Algoritmos de aprendizaje estadístico aplicados a la radiolocalización en interiores
Fecha
2009-06
Autores
Título de la revista
ISSN de la revista
Título del volumen
Editor
Universidad Rey Juan Carlos
Resumen
En los últimos años, los avances en la tecnología de computación ubicua han
dado lugar a numerosas aplicaciones informáticas móviles, en las que la posición
del usuario es una información relevante que permite ofrecer servicios dependientes
del contexto. Cuando estos servicios se desarrollan en el interior de un edificio,
estimar dicha posición se torna un problema complejo. Para resolverlo, los sistemas
de radiolocalización en interiores (SRLI) utilizan información extraída del campo
electromagnético propagado entre el terminal del usuario y una infraestructura
de comunicaciones. Entre las distintas opciones, la utilización de la potencia de
señal recibida (RSS) por los dispositivos de las redes WiFi existentes en multitud
de edificios, proporciona una alta precisión con un coste bajo en el despliegue
del sistema. El objetivo principal de esta Tesis es estudiar de forma integrada
los elementos que conforman un SRLI basado en tecnología WiFi, proponiendo
soluciones novedosas en tres aspectos clave.
En primer lugar, la revisión de la literatura existente revela una carencia de
metodología para evaluar y comparar diferentes SRLI. Por ello, se desarrolla una
técnica de evaluación y comparación de sistemas basada en cuatro indicadores de
calidad: la media y la desviación típica del error, y la incertidumbre y sesgo asociados
a la distribución del error, propuestos estos últimos en esta Tesis. Además, se
desarrolla una metodología para el cálculo de estos indicadores, basada en técnicas
no paramétricas de estimación estadística de tipo bootstrap, y un test comparativo
que permite distinguir dos sistemas en términos de cualquier indicador. En segundo
lugar, se realiza un estudio del procedimiento de medida de la RSS. En concreto,
se estudian los parámetros de dicho procedimiento, como el número de muestras
por posición o la densidad de localizaciones que son necesarios para caracterizar
la RSS. Por otro lado, cuando en un mismo SRLI esta magnitud es registrada por
dispositivos heterogéneos, se observa que las prestaciones del sistema disminuyen
drásticamente. Por ello, se proponen algoritmos de calibración de dispositivos basados
en técnicas de aprendizaje estadístico para solventar el problema. En tercer
lugar, se aborda el estudio del problema de localización como un problema de interpolación no uniforme. Bajo este enfoque, se estudian las máquinas de vectores
soporte como herramienta eficaz para resolver el problema en cuestión, debido a su
capacidad de incorporar información a priori y a sus propiedades de generalización.
Específicamente, se proponen algoritmos que utilizan como núcleo la autocorrelación de la señal y además proporcionan una salida compleja para modelar las
dos dimensiones de la posición, obteniéndose unas altas prestaciones en términos
de todos los indicadores de calidad. El estudio de estos tres bloques proporciona
resultados relevantes para afectan al diseño de los SRLI.
During the last years, the advances in ubiquitous computing have favored the development of numerous mobile aplications, for which the knowledge of the user position is a relevant information in order to provide context aware services. When these services take place in an indoor environment, locating the user becomes a very difficult task. In this context, indoor radiolocation systems (IRLS) make use of the information extracted from the electromagnetic field, propagated between the user terminal and a wireless communications infrastructure, to estimate the user position. Among the different options, the use of the received signal strength (RSS) read by the WiFi devices already deployed in most buildings, provides a high location accuracy with a low deployment cost. Then, the main objective of this Thesis, is to carry out a comprehensive study of the elements which form a WiFi-based IRLS, proposing novel solutions in three main aspects of the problem. Firstly, a literature review shows that there is a lack of a structured evaluation methodology for comparing IRLS¿s. Hence, we develop a non parametric technique for IRLS¿s evaluation and comparison, based on four quality indicators: the commonly used mean and standard deviation; and the uncertainty and bias parameters, computed from the error distribution and originally proposed in this work. Moreover, a non parametric bootstrap method is developed to robustly estimate those parameters, and a comparative test is proposed to evaluate the difference between two systems in terms of any quality indicator. Secondly, we study the RSS measurement procedure, analyzing its main parameters, like the number of samples recorded in each position or the spatial density used for characterizing the RSS. Additionaly, when different devices are used in the same IRLS, the performance of the system severely degrades. In order to solve this problem, several statistical learning algorithms are proposed and compared for device calibration. Thirdly, the indoor location problem is studied like a non-uniform interpolation problem, and a support vector machine algorithm is proposed to solve it, due to its capabilities for incorporating previous knowledge of the problem and its generalization properties. Specifically, the proposed algorithm uses the autocorrelation of the signal for constructing the kernel, and has a complex output to model the two dimensions of the user position, providing a high performance in terms of all the quality indicators. All together, the results obtained for the three studied issues, represent a relevant and usefull contribution for the design of IRLS¿s.
During the last years, the advances in ubiquitous computing have favored the development of numerous mobile aplications, for which the knowledge of the user position is a relevant information in order to provide context aware services. When these services take place in an indoor environment, locating the user becomes a very difficult task. In this context, indoor radiolocation systems (IRLS) make use of the information extracted from the electromagnetic field, propagated between the user terminal and a wireless communications infrastructure, to estimate the user position. Among the different options, the use of the received signal strength (RSS) read by the WiFi devices already deployed in most buildings, provides a high location accuracy with a low deployment cost. Then, the main objective of this Thesis, is to carry out a comprehensive study of the elements which form a WiFi-based IRLS, proposing novel solutions in three main aspects of the problem. Firstly, a literature review shows that there is a lack of a structured evaluation methodology for comparing IRLS¿s. Hence, we develop a non parametric technique for IRLS¿s evaluation and comparison, based on four quality indicators: the commonly used mean and standard deviation; and the uncertainty and bias parameters, computed from the error distribution and originally proposed in this work. Moreover, a non parametric bootstrap method is developed to robustly estimate those parameters, and a comparative test is proposed to evaluate the difference between two systems in terms of any quality indicator. Secondly, we study the RSS measurement procedure, analyzing its main parameters, like the number of samples recorded in each position or the spatial density used for characterizing the RSS. Additionaly, when different devices are used in the same IRLS, the performance of the system severely degrades. In order to solve this problem, several statistical learning algorithms are proposed and compared for device calibration. Thirdly, the indoor location problem is studied like a non-uniform interpolation problem, and a support vector machine algorithm is proposed to solve it, due to its capabilities for incorporating previous knowledge of the problem and its generalization properties. Specifically, the proposed algorithm uses the autocorrelation of the signal for constructing the kernel, and has a complex output to model the two dimensions of the user position, providing a high performance in terms of all the quality indicators. All together, the results obtained for the three studied issues, represent a relevant and usefull contribution for the design of IRLS¿s.
Descripción
Tesis Doctoral leída en la Universidad Rey Juan Carlos en junio de 2009. Directores de la Tesis: José Luis Rojo Álvarez y Javier Ramos López