Although research community have put lots of effort in improving localization in indoor environment through various of sensing techniques, outdoor localization still heavily relies on GPS solely. Despite GPS performs successfully in perfect outdoor scenario (e.g., large open air zone), many other outdoor scenarios in metropolitan area make GPS fail to achieve a satisfactory performance due to signal blocking or multi-path effect. Therefore, to improve localization performance in metropolitan area, an integrated outdoor localization solution is desired, which takes GPS as one part and other sources as supplementary. For designing such an integrated solution, we need to consider two questions: (1) what are the proper supplementary sources? (2) when are the proper moments to trigger a switch between GPS and other sources? To find a proper complementary source for outdoor localization in urban area, we have to consider the potential large-scale infrastructure and human effort in collecting and labelling data. The ever-expanding scale of WiFi deployments in urban area has made accurate GPS-free outdoor localization become possible. Additionally, there are many existing crowdsensing database available that collect WiFi signal of users' surrounding hotspots and users' locations. Leveraging the existing infrastructure and data source, we propose WOLoc (WiFi-only Outdoor Localization) as an outdoor localization solution using only WiFi hotspots labelled by crowdsensing. On one hand, we do not take these labels as fingerprints as it is almost impossible to extend indoor localization mechanisms by fingerprinting urban areas. On the other hand, we avoid the over-simplified local synthesis methods (e.g., centroid) that significantly lose the information contained in the labels. Instead, WOLoc adopts a semi-supervised manifold learning approach that accommodates all the labeled and unlabeled data for a given area, and the output concerning the unlabeled part will become the estimated locations for both unknown users and unknown WiFi hotspots. Empirical results show that WOLoc can achieve a meter-level localization accuracy in many urban areas. Despite the success of WOLoc in many cases, due to the lack of labelled data for hotspot location, WOLoc may suffer from an imbalanced alignment between the user manifold and hotspot manifold. Such misalignment may lead to extreme cases which are meaningless in some scenarios. To reduce such failures, we apply text mining techniques to analyze the SSIDs of hotspots, and develop a pipeline to incorporate SSID analysis in both pre-processing and hotspot labelling process of WOLoc. We make use of existing crowdsourced location services databases (e.g. Foursquare, Google Places) to give more accurate hotspot location labels. Empirical results show that the hotspot SSID analysis pipeline has successfully reduced extreme failure cases for WOLoc and lead it to a more consistent performance To find a proper trigger between GPS and WiFi, we need to understand how GPS performs in different urban outdoor scenarios. The latest developments in Android have granted us access to raw GNSS measurements, which contain far more information than commonly derived GPS location indicators. We explore these newly available measurements in order to better characterize the diversified urban scenarios and propose a deep learning model to identify representations for respective location contexts. With the deep profiling, we offer a more fine-grained semantic classification than binary indoor-outdoor detection and derive a GPS error indicator more meaningful than that provided by Google Maps. These results are all corroborated by our extensive data collection and trace-driven evaluations. To sum up, in this dissertation, we propose a series of data-driven approaches that aim for an integrated outdoor localization by exploiting WiFi RSSI data, WiFi SSID labels and GNSS raw measurements. We choose WiFi as supplementary sources and design a full version of system based on manifold learning technique to achieve WiFi-only outdoor localization in urban areas. We further develop the SSID-based text-mining techniques to improve both accuracy and robustness. We finally propose a GNSS profiling technique to evaluate GPS performance and differentiate different location context, which can be used as a trigger between GPS and WiFi in an integrated solution.
99
