Robust Dynamic Human Activity Recognition Based on Relative Energy Allocation

This paper develops an algorithm for robust human activity recognition in the face of imprecise sensor placement. It is motivated by the emerging body sensor networksthat monitor human activities (as opposed to environmental phenomena) for medical, entertainment, health-and-wellness, training, assisted-living, or entertainment reasons. Activities such as sitting, writing, and walking have been successfully inferred from data provided by body-worn accelerometers. A common concern with previous approaches is their sensitivity with respect to sensor placement. This paper makes two contributions. First, we explicitly address robustness of human activity recognition with respect to changes in accelerometer orientation. We develop a novel set of features based on relative activity-specific body-energy allocation and successfully apply them to recognize human activities in the presence of imprecise sensor placement. Second, we evaluate the accuracy of the approach using empirical data from body-worn sensors.