A system identification approach for recognition of personalized user motion patterns from mobile sensing data

Human activity can serve as an identifier of subject health, behavioral patterns, and personal preferences. With the sudden splurge in mobile and wearable devices, activity data has become more readily available to design useful applications that enhance the users' everyday lives without any obtrusive intervention. This paper focuses on the use of a system identification approach to characterize human movement dynamics from accelerometer sensors during physical activity and subsequently to construct user-specific models that can potentially be incorporated in personalized healthcare and safety. The study investigates the human wrist-to-ankle dynamic relationship using various linear prediction model structures. It was found that the ARMAX model structure is the most widely applicable across stereotypical activity patterns (e.g. walking, running). Importantly, and after conducting a series of order selection and validation tests, it was noted that specific activities across multiple individuals can be fit within a common model where the orders are fixed and only the parameters of that model are tuned to individual users. A potential application of these common models to user identification, as reflected through the models' frequency responses, is discussed.