Development of localization algorithms for a WiFi based indoor positioning system with machine learning techniques

Abstract

In recent years, the prevalence of mobile devices and the popularity of social networks have spurred extensive demands on Location Based Services (LBSs). Whereas GPS has been extensively adopted in outdoor positioning, it can't provide accurate enough indoor positions due to non-line-of-sight (NLOS) transmission channels between the receiver and the satellite in indoor environments. Thus, developing an Indoor Positioning System (IPS) that is capable of providing reliable and accurate LBSs is widely studied. A large body of WiFi-fingerprinting based indoor localization solutions emerge as WiFi is accessible for users and cost-efficient for developers. On the other hand, it has been demonstrated in literature that machine learning techniques can be applied to IPSs yielding satisfactory localization results in real-time. In this thesis, we aim to develop accurate and reliable localization algorithms for WiFi based IPS by machine learning (ML) techniques. We model the indoor positioning problem under a non-parametric stochastic framework, and modify the well-known ML tool, extreme learning machine (ELM), to achieve the above goal. Firstly, under the assumption that noises merely lie in input data, we modify ELM by introducing a dead zone, which is called DZ-ELM, and integrate it into our IPS. We analyse the consistency of DZ-ELM for different types of disturbances. Experimental results show that the DZ-ELM based IPS can not only provide higher accuracy, but also improve the repeatability of IPSs. Secondly, when assuming that noises lie in both input and output data, we exploit the fact that feature mapping in ELM is known to users to develop two kinds of robust ELM (RELM) based on second order cone programming. The simulation and real-world indoor localization experimental results both demonstrate that the proposed algorithm can not only improve the accuracy and repeatability, but also reduce the deviations and worst case errors of IPSs compared with other baseline algorithms. Lastly, beyond IPSs, we employ clustering algorithms to build up a clustering based zonal occupancy monitoring system by existing WiFi infrastructures to model the distribution of indoor occupants. The system is lightweight and free of calibration, which can be further incorporated in real applications such as optimization and control of building heating, ventilating, and air conditioning (HVAC) systems.