Learned fingerprinting-based models for reliable localization in large buildings

Laska, Marius; Blankenbach, Jörg (Thesis advisor); Krisp, Jukka (Thesis advisor)

Aachen : RWTH Aachen University (2023)
Dissertation / PhD Thesis

Dissertation, Rheinisch-Westfälische Technische Hochschule Aachen, 2023


Location-based services (LBS) such as pedestrian navigation or location-aware search are ubiquitous in our daily lives, but how can it be explained that - although we humans spend a large part of our time in buildings - LBS are hardly available indoors? The main reason is that existing technologies such as global navigation satellite systems (GNSS) do not function reliably indoors as the occurring signal attenuation and multi-path propagation distort the positioning. The search for alternative technological solutions has been initiated many years ago but still represents an ongoing research field with heterogenous approaches proposed. More recently, a technique called fingerprinting has paved the way towards mass-market pedestrian indoor localization. The assumption is that, depending on the position in the building, a characteristic fingerprint can be observed, which is commonly composed of the received signal strengths (RSS) of the available transmitting stations (e.g., WLAN access points). The relation between the fingerprints and their locations can be learned in a supervised setting. This caused the momentum in deep learning to be carried over to the field of fingerprinting-based localization recently. However, there are several theoretical as well as practical open challenges associated with RSS-based fingerprinting. On the theoretical side, RSS is unstable due to effects such as multi-path propagation. Even small position updates or minor environmental changes can cause the RSS to largely differ. Artificial neural networks (ANN) have the ability to learn approximations of the complex localization function using noisy features such as RSS. However, the nature of RSS limits the theoretically achievable positioning accuracy. On the practical side, the necessity of a labeled dataset increases the effort for setting up fingerprinting-based systems. Further, deploying trained models on end-user devices requires compatibility of models with the available hardware. Within this context the following research questions emerge that are addressed in the present cumulative dissertation. Considering the limitations of RSS-fingerprinting, how can the reliability of ANN-based indoor localization be increased? How can the digital building model be leveraged during training of the ANN? How can models be designed that allow for efficient positioning on end-user devices even for large-scale multi-floor buildings? Further, how can the collection of labeled fingerprints be simplified while preserving the required labeling accuracy for training supervised models? To increase the reliability of the location estimation, it is proposed to estimate the current location area instead of pin-pointing the exact position of the user. The assumption is that a coarser position estimate can be provided with a higher reliability. As an alternative to classifying the area based on a given floorplan segmentation, a novel family of models is introduced that learns to estimate the user's area only conditioned on a given input fingerprint. By integrating the building model directly within the learning phase of the model, predicted areas match with the underlying building structure while their individual shapes are learned rather than fixed by the presegmentation. Standard metrics such as the classification accuracy do not suffice for evaluating area localization models as the granularity of the estimated areas and their congruency with the underlying digital building model influence the information gain of the user (model expressiveness). Therefore, a metric is proposed to evaluate area localization models among the identified quality dimensions while preserving interpretability for classical point estimation models. After evaluation it can be concluded that with the introduced area localization models the reliability in the location estimate can be increased. The developed area localization models can be scaled towards large multi-floor environments by only requiring a single forward-pass of a single ANN. This avoids complex multi-model maintenance and eases deployability on smartphones. Further, large-scale localization is explicitly tackled via multi-task learning (joint building/floor classification and coordinate estimation). The proposed model performs grid cell classification and simultaneously estimates the coordinates within the local coordinate system of the chosen grid cell. To avoid large errors in case of grid cell misclassifications, the multi-cell encoding learning (m-CEL) technique is proposed, in which the model is trained to learn several redundant position representations within a slightly overlapping grid-cell encoding. The model is shown to outperform state-of-the-art multi-task learning models on several public benchmark datasets. One limitation of fingerprinting is its dependence on previously collected labeled data. To reduce the required data collection effort, a pipeline for auto-labeling the collected fingerprints is presented. Based on visual inertial simultaneous localization and mapping (VI-SLAM), a smartphone application is developed that logs its positions during fingerprint acquisition. In a subsequent case study, it is shown that with the developed pipeline the introduced positioning models can be efficiently deployed in a multi-floor university building. The introduced models as well as the data collection pipeline are available open-source to foster future research with respect to indoor positioning as well as for serving as black box input for applied research in the domain of indoor location based services.


  • Chair of Building Informatics and Geoinformation Systems [316210]