Particle filters for tracking in wireless sensor networks

• Autores: Katrin Achutegui Roncal
• Directores de la Tesis: Joaquín Miguez Arenas (dir. tes.)
• Lectura: En la Universidad Carlos III de Madrid ( España ) en 2013
• Idioma: inglés
• Tribunal Calificador de la Tesis: Robert Adrien Piche (presid.), Fernando Pérez Cruz (secret.), Carlos Fernández Prades (voc.)
• Materias:
  • Matemáticas
    • Probabilidad
      • Aplicación de la probabilidad
• Enlaces
  • Tesis en acceso abierto en: e-Archivo
• Resumen

  • The goal of this thesis is the development, implementation and assessment of efficient particle filters (PFs) for various target tracking applications on wireless sensor networks (WSNs). We first focus on developing efficient models and particle filters for indoor tracking using received signal strength (RSS) in WSNs. RSS is a very appealing type of measurement for indoor tracking because of its availability on many existing communication networks. In particular, most current wireless communication networks (WiFi, ZigBee or even cellular networks) provide radio signal strength (RSS) measurements for each radio transmission. Unfortunately, RSS in indoor scenarios is highly influenced by multipath propagation and, thus, it turns out very hard to adequately model the correspondence between the received power and the transmitterto- receiver distance. Further, the trajectories that the targets perform in indoor scenarios usually have abrupt changes that result from avoiding walls and furniture and consequently the target dynamics is also difficult to model. In Chapter 3 we propose a flexible probabilistic scheme that allows the description of different classes of target dynamics and propagation environments through the use of multiple switching models. The resulting state-space structure is termed a generalized switching multiple model (GSMM) system. The drawback of the GSMM system is the increase in the dimension of the system state and, hence, the number of variables that the tracking algorithm has to estimate. In order to handle the added difficulty, we propose two Rao-Blackwellized particle filtering (RBPF) algorithms in which a subset of the state variables is integrated out to improve the tracking accuracy. As the main drawback of the particle filters is their computational complexity we then move on to investigate how to reduce it via de distribution of the processing. Distributed applications of tracking are particularly interesting in situations where high-power centralized hardware cannot be used. For example, in deployments where computational infrastructure and power are not available or where there is no time or trivial way of connecting to it. The large majority of existing contributions related to particle filtering, however, only offer a theoretical perspective or computer simulation studies, owing in part to the complications of real-world deployment and testing on low-power hardware. In Chapter 4 we investigate the use of the distributed resampling with non-proportional allocation (DRNA) algorithm in order to obtain a distributed particle filtering (DPF) algorithm. The DRNA algorithm was devised to speed up the computations in particle filtering via the parallelization of the resampling step. The basic assumption is the availability of a set of processors interconnected by a high-speed network, in the manner of state-of-the-art graphical processing unit (GPU) based systems. In a typical WSN, the communications among nodes are subject to various constraints (i.e., transmission capacity, power consumption or error rates), hence the hardware setup is fundamentally different. We first revisit the standard PF and its combination with the DRNA algorithm, providing a formal description of the methodology. This includes a simple analysis showing that (a) the importance weights are proper and (b) the resampling scheme is unbiased. Then we address the practical implementation of a distributed PF for target tracking, based on the DRNA scheme, that runs in real time over a WSN. For the practical implementation of the methodology on a real-time WSN, we have developed a software and hardware testbed with the required algorithmic and communication modules, working on a network of wireless light-intensity sensors. The DPF scheme based on the DRNA algorithm guarantees the computation of proper weights and consistent estimators provided that the whole set of observations is available at every time instant at every node. Unfortunately, due to practical communication constraints, the technique described in Chapter 4 may turn out unrealistic for many WSNs of larger size. We thus investigate in Chapter 5 how to relax the communication requirements of the DPF algorithm using (a) a random model for the spread of data over the WSN and (b) methods that enable the out-of-sequence processing of sensor observations. The presented observation spread scheme is flexible and allows tuning of the observation spread over the network via the selection of a parameter. As the observation spread has a direct connection with the precision on the estimation, we have also introduced a methodology that allows the selection of the parameter a priori without the need of performing any kind of experiment. The performance of the proposed scheme is assessed by way of an extensive simulation study.