Physics-Informed Convolutional Neural Network for Indoor Localization

Abstract

The received signal strength indicator (RSSI) from wireless access points in indoor environments can be employed for user localization. The accuracy of RSSI-based localization can be greatly improved from advanced knowledge of the propagation characteristics of an environment, via extensive measurements or computationally costly simulations. This paper introduces a machine learning approach, leveraging a convolutional neural network, aimed at producing high-resolution power maps of complex indoor environments through low-cost ray-tracing simulations. The produced power maps are integrated with a k-nearest neighbors (kNN) algorithm that performs user localization. The proposed approach is successfully demonstrated in a localization case study across the floor of an office building at the University of Toronto campus.