Continuous Machine Learning for Assisting AR Indoor Navigation

Dimitrios Kalathas, Dionysios Koulouris, Andreas Menychtas, Panayiotis Tsanakas, Ilias Maglogiannis

Abstract: This study aims to facilitate the localization abilities of Mobile Augmented Reality Navigation systems by applying a Continuous Machine Learning approach. AR Navigation for mobile devices is accomplished by displaying navigation arrows in the frontal space as part of a visualized guidance path. It is achieved by extracting image Feature Points of the space using the Phone Camera, storing them in the Cloud, resolving them in the retrospect and comparing them with newly extracted Feature Points. A limitation of this approach, the solving of which is the motivation of this study, is the system’s lack of flexibility in changes or displacements of the environment. We studied the utilization of Artificial Intelligence methods and examined the integration of Machine Learning for optimizing such processes. The proposed system not only records a route and stores the path-related image Feature Points to the Cloud, but also has the ability to extract Feature Points each time a navigation is performed by a visitor to enrich the model of the path and record possible changes of the environment that occurred after the initial training. A proof-of-concept prototype named "NtuARguidance" was tested for both route recording and resolving at the National and Technical University of Athens Campus with the goal of deploying a final application available to the University visitors in the future. The prototype was shown to be less sensitive to environmental changes, compared to previous more simplistic approaches, highlighting a direction towards optimizing the AR Indoor Navigation experience.