Publicación: Detectores de caídas para teléfonos inteligentes basados en algoritmos de detección de novedad
Fecha
2015-02-02
Autores
Editor/a
Tutor/a
Coordinador/a
Prologuista
Revisor/a
Ilustrador/a
Derechos de acceso
Atribución-NoComercial-SinDerivadas 4.0 Internacional
info:eu-repo/semantics/openAccess
info:eu-repo/semantics/openAccess
Licencia Creative Commons
http://creativecommons.org/licenses/by-nc-nd/4.0
Título de la revista
ISSN de la revista
Título del volumen
Editor
Universidad Nacional de Educación a Distancia (España). Escuela Técnica Superior de Ingenieros Industriales. Departamento de Ingeniería Eléctrica, Electrónica y Control
Resumen
Falls are an important public health problem, leading to severe physical and psychological consequences among the elderly and economic consequences for health systems. A prompt detection of falls could alleviate these problems. Despite the large number of scientific studies, this kind of detectors has not become popular and a reliable and robust detector, usable by the elderly and at an affordable price, would be welcome. Computer Vision-based systems are restricted to a given environment and the variety of published algorithms indicates that the ultimate solution has not been found yet. Wearable sensors can be used at any place and time, but they are not comfortable and the user can forget to wear them. However, if the sensors were integrated in a smartphone, these drawbacks will disappear to a great extent. In this thesis, we propose to use novelty detection algorithms in mid-range smartphones. Smartphones are becoming very popular. Even though they are not adapted for older people, a suitable design could help to overcome this barrier. Besides, it is clear that the elderly in the future will get used to utilising them. Mid-range devices include accelerometers and communication functions at a reasonable price. On the other hand, several algorithms has been tested to detect falls from accelerometer data, either simple thresholds or more complex Machine Learning techniques. Novelty detection techniques model the normal behaviour (movement), so that a fall could be detected as an anomaly. This is interesting for several reasons. While real fall data are scarce, it is easy to record true data of activities of daily living (ADL), as much as needed. In addition, whenever a new user carries the phone, it can record new data and re-train the system. In this way the detector can adapt to conditions different from those of the first training phase, like the kind of movements or the place where the phone is worn. To carry on our study, we have registered a data set with ten volunteers, who simulated falls and carried the phone in their daily life for several days. The data set is publicly available, being one of the few that can be found for fall detection, and improving the others in terms of number of records. An off-line analysis has been made with our data set. We have compared some of the state-of-the-art novelty detection algorithms. We have selected the nearest neighbour (NN) as the most suitable. Then, we have compared it with a traditional classifier, a Support Vector Machine (SVM). SVM outperforms NN in a standar cross-validation, in a cross-validation by fall type or if the system operates at a different sampling frequency. Nevertheless, if the phone is worn in a different place (pocket - hand bag) or if it is personalized, NN reaches or even exceeds SVM in performance. Additional conditions regarding the inactivity or the orientation change after the fall can improve the results. We have estimated the number of false positives per week in a set of ADL of older people. We have found a reasonable number for some of them, but the system has still to ameliorate for others. A mobile application has been developed as a proof-ofconcept, checking its correct operation. Our future lines of work will be splitted in two sides. Firstly, from a more technical point of view, we would like to improve NN algorithms by smoothing the decision boundary. Secondly, we think that it would be highly desirable to record data from older people all over the time line for several days, without gaps, unlike our first data set, in order to take into account all the phases of a potential fall.
Descripción
Categorías UNESCO
Palabras clave
Citación
Centro
Facultades y escuelas::Escuela Internacional de Doctorado
Departamento
No procede