Smartphone-based human activity recognition
- Autores: Jorge Luis Reyes Ortiz
- Directores de la Tesis: Davide Anguita (dir. tes.), Francesc Xavier Parra Llanas (dir. tes.)
- Lectura: En la Universitat Politècnica de Catalunya (UPC) ( España ) en 2014
- Idioma: inglés
- Tribunal Calificador de la Tesis: Francisco Javier Ruiz Vegas (presid.), José Vehí Casellas (secret.), Ines Sousa (voc.)
- Materias:
- Enlaces
- Tesis en acceso abierto en: TDX
- Resumen
Human Activity Recognition (HAR) is a multidisciplinary research field that aims to gather data regarding people's behavior and their interaction with the environment in order to deliver valuable context-aware information. It has nowadays contributed to develop human-centered areas of study such as Ambient Intelligence and Ambient Assisted Living, which concentrate on the improvement of people's Quality of Life. The first stage to accomplish HAR requires to make observations from ambient or wearable sensor technologies. However, in the second case, the search for pervasive, unobtrusive, low-powered, and low-cost devices for achieving this challenging task still has not been fully addressed. In this thesis, we explore the use of smartphones as an alternative approach for performing the identification of physical activities. These self-contained devices, which are widely available in the market, are provided with embedded sensors, powerful computing capabilities and wireless communication technologies that make them highly suitable for this application. This work presents a series of contributions regarding the development of HAR systems with smartphones. In the first place we propose a fully operational system that recognizes in real-time six physical activities while also takes into account the effects of postural transitions that may occur between them. For achieving this, we cover some research topics from signal processing and feature selection of inertial data, to Machine Learning approaches for classification. We employ two sensors (the accelerometer and the gyroscope) for collecting inertial data. Their raw signals are the input of the system and are conditioned through filtering in order to reduce noise and allow the extraction of informative activity features. We also emphasize on the study of Support Vector Machines (SVMs), which are one of the state-of-the-art Machine Learning techniques for classification, and reformulate various of the standard multiclass linear and non-linear methods to find the best trade off between recognition performance, computational costs and energy requirements, which are essential aspects in battery-operated devices such as smartphones. In particular, we propose two multiclass SVMs for activity classification:one linear algorithm which allows to control over dimensionality reduction and system accuracy; and also a non-linear hardware-friendly algorithm that only uses fixed-point arithmetic in the prediction phase and enables a model complexity reduction while maintaining the system performance. The efficiency of the proposed system is verified through extensive experimentation over a HAR dataset which we have generated and made publicly available. It is composed of inertial data collected from a group of 30 participants which performed a set of common daily activities while carrying a smartphone as a wearable device. The results achieved in this research show that it is possible to perform HAR in real-time with a precision near 97\% with smartphones. In this way, we can employ the proposed methodology in several higher-level applications that require HAR such as ambulatory monitoring of the disabled and the elderly during periods above five days without the need of a battery recharge. Moreover, the proposed algorithms can be adapted to other commercial wearable devices recently introduced in the market (e.g. smartwatches, phablets, and glasses). This will open up new opportunities for developing practical and innovative HAR applications.
