Analytical Study on Raw CSI for Posture-Based Wi-Fi Sensing: Insights for Efficient Machine Learning Classifiers

Abstract

Device-free posture recognition using WiFi Channel State Information (CSI) is attractive because it leverages existing infrastructure, preserves privacy, and captures rich propagation dynamics. Yet most work optimizes classifiers directly on CSI without first establishing which raw signal components and spatial factors are intrinsically discriminative. This paper provides a constructive, data-centric analysis of raw CSI collected with two receivers over 114 subcarriers in a controlled anechoic chamber, across 28 spatial coordinates and two posture classes (sitting and standing) performed by one subject. We present an analytical framework covering multipath propagation geometry, phase discriminability, and preprocessing transforms, with the assumptions and the level of each result stated explicitly: the phase-dominance result is conditional on a dominant-path channel, the spatial result is a property of the discriminability metric rather than of classifier error, and the position/posture decomposition is a modeling interpretation rather than a derived classifier-level guarantee. We then report a systematic comparison of seven preprocessing methods across all 28 coordinates using a defined mean-offset discriminability metric, and a spatial heatmap analysis of mean phase for both posture classes. We further evaluate three feature representations under both 10-fold stratified and Leave-One-Position-Out (LOPO) validation, and report separately that residual and recording-level PCA-STFT features improve spatial generalization over raw CSI at unseen coordinates while using compact, interpretable descriptors. We also perform complementary cross-dataset validation on two public CSI benchmarks, showing that preprocessing behavior is regime-dependent under room-, subject-, and environment–cohort shifts. Within the scope of this single-subject, two-posture, single-geometry experiment, our findings motivate posture detection pipelines that are phase-priority, spatially aware, and preprocessing-selective, supporting compact and interpretable feature representations for efficient machine learning classifiers.