Multiple Self-Supervised Tiny Neural Networks for mmWave Localization in Complex Indoor Areas

We propose to employ multiple tiny neural network (NN) models to localize a mobile client in large or complex indoor environments. These models are trained in a self-supervised fashion by exploiting the location labels obtained from a geometry-based bootstrapping localization algorithm, thus relieving the burden of training data collection. We further propose a scheme to switch to the right NN model in order to keep localizing a mobile client accurately as it moves through the indoor space.