Crowdsourcing Mobile Data for Indoor Positioning

Indoor positioning is a key enabler for context-aware computing. Applications such as navigation, tracking, monitoring, spatial programming and location-based information sharing require the accurate and timely inputs of users' indoor positions. Signal fingerprinting, as a popular infrastructure-free solution for indoor positioning, conventionally builds radio maps that can be used to position a user indoors by manual surveying. Surveying is typically very costly in terms of human effort. This thesis is about crowdsourcing fingerprints so that traditional surveying tasks can be distributed to individual users, especially in a passive way so that the users do not have to take extra actions in order to contribute. Much existing research is based on the assumption that signals collected from multiple devices have to be unified before being merged into the radio map. For example, the signal strength reported by a new device has to be calibrated to a reference device. However, by empirical experiments conducted on two popular crowd-built heterogenous datasets, this thesis finds that the heterogeneity effect is negligible compared to noise, meaning that the heterogeneity will not be as much problematic as in crowdsourcing as it is in traditional fingerprinting research. Furthermore, using heterogenous fingerprints directly in map building outperforms using homogeneous data because the larger quantity of data compensates for the heterogeneity effect. Pedestrian Dead Reckoning (\GLS{PDR}) algorithms are powerful tools to recover walking trajectories from the inertial data. However, in crowdsourcing, inertial data is typically noisy and, thus, error-prone when processed by traditional PDR algorithms. Crowdsourcing can afford to filter out trajectories that are of poor quality and only merge the trustworthy trajectories. Existing PDR algorithms do not handle suboptimal cases properly, especially when the quality of trajectories deteriorates. This thesis proposes a gravity-verified PDR that exploits the observation that the estimated gravity is a good indicator of the quality of the recovered trajectory. The gravity-verified PDR segments trajectories into trustworthy pieces and discards those that cannot be further handled. This thesis proposed two approaches to match trustworthy trajectories to the floor plan in order to build the radio maps based on the idea of maximising the likelihood of observing the recorded fingerprints along the trajectories using Gaussian Processes. The first approach was verified in an office environment with a corridor as its major landscape feature, and achieved a mean positioning accuracy of 2.47 metres, comparable to 2.65 metres, the accuracy of the traditional fingerprinting. The second approach was verified in a public museum with actual visitors as the data contributor and a particle filter as alignment bootstrap aid, and achieved a mean positioning accuracy of 1.66 metres, comparable to 1.87 metres, the accuracy of the traditional fingerprinting. Both approaches achieved equivalent positioning performance to traditional manual surveying methods. The crowdsourced map grows gradually and may not cover all areas, which may lead to systematic errors in positioning when the fingerprint to be positioned is outside the covered area. This thesis also introduces fingerprint anomaly detection algorithm that distinguishes localisable fingerprints from noisy and uninformative fingerprints. With the anomaly detection, positioning algorithm can report the reliability of the positioning result to the users, improving the user experience. This thesis demonstrated a working prototype of the passive crowdsourced indoor positioning system that can greatly lower the human intervention needed for radio map building and maintenance.China Scholarship Counci