Algorithm-Supervised Millimeter Wave Indoor Localization using Tiny Neural Networks

Published

January 17, 2025

Author(s)

Anish Shastri, Steve Blandino, Camillo Gentile, Chiehping Lai, Paolo Casari

Abstract

The quasi-optical propagation of mmw signals has been instrumental in designing high-accuracy localization algorithms. These algorithms either exploit the geometric properties of the mmw signals or are based on an environment's radio footprints, leading to automated learning models to infer the location of a device. However, most of these algorithms pose stringent constraints: they may require information on the indoor environment, may entail the generation and collection of large datasets to train large learning models or their computational complexity may prevent implementation on cots devices. In this work, we propose to use tiny neural network models to learn the complex relationship between the adoa measurements and the locations of a receiver moving in an indoor environment. In order to relieve extensive training dataset collection efforts, we propose a self-supervised strategy to train our models using location estimates of the receiver obtained using a geometry-based localization algorithm. We evaluate our models using simulations as well as measurements obtained from an extensive measurement campaign using a 60-GHz mmw double-directional channel sounder. In order to associate the aoa of the reflected mpc to the reflecting surfaces, we propose an algorithm that recursively clusters the mpc to obtain the dominant mpc. The angles associated with the mpc are then used to generate adoa measurements, and in turn localize the receiver using both the geometry-based algorithm and the neural network model. The results of our experimental evaluations show sub-meter accuracy in 74% of the cases, when our model is trained with location estimates obtained using the real channel measurements, thus performing as good as or even better than the state-of-the art algorithms. The lower computational complexity of the model also caters to an easier implementation on cots device.

Citation

IEEE Transactions on Wireless Communications

Pub Type

Journals

Download Paper

Local Download

Citation

Shastri, A. , Blandino, S. , Gentile, C. , Lai, C. and Casari, P. (2025), Algorithm-Supervised Millimeter Wave Indoor Localization using Tiny Neural Networks, IEEE Transactions on Wireless Communications, [online], https://tsapps.nist.gov/publication/get_pdf.cfm?pub_id=936955 (Accessed February 5, 2025)