A Deep Learning Approach for Recognizing Daily Movement Patterns through Accelerometer Data

Abstract

Physical activity is a key factor in the treatment of chronic diseases such asdiabetes, cardiovascular disease, and depression. Doctors and personal trainershave limited methods to accurately monitor and classify a patients actual activi-ties based on training diaries and logs that are commonly used today. In this thesis,we apply a tri-axial accelerometer carried by a patient to collect data associated todifferent activities of daily life (ADL) and utilize deep learning (DL) algorithmsfor classifying distinct activities based on the data obtained from the accelerome-ter. Among various DL methods and algorithms, we adopt specifically deep neuralnetworks (DNN) and recurrent neural networks (RNN) to classify movement pat-terns. In addition, we compare our proposed structures with the state-of-the-artmethods via extensive experiments. Numerical results show that our proposedDNN model slightly exceeds, and our RNN model vastly outperforms the state-of-the-art methods in classification of basic movement patterns. The overall solu-tion for data collection and movement classification provides medical doctors andtrainers a promising way to precisely track and understand the physical activitiesof a patient for a better treatment.