Grid-Supervised Flow Refinement for Wireless Localization

Abstract

This paper introduces FlowLoc, a framework for high-accuracy user equipment positioning from channel state information observed at multiple base stations. The input is a magnitude channel impulse response fingerprint derived from channel estimates. FlowLoc combines a learned initial position estimate with a flow model that corrects the remaining position error. First, FlowLoc trains base station encoders using grid supervision and then removes the grid classifiers. Second, it freezes the encoders and trains a network to estimate an initial position, called the anchor, from their outputs. Third, it freezes the anchor network and trains a flow model to correct the remaining position error. We evaluate FlowLoc using ray-traced channels from an outdoor urban region, with separate trajectories for training and testing. FlowLoc reduces mean localization error from 0.917m for a grid-based localization baseline adapted from prior work to 0.537m, a 41.4% reduction. It also outperforms supervised regression and direct-coordinate generative baselines. FlowLoc retains sub-meter mean error on independently sampled channels across mobility settings from pedestrian speeds to 25m/s. Further experiments quantify sensitivity to channel noise and missing base station observations.

arXiv comment13 pages. Substantially revised from v1, with a new title, method, and experimental evaluation