Cloud-based Indoor Positioning Platform for Context-adaptivity in GNSS-denied Scenarios

The demand for positioning, localisation and navigation services is on the rise, largely owing to the fact that such services form an integral part of applications in areas such as human activity recognition, robotics, and eHealth. Depending on the field of application, these services must accomplish high levels of accuracy, massive device connectivity, real-time response, flexibility, and integrability. Although many current solutions have succeeded in fulfilling these requirements, numerous challenges remain in terms of providing robust and reliable indoor positioning solutions.

This dissertation has a core focus on improving computing efficiency, data pre-processing, and software architecture for Indoor Positioning Systems (IPSs), without throwing out position and location accuracy. Fingerprinting is the main positioning technique used in this dissertation, as it is one of the approaches used most frequently in indoor positioning solutions. The dissertation begins by presenting a systematic review of current cloud-based indoor positioning solutions for Global Navigation Satellite System (GNSS) denied scenarios. This first contribution identifies the current challenges and trends in indoor positioning applications over the last seven years (from January 2015 to May 2022).

Secondly, we focus on the study of data optimisation techniques such as data cleansing and data augmentation. This second contribution is devoted to reducing the number of outliers fingerprints in radio maps and, therefore, reducing the error in position estimation. The data cleansing algorithm relies on the correlation between fingerprints, taking into account the maximum Received Signal Strength (RSS) values, whereas the Generative Adversarial Network (GAN) network is used for data augmentation in order to generate synthetic fingerprints that are barely distinguishable from real ones. Consequently, the positioning error is reduced by more than 3.5% after applying the data cleansing. Similarly, the positioning error is reduced in 8 from 11 datasets after generating new synthetic fingerprints.

The third contribution suggests two algorithms which group similar fingerprints into clusters. To that end, a new post-processing algorithm for Density-based Spatial Clustering of Applications with Noise (DBSCAN) clustering is developed to redistribute noisy fingerprints to the formed clusters, enhancing the mean positioning accuracy by more than 20% in comparison with the plain DBSCAN. A new lightweight clustering algorithm is also introduced, which joins similar fingerprints based on the maximum RSS values and Access Point (AP) identifiers. This new clustering algorithm reduces the time required to form the clusters by more than 60% compared with two traditional clustering algorithms.

The fourth contribution explores the use of Machine Learning (ML) models to enhance the accuracy of position estimation. These models are based on Deep Neural Network (DNN) and Extreme Learning Machine (ELM). The first combines Convolutional Neural Network (CNN) and Long short-term memory (LSTM) to learn the complex patterns in fingerprinting radio maps and improve position accuracy. The second model uses CNN and ELM to provide a fast and accurate solution for the classification of fingerprints into buildings and floors. Both models offer better performance in terms of floor hit rate than the baseline (more than 8% on average), and also outperform some machine learning models from the literature.

Finally, this dissertation summarises the key findings of the previous chapters in an open-source cloud platform for indoor positioning. This software developed in this dissertation follows the guidelines provided by current standards in positioning, mapping, and software architecture to provide a reliable and scalable system.