Channel Charting assisted Radio Localization: Towards Reliability and Robustness

Indoor localization plays an important role in various fields, such as health care, industrial production or smart shopping centers. Often, radio localization is employed due to its high accuracy, scalability, and energy efficiency. Classical ToA based localization may achieve centimeter accuracy, with LoS between receiver and transmitter. However, due to the complexity of indoor environments, the LoS is often blocked, lowering the accuracy and reliability dramatically. To enable localization in such challenging environments, fingerprint-based models can be employed. Fingerprint-based models learn a function mapping distinct radio signatures to positions given a labeled database. Those models are immune to NLoS propagation as they only require similar radio signatures at certain positions. However, due to the dynamic nature of productive environments, e.g. industrial facilities, fingerprints regularly change, rendering a maintenance process necessary, which includes an expensive labeling process of the recorded signals, i.e., CSI. Thus, in this thesis, novel fingerprinting concepts are contributed to reduce or eliminate the requirement for labeling and with the ability to identify corrupted fingerprints or signals not included in the training area. This cumulative thesis is built around three main publications. In the first publication a novel channel charting algorithm for ToA based localization systems is contributed. Channel charting allows to model the radio geometry, i.e., relative coordinates of CSI in space, to enable self-supervised fingerprinting. We formalized a novel geodesic distance metric between pairs of CIR measurements for high bandwidth radio systems. Our metric allows to model global channel charts, enabling self-supervised fingerprinting. The efficiency of our approach is investigated on two real-world datasets, highly outperforming supervised fingerprinting with only a few data samples available. The second publication contributes a radio system independent channel charting approach based on velocity information. Given noisy velocity estimation derived from, e.g., odometry systems or PDR systems, data-independent distances between CSI can be derived by simple integration over time. We also propose an adaptive map matching algorithm to entirely avoid reference positions. Our evaluations show that we achieve comparable results to supervised fingerprinting without the requirement of expensive ground truth reference positions. To ensure reliable fingerprinting, in the last publication a model monitoring approach is contributed. Our approach is based on the uncertainty estimation of neural networks to identify corrupted fingerprints and the spatial limitations of a fingerprinting model. This enables an efficient maintenance process and allows the deployment of fingerprinting models only in NLoS dominated areas. In conclusion, a novel fingerprinting concept is proposed, which exploits the intrinsic structure of measured CSI to lower the effort for labeling, while our model monitoring approaches enhance the reliability. In combination, the proposed algorithms help to lower the expenses for deployment and maintenance and enable robust radio localization in infrastructure-restricted, NLoS dominated areas, rendering radio-based fingerprinting a viable alternative to classical ToA radio localization.