Device Transparent RSS-Based Indoor Localization

Signal strength variation across diverse devices is a major problem with RSS-based indoor localization. This thesis aims to solve this problem by considering two factors: the instability of collected RSS samples and linear shift of RSS patterns collected by different devices. We propose different techniques to handle the uncertainty of samples. Furthermore, we propose four different algorithms that make the localization system device-transparent. All solutions rely on the linear relationship across diverse devices. The first algorithm is the independent linear transformation algorithm. It starts by finding a set of candidate nearest neighbour fingerprints for an online point through correlation coefficient. Then, a series of linear transformations is applied to find nearest neighbours with high accuracy. The algorithm is justified through an analytical study of probability of error. The second algorithm is the cooperative linear transformation algorithm. This algorithm considers the case where multiple online points are available. The algorithm uses independent linear transformation algorithm to find the set of nearest neighbour fingerprints for each online point independently. Next, the online points and their detected nearest neighbours are used cooperatively to enhance the transformation model. The probability of error is also analyzed. The third algorithm utilizes label propagation algorithm for localization. Label propagation is a graph-based semi-supervised algorithm that requires the collection of some labeled points by the localization device. We propose to use label propagation in an un-supervised way. The algorithm transforms all fingerprints to the signal space of the localization device. All available online points are used cooperatively to find the transformation parameters. Next, it finds the weights of all edges connecting RSS points through Euclidean distances. Finally, label propagation is applied for localization. Those three algorithms utilize ordinary least squares as the linear regressor. The fourth algorithm is the robust linear transformation algorithm that utilizes robust linear regression for localization. Robust regression minimizes the effect of outliers by giving less weights to potential outliers. KNN is used as the localization engine. All proposed systems are automatic and do not require any training period. Experimental results indicate the proposed systems are reliable with very high positional accuracy.