Deep-Learning-Based Human Activity Recognition Using Wearable Sensors

Accurate activity recognition has multiple useful health applications, including home-based monitoring in various chronic disease applications. This paper utilizes a deep-learning-based algorithm for recognition of various daily living activities. A nonlinear observer that estimates body segment tilt angles and sensor bias parameters, using inertial sensors provides the inputs to the deep-learning-based algorithm. The observer is designed using Lyapunov analysis and an LMI to obtain the observer gain. Switched gains are utilized for each monotonic region of the nonlinear dynamics to ensure global stability. A convolutional neural network long short-term memory network (CNN-LSTM) is trained as the state-of-the-art deep learning-based activity recognition algorithm. Experimental results are presented on the performance of the CNN-LSTM network with the raw inertial sensor data and compared with results when inputs are the angles estimated by the nonlinear observer instead of the raw data. The activity recognition algorithm that utilizes 18 raw inertial signal data shows an average accuracy of 96.47% while the algorithm that utilizes only the 3 tilt angles estimated by the nonlinear observer provides an average accuracy of 99.81%.