University of Central Florida, United States
Macro-class Selection, Hierarchical Classification, Human Activity Recognition.
Embedded Communications Systems
Mobile and Pervasive Computing
Pervasive Embedded Devices
Ubiquitous Computing Systems and Services
Quality classifiers can be difficult to implement on the limited resources of an embedded system, especially if the data contains many confusing classes. This can be overcome by using a hierarchical set of classifiers in which specialized feature sets are used at each node to distinguish within the macro-classes defined by the hierarchy. This method exploits the fact that similar classes according to one feature set may be dissimilar according to another, allowing normally confused classes to be grouped and handled separately. However, determining these macro-classes of similarity is not straightforward when the selected feature set has yet to be determined. In this paper, we present a new greedy forward selection algorithm to simultaneously determine good macro-classes and the features that best distinguish them. The algorithm is tested on two human activity recognition datasets: CMU-MMAC (29 classes), and a custom dataset collected from a commodity smartphone for this paper (9 class
es). In both datasets, we employ statistical features obtained from on-body IMU sensors. Classification accuracy using the selected macro-classes was increased 69% and 12% respectively over our non-hierarchical baselines.