Real-Time Activity Classification by Matched Filtering Using Body-Worn Accelerometers

Having information of a user's activity can provide great use in modern-day devices such as ones that track and monitor the user's activity and fitness. In this paper, we demon-strate activity classification performance of the matched-filtering method with data obtained from a three-axis accelerometer worn by the user. We also show the real-time processing capability of our algorithms on the MSP432P401R low-powered micro-controller. Dimensionality reduction with principal component analysis (PCA) [1] is a data compression technique we use to improve our processing throughput which, inherently, has the added benefit of making our data invariant to sensor orientation. Data decimation in time is an additional throughput enhancement that we apply early to our data. We make use of an instance-based learning algorithm to train the device to learn the individual's motion patterns and store that information as activity templates for use in our matched-filter.