Abstract

Ubiquitous presence of smart connected devices coupled with evolution of Artificial Intelligence (AI) within the field of Internet of Things (IoT) have resulted in emergence of innovative ambience awareness concepts such as smart homes and smart cities. In particular, IoT-based indoor localization has gained significant popularity, given the expected widespread implementation of 5G network, to satisfy the ever increasing requirements of Location-based Services (LBS) and Proximity Based Services (PBS). LBSs and PBSs have found several applications under different circumstances such as localization profiling for human resource management; navigation assistant applications in smart buildings/hospitals, and; proximity based advertisement and marketing. The focus of this thesis is, therefore, on design and implementation of efficient and accurate indoor localization processing and learning techniques. In particular, the thesis focuses on the following three positioning frameworks: (i) \textit{Bluetooth Low Energy (BLE)-based Indoor Localization}, which uses the pathloss model to estimate the user's location; (ii) \textit{Inertial Measurement Unit (IMU)-based Indoor Positioning}, where smart phone's $3$ axis inertial sensors are utilized to iteratively estimate the headings and steps of the target, and; (iii) \textit{Pattern Recognition-based Indoor Localization}, which uses Deep Neural Networks (DNNs) to estimate the performed actions and find the user's location. With regards to Item (i), the thesis evaluates effects of the orientation of target's phone, Line of Sight (LOS) / Non Line of Sight (NLOS) signal propagation, and presence of obstacles in the environment on the BLE-based distance estimates. Additionally, a fusion framework, combining Particle Filtering with K-Nearest Neighbors (K-NN) algorithm, is proposed and evaluated based on real datasets collected through an implemented LBS platform. With regards to Item (ii), an orientation detection and multiple-modeling framework is proposed to refine Received Signal Strength Indicator (RSSI) fluctuations by compensating negative orientation effects. The proposed data-driven and orientation-free modeling framework provides improved localization results. With regards to Item (iii), the focus is on classifying actions performed by a user using Long Short Term Memory (LSTM) architectures. To address issues related to cumulative error of Pedestrian Dead Reckoning (PDR) solutions, three Online Dynamic Window (ODW) assisted LSTM positioning frameworks are proposed. The first model, uses a Natural Language Processing (NLP)-inspired Dynamic Window (DW) approach, which significantly reduces the computation time required for Real Time Localization Systems (RTLS). The second framework is developed based on a Signal Processing Dynamic Window (SP-DW) approach to further reduce the required processing time of the two stage LSTM based indoor localization. The third model, referred to as the SP-NLP, combines the first two models to further improve the overall achieved accuracy.