Channel Covariance based Fingerprint Localization

We study performance and complexity of fingerprint localization based on 5G signaling. We concentrate on channel covariance and Channel Impulse Response (CIR) features, studying the effect of several factors on the localization performance such as the channel bandwidth, the number of Base Stations (BSs), the number of antennas at each BS, and the number of time samples. We consider Weighted K Nearest Neighbour (WKNN) as well as Deep Neural Network (DNN) localization. We adopt DNNs based on the Rel-18 3GPP Study Item AI/ML for positioning accuracy enhancement. Simulation results show that channel covariance features outperform CIR in terms of localization accuracy. Furthermore, covariance-based features are robust with respect to bandwidth reduction, allowing for more power-efficient implementations. However, a noticeable dependency on the number of BSs, BS antennas, and time samples, is found. Results also show that increasing sampling density is much more beneficial for improving performance with CIR-based features. Again this highlights the power saving virtues of using covariance based features as input. Finally, results show that WKNN performs better with covariance-based features, with noticeable degradation in performance, when CIR features are used instead.