Author:

Abstract:

The development of indoor localization technologies enables location based services in environments where Global Navigation Satellite Systems like for example the Global Positioning System fail to provide reliable location information. These technologies have a large application potential in many different sectors. One use case can for example be found in the logistics sector where asset tracking and autonomous vehicle navigation is of interest. Other use cases can be indoor navigation assistance in for example airports or subterranean parking lots. Many indoor technologies have been discussed in literature or are already commercially available, but due to the complexity of indoor environments and its dynamics, the performance of different indoor positioning technologies is strongly dependent on the application scenario. As a consequence no default solution, such as the Global Positioning System in the outdoor scenario, is available. This research focuses on an indoor positioning solution that uses Optical Wireless Communication which relies on the advances in solid state lighting. More specifically, it uses visible light signals which originate from LEDs and are captured by a photodetector to infer location estimates. This is generally known as Visible Light Positioning. The focus of this work is on the application of Machine Learning methods such as Artificial Neural Networks and Gaussian Processes to enhance the performance of Received Signal Strength-based Visible Light Positioning. The achievable accuracy of classical RSS-based localization methods is limited by flaws in the propagation model, aberrant radiation patterns and the need of exactly quantifying every parameter in the system which is not feasible in practice. The use of ML-methods is a major improvement because they instead learn the underlying physical model by observing a training dataset. It is shown that they exhibit superior performance compared to the classical localization methods even when only a sparse amount of training data is available. The data-driven ML-methods were initially evaluated in a small scale setup and compared to the baseline localization methods such as multilateration. Finally, the data-driven strategy was demonstrated in an industrial scale setup where robustness and scalability were evaluated in this challenging and representative environment.