AbstractTraditional approaches to three-dimensional (3-D) visible light positioning (VLP) suffers significantly in the presence of multipath propagation. This thesis overcomes such challenges by adopting a novel spatial and angular diversity receivers and combining
them with various machine learning (ML) algorithms for indoor, dark, passive and outdoor VLP. This thesis uses light emitting diode (LED)s as transmitters and photodiode (PD)s as receivers. To ensure that realistic channel models are used, the VLP
model includes line-of-sight (LOS), non-LOS (NLOS) for all indoor applications. Only LOS path is considered in the outdoor as the effect of NLOS from the road is ignored. However, the outdoor applications consider the impact of weather condition. A range
of ML approaches were considered, however, it is found that multi-layer perceptron (MLP) network offers the best performance and the lowest complexity for VLP applications. Using Levenberg Marquardt, the MLP hyper-parameters are tuned for each
application to ensure good performance and generalisability.
The results for each indoor application demonstrates the benefit of combining ML technique with received signal strength (RSS) based VLP. The ML technique offers good 3-D indoor VLP that is further improved using spatial receiver diversity, resulting
in a positioning error of 0.021 m in a 5 m3 room. The application to dark VLP, which uses a very low duty cycle pulse width modulation (PWM), resulted in a slightly higher positioning error of 0.06 m, which is still a 52% positioning accuracy improvement
compared to state-of-the-art dark VLP techniques. The more challenging passive VLP led to an RMS error of 0.23 m for a solution involving 9 transmitters and 21 receivers placed on the ceiling and walls in a 5 m ×5 m ×3 m room dimension.
MLs are demonstrated for outdoor vehicular applications with traffic lights and streetlights. The two-dimensional VLP using angular and spatial receiver diversity is able to overcome the streetlight collinearity condition resulting in 0.22 m RMS error in
the presence of direct sunlight, 0.29 m for dense fog and 0.14 m at night. Traffic light VLP with two receivers facing the direction of travel led to positioning errors of 1.33 m and 0.21 m using a single and double traffic light on the road, respectively. This represented a 77% and 47% improvement with the state-of-the-art traffic light-based VLP technique. These results highlight the degrading effect of NLOS and weather conditions in VLP and how ML techniques, together with spatial and angular receiver diversity
scheme can be used to offer improved accuracy for outdoor and indoor applications.
|Date of Award||2022|
|Supervisor||Zahir Ahmad (Supervisor), Olivier Haas (Supervisor) & Sujan Rajbhandari (Supervisor)|