Human Activity and Correlated Posture Monitoring Using Earlobe-Worn Wearable Sensor System and Deep Learning Algorithm

An approach for monitoring human activities and correlated postures using an earlobe-worn wearable sensor and a deep learning algorithm is proposed. The herein-used miniaturized wearable is called TRACE and is to be mounted on an earlobe, for which smaller movements are expected compared with common locations for wearable devices. This work adopts both biological [heart rate (HR)] and physical [acceleration (ACC)] signals that are simultaneously recorded in real time. The recorded raw data are sensor-fused, adequately labeled, processed, and trained by a deep learning algorithm. Using the data, several deep learning models based on different deep learning algorithms are developed and compared. Among our experimental conditions, the best classification accuracy is achieved when the biological and physical sensor signals are fused and a long-short term memory (LSTM) algorithm is used. The proposed approach successfully classified human activities and postures with high accuracy reaching
$>$
95% for two patterns of daily activities. More specifically, the proposed method achieved 97.83%, 97.31%, and 98.19% accuracy for the deep neural network (DNN), convolutional neural network (CNN), and LSTM models, respectively.