Deep Neural Networks For Human Activity Recognition With Wearable Sensors
Abstract
Human Activity Recognition (HAR) has been attracting significant research attention because of a wide range of applications from healthcare to security. Recently, deep learning approaches have demonstrated great success in the HAR area. However, these models are often evaluated on the same subjects as those used to train the model; thus, the provided accuracy estimates do not pertain to new subjects. Consequently, this thesis examines the generalization capability of different machine learning architectures using Leave-One-Subject-Out Cross-Validation (LOSOCV) and then proposes a personalized model. The accuracy is improved by considering two feature selection directions, time- and frequency-domain, and by dynamically selecting the best model for that subject. In the time-domain, the best results are achieved by CNN with two convolutions and a one-dimensional filter. In the frequency domain, the results are improved through Empirical Mode Decomposition (EMD) with cubic spline and the personalization technique.