الفهرس 
Table of ContentsOnly 14 pages are availabe for public view
 |  | from | 105 |  |  |
| | | from | 105 | | |
Abstract
Localization of moving targets in indoor environment is increasingly becoming an integral part of many emerging business applications. Localization techniques based on Received Signal Strength (RSS) are known to be one of the most common techniques for indoor localization due to their relative low cost as no expensive devices (e.g., GPS) are needed. However, estimation accuracy, using RSS suffers from two shortcomings. First, the intensive human effort is needed in order to prepare the indoor environment for localization with high accuracy. Second, RSS estimation accuracy is notable impacted by the obstacles that natural exist in the real-life indoor environment.
Accordingly, in this dissertation, we attempt to address the aforementioned challenges in RSS-based estimation for indoor environments. To this end, the dissertation proposes and evaluates three techniques to estimate the observation model parameters for simulated and real time collected RSS data sets. The fundamental concept of the three techniques is to model the motion dynamics of a mobile target as a hidden Markov model in which each state corresponds to geographic location, then estimate the mobile target location with a position estimator using Bayesian minimum mean square error (BMMSE) within the fingerprint locations.
The first proposed technique, namely posterior state distribution of HMM with Gaussian observation model per state (GOMS), uses the Gaussian observation model in a uniform RSS fingerprints environment. The second technique, namely posterior state distribution of HMM with Gaussian observation model for sparse fingerprint data set (GOMSF), uses the Gaussian observation modelin a non-uniform sparse RSS fingerprints environment to achieve a reasonable localization accuracy while requiring a less human effort using linear interpolation for offline network setting. Finally, the third technique, namely posterior state distribution of HMM with Gaussian mixture observation model (GMOM) and posterior state distribution of HMM with Gaussian mixture component
16
optimization observation model (GMOOM), use Gaussian mixture model for multiple uniform RSS fingerprints to achieve a lower localization error in a complex indoor environment by predicting the RSS characteristics among the anchor nodes with a much easier way to construct the offline network setting.
The three proposed techniques are implemented and evaluated under three different RSS data set environment models: (1) free space path loss propagation model, (2) a real time RSS collected data set from the fifth floor of a five story building of the faculty of computer science and information technology, riga technical university, and (3) a ray tracing radio wave propagation simulator tool to construct a harsh and critical indoor environment. Evaluation results show that GOMS achieves 35% localization accuracy improvement compared to the Bayesian filtering used in [1]. Whereas GOMSF achieves 19% localization accuracy compared to weighted nearest neighborhood (WKNN). Moreover, in critical environment with obstacles, GOMSF shows 44% localization accuracy improvement compared to the Bayesian filtering with only 30% of the fingerprint data set which means, 70% human calibration effort removed with accuracy improvement. GMOOM shows higher performance with 60% localization accuracy improvement.