Date of Award
Campus Access Dissertation
Doctor of Philosophy (PhD)
Duc A. Tran
GPS using satellite signals is the most obvious way to get location information. It does not work indoors. Even for outdoor environments, it is not energy-efficient to turn it on continuously all the time. Consequently, a GPS-free localization solution is desirable. A common approach in this direction is location fingerprinting. The idea is to precompute a "fingerprint", which is a set of local measurements, at each of a set of sample locations and use this information as training data to predict the location of a mobile device given its real-time fingerprint. For good accuracy, the training set should be large enough to appropriately cover the area. A quality training set, however, is not easy to obtain in practice. How can one maximize the localization accuracy with limited fingerprint information? The dissertation addresses this question in two settings. In the first, a fingerprint is assumed to be a vector of device-to-reference measurements with respect to a universal set of references. In the second, a fingerprint is a list of device-to-device measurements corresponding to a size-varying set of devices nearby. The following contributions are made. For the first setting, which is the most common in the literature of fingerprint localization, the contribution is a framework using mathematical regularization to enrich the training dataset with "unlabeled" fingerprints, i.e., those available without location information. For the second setting, which is quite new to the literature, the contribution is a framework based on geometric embedding to transform the problem to the first setting and demonstrates its effectiveness with several k-nearest-neighbor localization solutions. The theoretical findings are substantiated by evaluation studies using synthetic and real-world data.
Gong, Siyuan, "Fingerprint-Based Localization with Limited Training Data" (2019). Graduate Doctoral Dissertations. 521.