A context-aware hierarchical approach for activity recognition based on mobile devices

Various sensors integrated in the wearable device provide massive data for activity recognition. In this paper, a context-aware hierarchical approach is proposed for the recognition of activities using accelerometers on smartphones and smartwatches. We adopt a simple variance threshold based method and separate the activities into two major categories named body-fixed set and body-unfixed set according to the inherent characteristics of these activities in the first layer. Next the Support Vector Machine approach is used respectively for the two sets in the second layer. A probability distribution over activity labels instead of a single activity result is generated in this layer. In the third layer, the contextual information is introduced to improve the classification result. Our comparative study with ordinary Support Vector Machines and other alternative methods has shown that our method is more robust and accurate.