David Bergh presents his Master's Thesis: Scalable Gaussian Process Assisted Positioning Using Cell Fingerprinting, Wednesday 13/3 at 10:15 in MH:332B (and on zoom).

Abstract:
GNSS is currently the most prominent method for outdoor positioning, providing an accuracy of approximately 5 metres in good conditions. However, for environments such as dense city areas or indoors a GNSS signal reception might not be possible. An alternative method of positioning is the fingerprint-based method, which has been widely adopted for indoor positioning by sampling the signal strength of WiFi beacons. This method can also be deployed for outdoor positioning by using the signal strength of mobile network cells. 

This thesis presents several fingerprint-based positioning models using Gaussian processes (GPs) and Sparse Gaussian Processes (SGPs) to model the signal strength propagation. A combination of global and local optimisation is used to find the position of highest likelihood, which yields our predicted position.

Two baseline positioning methods were used to compare with our SGP models, namely the Weighted K-Nearest Neighbours (WKNN) algorithm and the Weighted Centroid Localisation (WCL) algorithm. The WKNN achieved an approximately 20% smaller mean error when validated against the most extensive and realistic dataset. When validating with data not suited for WKNN our proposed model performed with approximately 33% better accuracy than that of WCL.

Supervisors:
Gabrielle Flood, Centre for Mathematical Sciences
Magnus Oskarsson, Centre for Mathematical Sciences
Martin Larsson, Combain

Examiner:
Alexandros Sopasakis, Centre for Mathematical Sciences

Zoom link: https://lu-se.zoom.us/j/67684654204