ENHANCING THE PERFORMANCE OF VISIBLE LED LIGHTS BASED INDOOR POSITIONING SYSTEM USING MACHINE LEARNING

Abstract

Recently, visible light positioning (VLP) applications have attracted a great deal of research attention because of the extremely high positioning accuracy that this system can provide. VLP continues to emerge as a leading candidate in the field of indoor positioning because of its advantages of accuracy, cost effectiveness, and simplicity even though it faces some difficulties, including multipath reflection, light interference between Light Emitting Diode (LED) lights, and noises from sunlight and artificial light sources. The main target of this thesis is to gradually develop VLP systems from simulation to experiment by building and improving machine learning (ML) algorithms as well as solving some inherent limitations of the LED-based VLP system. To minimize the computational time when applying ML algorithms, the novel adoption of dual-function ML is employed. These algorithms have a common feature that contains two functions: classification and regression. The division of the experimental testbed into specific areas by classification function significantly reduces the execution time for the positioning process. On the other hand, the regression function of the proposed dual-function ML algorithms plays an important role in estimating and improving the positioning accuracy. To solve the multipath reflection effects, a combination of random forest and K-nearest neighbor (KNN) algorithms is used to improve the positioning accuracy at the areas outside the center where the appearance of noise due to reflection is a serious problem. In addition to simulation, an improved algorithm of KNN, namely weighted optimum KNN (WOKNN), is applied to the real model to prove the feasibility and practicality of the proposed VLP system. The results show that, the proposed solution enhances the positioning accuracy to the millimeter level. During the implementation of the WOKNN algorithm, the fact is that the number of initially labeled fingerprints greatly affects the positioning performance. The more the offline fingerprints the system has, the more accurate the system achieves. It is an obstacle to leverage the fingerprinting method in larger spaces. To eliminate this weakness, the Ada-XCoReg algorithm is suggested. This is a combination of co-training semi-regression learning and adaptive boosting algorithms. The experimental results show that a mean positioning error of approximately 6 cm is achieved although the number of labeled fingerprints is reduced by roughly 90 percent. By applying this approach, a high positioning accuracy is maintained while the number of fingerprints became more feasible for real applications.