AbstractsComputer Science

Particle filters (PFs) are efficient tools for nonlinear state estimation especially for a wireless indoor target tracking system, who estimates the target's position using wireless facilities in the buildings or houses. Particle filters are recursive Monte-Carlo methods based on Bayes' theorem, which fuse the previous state information and the current measurement data to obtain the target position. The advantage of using PFs is that the heterogeneous information can be combined effectively within the PFs to estimate the unknown state. The purpose is to design PF algorithms with high estimation accuracy, where average absolute error approaches 0, and make PFs robust to hybrid line-of-sight (LOS) and non-line-of-sight (NLOS) errors. Particle filters should also have a low computational complexity. The major contributions are five folds: 1. The impact of instantaneous measurement error is firstly found and analyzed in this thesis. It is the major source of the estimation error of PFs. According to the analysis, a likelihood adaptation method is proposed to reduce the instantaneous measurement error. Then, adaptive PFs integrated with the likelihood adaptation method are developed. 2. Due to the NLOS and multi-path effect, the ranging error is difficult to model in indoor environment. Therefore, a dynamic Gaussian model (DGM) is proposed to describe the distribution of hybrid LOS/NLOS ranging errors. Then, adaptive PFs using DGM are extended in the high dynamic indoor environment still with the accurate estimation. 3. Design the architecture of a real world tracking system for adaptive PFs. In addition, adaptive PF fusing building layout information to improve the estimation accuracy is proposed. 4. Propose distributed PFs protocols to collaboratively estimate the target using local anchors. Selective gossip algorithms are applied in the distributed PF design. 5. Theoretical analysis of the adaptive PFs for the estimation performance based on the Cramer-Rao lower bound. It is proved that with reliable priori information, the adaptive PFs outperform conventional PFs. The performance of the adaptive PFs are evaluated in various simulations and real world experiment. The estimation results are compared with the conventional PFs and other localization algorithms. For the conventional PFs, the average absolute errors have biases, which are about 0.5m to 1m, and the root mean square error (RMSE) is more than 2m, but the average absolute errors of proposed adaptive PFs can approach to 0 and RMSEs are only about 1.5m error, which meets the requirements of the indoor location based services. In addition, only 30 particles are required in the adaptive PFs, which significantly reduce the computational complexity. In real world experiments, adaptive PFs also outperform other localization algorithms, e.g. linear least square method, nonlinear least square method, min-max algorithm, extended Kalman filter, etc. The estimation accuracy of distributed PFs using selective gossip algorithms are guaranteed. The…