Source-linked AI summary
Through-Wall Detection using Software-Defined Radio based on adaptive Principal Component Analysis
Dinuli Naotunna, Wenchao Li, Sanka Piyaratna, Phil Wandel
TL;DR
Detecting motion behind walls from WiFi is complicated by direct-path power and noise masking Doppler effects, while fixed PCA component assumptions may fail under broadband noise. The paper uses spectral scoring to adaptively select CSI principal components, and experiments report improved performance and clearer, more localized wavelet energy structures than conventional PCA.
Problem
Direct-path power and noise can mask Doppler effects in through-wall detection, and conventional PCA's fixed component assumption may fail with broadband noise.
Source: §2 · p. 3, §3.3 · p. 4Method
The method scores PCA components using spectral features to adaptively select motion-relevant components, then uses wavelet scalograms to identify localized energy bursts.
Source: §1 · p. 2, §3.4 · p. 4, §4.1 · p. 5Results
The adaptive method improves performance over conventional PCA and produces clearer, more localized wavelet-domain energy structures across scenarios.
Source: §6 · p. 13, §5.3 · p. 13Takeaways & Limitations
The no-crossing control shows that component relevance varies by scenario: the method can select the first principal component rather than impose fixed motion-component assumptions.
Source: §5.3 · p. 13The reference band depends on the dominant spectral peak of the second principal component, so peak errors may affect all component scores; the fixed component count also limits use of available information.
Source: §6 · p. 13
Abstract
from arXiv · showhide
Through-Wall Detection (TWD) using opportunistic WiFi signals enables non-invasive sensing for security and rescue applications; however many existing approaches rely on controlled access points or specialised hardware. This paper presents a TWD system that extracts Channel State Information (CSI) from ambient WiFi packets using a customised software-defined radio (SDR), Bluebottle, without requiring transmitter control. The key contribution is a spectral-domain scoring mechanism for adaptively selecting motion-relevant principal components from a Principal Component Analysis (PCA) decomposition of the CSI, using Welch power spectral density estimates to quantify each component's signal-to-noise ratio and spectral concentration within the frequency band associated with human motion. The selected components are then analysed using a continuous wavelet transform to robustly identify time-localised motion events. Experimental results demonstrate that the proposed adaptive component-selection method consistently reduces false detections and produces sharper time-frequency energy ridges compared to conventional fixed-component PCA.
1 Introduction
Ubiquitous WiFi signals can support non-invasive detection of human motion or presence behind walls. The paper addresses the challenge of automatically selecting motion-relevant CSI components in opportunistic sensing and proposes adaptive spectral scoring followed by wavelet-based detection.
1 Introduction
- WiFi signal variations can reveal human motion or presence behind walls, with potential applications in non-invasive security and rescue sensing.WiFi signals can penetrate walls, furniture, and glass.Source: §1 · p. 1
- Existing TWD systems may require controlled access points, specialized configurations, or calibration, while fixed-component and explained-variance PCA can miss motion-relevant components.Automatic component identification in opportunistic sensing without transmitter control remains an open challenge.Source: §1 · p. 1
- The proposed framework ranks PCA components by spectral concentration in a motion band and applies wavelet time-frequency analysis to detect transient motion without transmitter control.The paper also reviews TWD and PCA background before presenting its processing pipeline and experiments.Source: §1 · p. 2
2 Problem formulation
The problem formulation represents the wireless channel through frequency- and time-dependent responses sampled as CSI across subcarriers. It explains why motion-induced Doppler can be difficult to detect directly when the direct path dominates or deep fading produces rapid local-frequency variation.
2 Problem formulation
- The received signal is modeled as the transmitted signal multiplied by the channel frequency response, whose inverse Fourier transform gives the channel impulse response.CSI is the sampled channel frequency response at each subcarrier.Source: §2 · p. 2
- The channel model distinguishes static and dynamic paths using their CSI contributions, path collections, amplitudes, and path lengths.Figure 1 illustrates static and dynamic paths.Source: §2 · p. 2, Figure 1 · p. 2
- The Doppler frequency is associated with a human target's velocity and indexed by dynamic path and subcarrier.The formulation also considers a simplified case with one motion-induced dynamic path when expressing CSI amplitude.Source: §2 · pp. 2–3
- Direct-path power can mask Doppler-induced oscillations in CSI amplitude, while deep fading causes rapid local-frequency variations that make direct Doppler estimation difficult.The paper uses the deep-fading effect as a feature for detecting human presence through motion-related spectral structure and temporal variability.Source: §2 · p. 3
3 Conventional detection algorithm
The conventional detection pipeline prepares CSI amplitudes, applies PCA, and uses a continuous wavelet transform and CFAR to identify motion events. Fixed consecutive-component retention can be unreliable when noise and interference disrupt the assumed component ordering.
