RF fingerprinting can solve the indoor positioning problem with satisfactory accuracy, but the methodology depends on the so-called radio map calibrated in the offline phase via manual site-survey, which is costly, time-consuming and somewhat error-prone. It also assumes the RF fingerprint’s signal-spatial correlations to remain static throughout the online positioning phase, which generally does not hold in practice. This is because indoor environments constantly experience dynamic changes, causing the radio signal strengths to fluctuate over time, which weakens the signal-spatial correlations of the RF fingerprints. State-of-the-arts have proposed adaptive RF fingerprint methodology capable of calibrating the radio map in real-time and on-demand to address these drawbacks. However, existing implementations are highly server-centric, which is less robust, does not scale well, and not privacy-friendly. This thesis aims to address these drawbacks by exploring the feasibility of implementing an adaptive RF fingerprint indoor positioning system in a distributed and client-centric architecture using only commodity Wi-Fi hardware, so it can seamlessly integrate with existing Wi-Fi network and allow it to offer both networking and positioning services. Such approach has not been explored in previous works, which forms the basis of this thesis’ main contribution. The proposed methodology utilizes a network of distributed location beacons as its reference infrastructure; hence the system is more robust since it does not have any single point-of-failure. Each location beacon periodically broadcasts its coordinate to announce its presence in the area, plus coefficients that model its real-time RSS distribution around the transmitting antenna. These coefficients are constantly self-calibrated by the location beacon using empirical RSS measurements obtained from neighbouring location beacons in a collaborative fashion, and fitting the values using path loss with log-normal shadowing model as a function of inter-beacon distances while minimizing the error in a least-squared sense. By self-modelling its RSS distribution in real-time, the location beacon becomes aware of its dynamically fluctuating signal levels caused by physical, environmental and temporal characteristics of the indoor environment. The implementation of this self-modelling feature on commodity Wi-Fi hardware is another original contribution of this thesis. Location discovery is managed locally by the clients, which means the proposed system can support unlimited number of client devices simultaneously while also protect user’s privacy because no information is shared with external parties. It starts by listening for beacon frames broadcasted by nearby location beacons and measuring their RSS values to establish the RF fingerprint of the unknown point. Next, it simulates the reference RF fingerprints of predetermined points inside the target area, effectively calibrating the site’s radio map, by computing the RSS values of all detected location beacons using their respective coordinates and path loss coefficients embedded inside the received beacon frames. Note that the coefficients model the real-time RSS distribution of each location beacon around its transmitting antenna; hence, the radio map is able to adapt itself to the dynamic fluctuations of the radio signal to maintain its signal-spatial correlations. The final step is to search the radio map to find the reference RF fingerprint that most closely resembles the unknown sample, where its coordinate is returned as the location result. One positioning approach would be to first construct a full radio map by computing the RSS of all detected location beacons at all predetermined calibration points, then followed by an exhaustive search over all reference RF fingerprints to find the best match. Generally, RF fingerprint algorithm performs better with higher number of calibration points per unit area since more locations can be classified, while extra RSS components can help to better distinguish between nearby calibration points. However, to calibrate and search many RF fingerprints will incur substantial computing costs, which is unsuitable for power and resource limited client devices. To address this challenge, this thesis introduces a novel algorithm suitable for client-centric positioning as another contribution. Given an unknown RF fingerprint to solve for location, the proposed algorithm first sorts the RSS in descending order. It then iterates over this list, first selecting the location beacon with the strongest RSS because this implies the unknown location is closest to the said location beacon. Next, it computes the beacon’s RSS using its path loss coefficients and coordinate information one calibration point at a time while simultaneously compares the result with the measured value. If they are similar, the algorithm keeps this location for subsequent processing; else it is removed because distant points relative to the unknown location would exhibit vastly different RSS values due to the different site-specific obstructions encountered by the radio signal propagation. The algorithm repeats the process by selecting the next strongest location beacon, but this time it only computes its RSS for those points identified in the previous iteration. After the last iteration completes, the average coordinate of remaining calibration points is returned as the location result. Matlab simulation shows the proposed algorithm only takes about half of the time to produce a location estimate with similar positioning accuracy compared to conventional algorithm that does a full radio map calibration and exhaustive RF fingerprint search. As part of the thesis’ contribution, a prototype of the proposed indoor positioning system is developed using only commodity Wi-Fi hardware and open-source software to evaluate its usability in real-world settings and to demonstrate possible implementation on existing Wi-Fi installations. Experimental results verify the proposed system yields consistent positioning accuracy, even in highly dynamic indoor environments and changing location beacon topologies.

Gaze tracking has already shown to be a popular technology for desktop devices. When it comes to gaze tracking for mobile devices, however, there is still a lot of progress to be made. There’s still no high accuracy gaze tracking available that works in an unconstrained setting for mobile devices. This work makes contributions in the area of appearance-based unconstrained gaze estimation. Artificial neural networks are trained on GazeCapture, a publicly available dataset for mobile gaze estimation containing over 2 million face images and corresponding gaze labels. In this work, Siamese neural networks are trained to learn linear distances between face images for different gaze points. Then, during inference, calibration points are used to estimate gaze points. This approach is shown to be an effective way of utilizing calibration points in order to improve the result of gaze estimation. / Ögonblickspårning har redan etablerat sig som en populär teknologi för stationära enheter. När det dock gäller mobila enheter så finns det framsteg att göra. Det saknas fortfarande en lösning för ögonblickspårning som fungerar i en undantagsfri miljö för mobila enheter. Detta examensarbete ämnar att bidra till en sådan lösning. Artificiella neurala nätverk tränas på GazeCapture, en allmänt tillgänglig datasamling som består av över 2 miljoner ansiktsbilder samt korresponderande etikett för ögonblickspunkt. I detta examensarbete tränas Siamesiska neurala nätverk för att lära sig det linjära avståndet mellan två ögonblickspunkter. Sedan utnyttjas en samling med kalibreringsbilder för att estimera ögonblickspunkter. Denna teknik visar sig vara ett effektivt sätt att nyttja kalibreringsbilder med målet att förbättra resultatet för ögonblicksestimering.