Learned Multi-Sensor Indoor Positioning of Mobile Devices

Abstract: Attenuation of the microwave signal used in the Global Positioning System (GPS) due to interaction with building structures complicates the task of accurate indoor positioning. For this reason, research on alternative approaches is being performed at universities and companies worldwide. The Global Indoor Navigation (GIN) research project in Lund is one such initiative, where the main goal is to perform indoor positioning in a vast number of buildings globally rather than in specific buildings locally. This thesis proposes and investigates several methods and algorithms for indoor positioning with applications in the GIN project that are adaptive to different buildings, users and devices. The main contribution is a novel multi-purpose deep learning architecture and training procedure which leverages received signal strength indications (RSSI) from Wi-Fi and Bluetooth beacons. The approach achieves a mean positioning error that is comparable with state of the art WKNN methods, offering improved inference speed. During one training session, three multi-layer perceptrons are generated, which all can be used separately in different applications. In addition to the development of this model, it is also investigated how the resulting position estimates can be combined with estimated displacements in position that are based on inertial sensors in mobile devices. This part also shows promising results, as the investigated approach decreases the mean positioning error from 7.2 m to 5.4 m for the largest dataset used. The models and algorithms were evaluated on two fingerprinting datasets, and in a Kaggle competition where these contributed to the second place entry.