3.1 Data collection and preparation
- For each frame, CSI data are arranged as an N × (2K+1) matrix across subcarriers; each subcarrier time series is independently filtered with a fourth-order Butterworth filter using a [10, 35] Hz passband.Zero-phase filtering using signal.filtfilt was applied to avoid phase distortion.Source: §3.1 · p. 3, §3.2 · p. 3
- Each frame’s CSI is organized as a matrix with one row per packet sample and one column per subcarrier.The packet count per frame is determined by the SDR sampling rate and collection duration.Source: §3.1 · p. 3
3.3 Principal Component Analysis
- PCA is applied to each frame’s CSI-amplitude matrix after column means are removed; its ordered eigenvalues represent explained variance.Conventional PCA retains a fixed number of consecutive components beginning with the second, although broadband noise and interference can undermine that assumption.Source: §3.3 · pp. 3–4
3.4 Wavelet transform
- The continuous wavelet transform produces a time–frequency scalogram where transient, localized energy bursts appear as bright vertical ridges indicating candidate human-presence events.Log-spaced scales span the frequency range of interest to localize motion-induced events.Source: §3.4 · p. 4
3.5 Detection
- The conventional algorithm aggregates scalogram energy over time and applies CFAR to detect human presence.The time-axis energy vector reduces the two-dimensional scalogram to a form suitable for CFAR.Source: §3.5 · p. 5
4 Adaptive Principal Component Analysis
Fixed consecutive PCA components may not capture motion consistently across environments because motion energy is unevenly distributed and affected by noise. The adaptive method uses Welch PSD estimates to score components by motion-band concentration and robustness, then selects the highest-scoring components.
4.1 Introduction
- Fixed consecutive-component selection can miss motion energy because component energy varies with environment, noise, and multipath conditions.In the example, motion is clear in components 2, 4, 6, 7, and 8, but obscure in 3, 5, 9, and 11.Source: §4.1 · p. 5, Figure 2 · p. 6
4.2 Power Spectral Density analysis for PCA components selection
- Welch PSD estimates quantify each component’s motion relevance, using the second component’s dominant spectral peak to define a frequency window applied across components.The first component is discarded as it typically captures static background structure; the experimental window bandwidth is 7 Hz.Source: §4.2 · pp. 5–6
- The method ranks components by scores indicating spectral localization and robustness to broadband noise, then selects the top five.The score is based on a band ratio and an equivalent SNR; the selected indices are the five highest-scoring components.Source: §4.2 · pp. 6–7
5 Experimental validation
Experiments compare fixed cPCA selection with adaptive aPCA across static, crossing, and separate-office scenarios. Across these conditions, aPCA suppresses noise while selecting motion-associated components, and avoids spurious detections in the no-crossing control.
5.1 Experimental setup and data collection
- The passive Bluebottle receiver collected ambient WiFi packets in two offices; trials varied walking speed and dwell time, with Office 2 presenting greater distance and more complex obstacles.Irregular packet CSI was interpolated and resampled to 100 Hz for analysis.Source: §5.1 · pp. 7–8, Figure 3 · p. 8
5.2.1 No DP crossing (S1)
- In the largely static no-crossing trial, cPCA and aPCA produced qualitatively similar results and neither introduced artificial detections.The component selections and scores are compared in Table 1.Source: §5.2.1 · p. 9, Table 1 · p. 9
5.2.2 Two DP crossings (S1)
- For two DP crossings, cPCA produced blurred energy ridges and several false detections, whereas aPCA yielded clear ridges and correctly identified both crossings without false detections.The detection results are shown in Figure 5.Source: §5.2.2 · p. 9, Figure 5 · p. 10
5.2.3 Four DP crossings (S1)
- For four DP crossings, cPCA showed spurious intermediate ridges and false detections, while aPCA retained four salient event ridges, including one weaker ridge, and correctly identified the crossings.The four aPCA ridges were the only salient scalogram features.Source: §5.2.3 · pp. 10–11, Figure 6 · p. 11
5.2.4 Experiment over separate offices (S2)
- In the separate-office scenario, aPCA offered only a slight improvement because most components were already low-noise, but it increased SNR and produced two sharp peaks around the crossings.The adjacent peaks correspond to the volunteer entering and leaving the direct path and could potentially indicate walking speed.Source: Figure 7 · p. 12, §5.2.4 · p. 12
5.3 Summary of the experiments
- Across scenarios, aPCA consistently selects components associated with motion structure and suppresses noise in crossing trials, while avoiding spurious detections in the no-crossing control.The component selections and scores across scenarios are summarized in Table 5.Source: Table 5 · p. 13, §5.3 · p. 13
- Across motion-present scenarios, the second principal component was selected consistently, while the first was rejected as often dominated by static reflections and broadband noise.For the no-crossing control, aPCA included the first component among its selections.Source: §5.3 · p. 13
6 Discussion
The discussion presents spectral-domain adaptive PCA as a flexible alternative to fixed component selection, reporting clearer wavelet energy structures and fewer false detections. It also identifies reference-band estimation and fixed component count as limitations.
6 Discussion
- The adaptive framework remains flexible in control scenarios and reduces false detections compared with rigid component assumptions.It prioritizes motion-relevant components while suppressing noise-dominated ones.Source: §6 · p. 13
- Spectral-domain features guide selection of principal components, producing clearer and more localized energy structures in the wavelet domain.The second principal component is consistently selected as capturing motion-induced dynamics.Source: §6 · p. 13
- Using the second component’s dominant spectral peak to define the reference band may propagate errors across component scores, while a fixed component count limits use of available information.The authors identify an adaptive threshold for component number as future work.Source: §6 · p. 13