Deep learning and model personalization in sensor-based human activity recognition

Abstract

AbstractHuman activity recognition (HAR) is a line of research whose goal is to design and develop automatic techniques for recognizing activities of daily living (ADLs) using signals from sensors. HAR is an active research filed in response to the ever-increasing need to collect information remotely related to ADLs for diagnostic and therapeutic purposes. Traditionally, HAR used environmental or wearable sensors to acquire signals and relied on traditional machine-learning techniques to classify ADLs. In recent years, HAR is moving towards the use of both wearable devices (such as smartphones or fitness trackers, since they are daily used by people and they include reliable inertial sensors), and deep learning techniques (given the encouraging results obtained in the area of computer vision). One of the major challenges related to HAR is population diversity, which makes difficult traditional machine-learning algorithms to generalize. Recently, researchers successfully attempted to address the problem by proposing techniques based on personalization combined with traditional machine learning. To date, no effort has been directed at investigating the benefits that personalization can bring in deep learning techniques in the HAR domain. The goal of our research is to verify if personalization applied to both traditional and deep learning techniques can lead to better performance than classical approaches (i.e., without personalization). The experiments were conducted on three datasets that are extensively used in the literature and that contain metadata related to the subjects. AdaBoost is the technique chosen for traditional machine learning, while convolutional neural network is the one chosen for deep learning. These techniques have shown to offer good performance. Personalization considers both the physical characteristics of the subjects and the inertial signals generated by the subjects. Results suggest that personalization is most effective when applied to traditional machine-learning techniques rather than to deep learning ones. Moreover, results show that deep learning without personalization performs better than any other methods experimented in the paper in those cases where the number of training samples is high and samples are heterogeneous (i.e., they represent a wider spectrum of the population). This suggests that traditional deep learning can be more effective, provided you have a large and heterogeneous dataset, intrinsically modeling the population diversity in the training process.