More specifically, we made use a personal device providing a list of candidate user activities instead of only the most probable one. Then, from the GPS location of the user, we (i) extract a list of neighboring commercial activities using a reverse geo-coding service and (ii) classify the satellite imagery of the area with state-of-the-art techniques. The proposed approach uses the ConceptNet network for ranking the list of candidate activities using both additional information. Results show an improvement in activity recognition accuracy