Algorithm-Supervised Millimeter Wave Indoor Localization using Tiny Neural Networks

TL;DR

The paper tackles indoor mmWave localization under limited environmental knowledge and training data by proposing a tiny neural network trained in an algorithm-supervised manner. It maps $ADOA$ inputs derived from MPC clusters to 2D receiver coordinates, with bootstrapped labels provided by JADE to avoid costly ground-truth data collection. A recursive DBSCAN-based clustering pipeline yields robust dominant MPCs, and a four-layer network with ReLU activations achieves sub-meter accuracy in a majority of cases while offering much lower computational load than JADE. Real 60 GHz measurements and supplementary simulations demonstrate practical viability for device-centric localization on resource-constrained hardware, enabling scalable, edge-friendly indoor localization. The approach balances accuracy, data-effort, and compute, and points to future work on adaptive switching between bootstrapping and NN inference in dynamic environments.

Abstract

The quasi-optical propagation of millimeter-wave signals enables high-accuracy localization algorithms that employ geometric approaches or machine learning models. However, most algorithms require information on the indoor environment, may entail the collection of large training datasets, or bear an infeasible computational burden for commercial off-the-shelf (COTS) devices. In this work, we propose to use tiny neural networks (NNs) to learn the relationship between angle difference-of-arrival (ADoA) measurements and locations of a receiver in an indoor environment. To relieve training data collection efforts, we resort to a self-supervised approach by bootstrapping the training of our neural network through location estimates obtained from a state-of-the-art localization algorithm. We evaluate our scheme via mmWave measurements from indoor 60-GHz double-directional channel sounding. We process the measurements to yield dominant multipath components, use the corresponding angles to compute ADoA values, and finally obtain location fixes. Results show that the tiny NN achieves sub-meter errors in 74% of the cases, thus performing as good as or even better than the state-of-the-art algorithm, with significantly lower computational complexity.

Figures (12)

• Figure 1: Workflow of our proposed localization scheme.
• Figure 2: Illustration of virtual anchor and first-order reflections. Here, $\theta_1$ and $\theta_2$ represent the aoa of the los and the nlos paths, and $\alpha = \theta_1 - \theta_2$ represents the adoa with respect to the los path. Note that first-order reflections have the same geometrical properties (delay and AoA) of direct rays generated from virtual anchors.
• Figure 3: (a) Measurement setup and (b) floor plan of the room for the 3D double directional 60 GHz channel sounder experiment conducted at NIST to collect indoor mmWave channel responses. (c) Dataset resulting from processing the channel estimation data. The areas are divided and annotated based on the number of tx from which the rx captures the signal. Axis unit: [m]. Note the different axis scale.
• Figure 4: mpc clustering using our proposed recursive approach for a sample rx location in area 4. Note that the triangles represent the cluster centroids from the initial dbscan procedure and the squares represent the centroids obtained after recursive clustering. (a) Reference ray traces of the LoS and first-order mpc from tx-1 to the rx. (b) Clustering of the mpc from tx-1.
• Figure 5: CDF of the location estimation errors for our proposed 3-hidden layer (HL) NN (referred to as AS-TNN) against the previous 2-hidden layer model and other state-of-the-art algorithms. Note that our tiny NN trained with ground truth is referred to as TNN.