Project

Back to overview

Eratosthenes: Deep generative modeling for indoor and outdoor positioning with fingerprinting methods

Applicant Anagnostopoulos Grigorios
Number 195964
Funding scheme Spark
Research institution
Institution of higher education University of Applied Sciences and Arts Western Switzerland - HES-SO
Main discipline Information Technology
Start/End 01.10.2020 - 31.12.2021
Approved amount 142'763.00
Show all

All Disciplines (2)

Discipline
Information Technology
Electrical Engineering

Keywords (8)

outdoor positioning; indoor positioning; localisation; generative modelling; deep learning; machine learning; IoT; variational autoencoders

Lay Summary (French)

Lead
L'ambition du projet Eratosthenes est d'intégrer les avancées récentes du domaine de la modélisation générative dans le domaine du positionnement intérieur et extérieur. Les méthodes de localisation «fingeprinting» (empreintes digitales) utilisant l’apprentissage automatique offrent généralement une haute précision de localisation mais, en revanche, ils souffrent de l'inconvénient de nécessiter d'une phase de collection de données qui est très coûteuse. Dans le cadre de ce projet, nous proposons, donc, de s'affranchir en partie de cette contrainte par l’introduction de la modélisation générative.
Lay summary

La recherche dans le domaine du positionnement intérieur et du positionnement des objets IoT n'a pas systématiquement intégré les récentes avancées de la modélisation générative. Dans le cadre du projet Eratosthenes nous allons, premièrement, étudier le concept d'augmentation des «fingeprints» grâce à la modélisation générative, car atteindre une génération fiable des «fingeprints» pourraient avoir un grand impact sur la façon dont les systèmes de positionnement des empreintes digitales sont conçus, réglés et déployés. Deuxièmement, nous examinerons les avantages potentiels du codage des «fingeprints» au travers d'une représentation plus compact, appelé espace latent, modélisé par des auto-encodeurs variationnels (VAEs). Finalement, la partie la plus ambitieuse du projet Eratosthenes est l'incorporation de crowdsourcing, visant à l’utilisation des données non étiquetées («fingeprints» à des positions inconnues), afin de renforcer les capacités prédictives des modèles de positionnement en utilisant une approche semi-supervisée. Actuellement, tout le volume de données non étiquetées qui est généré pendant le fonctionnement des systèmes de positionnement est uniquement utilisé pour alimenter les modèles qui donnent les estimations de position présentées aux utilisateurs. L’utilisation de ces énormes volumes de données pour apprendre une meilleure représentation latente et pour améliorer la performance du modèle prédictif de localisation constituerait un changement de paradigme pour le domaine.

 

Direct link to Lay Summary Last update: 31.08.2020

Responsible applicant and co-applicants

Employees

Abstract

The proliferation of Location Based Services through smart devices has supported a variety of services. Unlike outdoor positioning, where systems such as the GPS or Galileo are considered a de facto standard, indoor positioning has presented systems utilizing a big variety of technologies and positioning methods. The main categorical division of these methods is set between ranging and fingerprinting methods. Moreover, the recent emergence of the usage of Internet-of-Things (IoT) technologies and their usage in creating Low Power Wide Area Networks (LPWAN) has shaped a new landscape in the field of outdoor localization. The low-power devices used in LPWANs cannot afford the battery consumption of a Global Navigation Satellite System (GNSS) chip set, such as the GPS. Therefore, an alternative approach is needed in order to localize these low power devices. Similarly, to the indoor positioning, ranging and fingerprinting methods are two main competing approaches for LPWAN positioning as well. Fingerprinting methods generally offer a higher accuracy of localization but suffer from the disadvantage of requiring a tedious and costly data collection phase. Since fingerprinting is a purely data-driven approach, the way data are collected, the volume of the collected data and the way they are utilized in the modeling phase determine the cost and the performance of such methods. In practice, the main barrier that often prevents the selection of fingerprinting is the data collection requirement.In the Eratosthenes (deep gEneRative modeling for indoor And ouTdoor pOSitioning wiTH fingErprintiNg mEthodS) project we propose a reproducible and systematic study of the potential contributions of generative modeling in fingerprinting localization methods. Our envisaged utilization of generative models is threefold. Initially, (i) we will investigate the concept of fingerprint augmentation through generative modeling. Achieving a reliable generation of fingerprints could have a great impact on how fingerprinting positioning systems are designed, tuned and deployed. Secondly, (ii) we will examine the potential benefits of encoding the fingerprints in the more compact representation of a latent space provided by (variational) autoencoders and operating lightweight ML models on this latent space. Lastly, the most ambitious part of Eratosthenes is (iii) the incorporation of crowdsourced, unlabeled data, in order to strengthen the predictive capabilities of the positioning model, by using Semi-Supervised Learning with deep generative models. Currently, all the volume of unlabeled data that occurs during the operation of positioning systems is only used to feed the model which produces position estimates. Utilizing these huge volumes of data for learning a better latent representation and improving the positioning model would constitute a paradigm shift for the field. The indoor positioning research hasn't integrated in a systematic way the recent breakthroughs of the machine learning field. Moreover, during the last few years, there has been an effort to strengthen the consistent evaluation, the comparability and the reproducibility of published works of the field. In this regard, and in terms of the project’s implementation, we will use the most relevant public datasets, that function as benchmarks and will proceed with an open-data, open-code policy. The ambition of the project is to offer a systematic, rigorous, unambiguous, reproducible and consistent overview of the capabilities of generative models to improve fingerprinting techniques, which is a missing piece of the puzzle of the field. The unconventionality of the approach and the potential of a scale-up improvement of the trade-off between improved accuracy and reduced data collection effort is truly appealing. Moreover, the impact of potentially mitigating the barrier of the data collection effort would be the shaping of a new landscape in deployed positioning systems.
-