Device-Free Wi-Fi Localization via Deep Learning on Two-Way CSI

We propose a novel method for device-free human detection and localization relative to reference points, leveraging two-way Wi-Fi channel state information (CSI) obtained from DATA/ACK exchanges and captured by a single-antenna passive receiver. By constructing a two-way CSI cross-correlation matrix and feeding its magnitude and phase into a compact one-dimensional convolutional neural network (CNN), the method amplifies subtle human-induced multipath variations while remaining fully network-independent. The system is evaluated in diverse environments and compared with conventional classifiers, including k-nearest neighbors (KNN) and support vector machines (SVM). Outdoor experiments over 27 reference points (RPs) achieve a best test accuracy of 96.22% (mean error ≈ 1m), outperforming KNN (37.83%) and SVM (91.88%). In two indoor scenarios (passive receiver inside vs. outside), best test accuracies of 98.36% and 98.86% are obtained, respectively. The approach requires no AP/client modifications, generalizes across device positions and days, and is robust to ACK timing jitter around SIFS.