PaperlayerView exact revision on arXiv

Source-linked AI summary

FreeSense:Indoor Human Identification with WiFi Signals

Tong Xin, Bin Guo, Zhu Wang, Mingyang Li, Zhiwen Yu

arXiv:1608.03430v1cs.HC

TL;DR

Existing identification methods can involve privacy or sensing-coverage constraints, motivating an indoor approach using WiFi CSI. FreeSense uses CSI waveform features for identification and reports 88.9% to 94.5% accuracy as the candidate set changes from 6 to 2.

  • Problem

    Existing identification methods may require close sensor proximity, line-of-sight coverage, rich lighting, or wearable sensors, and vision-based approaches raise privacy concerns.

    Source: §I · p. 1

  • Method

    FreeSense identifies people indoors from person-specific effects on WiFi CSI, using PCA and DWT to extract LOS waveform shape features.

    Source: §I · pp. 1–2

  • Results

    88.9% to 94.5% identification accuracy was reported as the candidate user set changed from 6 to 2.

    Source: §I · p. 2

  • Takeaways & Limitations

    The reported results support the method’s effectiveness in domestic environments.

    Source: §VI · p. 6

Abstract

from arXiv · show

Human identification plays an important role in human-computer interaction. There have been numerous methods proposed for human identification (e.g., face recognition, gait recognition, fingerprint identification, etc.). While these methods could be very useful under different conditions, they also suffer from certain shortcomings (e.g., user privacy, sensing coverage range). In this paper, we propose a novel approach for human identification, which leverages WIFI signals to enable non-intrusive human identification in domestic environments. It is based on the observation that each person has specific influence patterns to the surrounding WIFI signal while moving indoors, regarding their body shape characteristics and motion patterns. The influence can be captured by the Channel State Information (CSI) time series of WIFI. Specifically, a combination of Principal Component Analysis (PCA), Discrete Wavelet Transform (DWT) and Dynamic Time Warping (DTW) techniques is used for CSI waveform-based human identification. We implemented the system in a 6m*5m smart home environment and recruited 9 users for data collection and evaluation. Experimental results indicate that the identification accuracy is about 88.9% to 94.5% when the candidate user set changes from 6 to 2, showing that the proposed human identification method is effective in domestic environments.

I. INTRODUCTION

FreeSense proposes indoor identification from person-specific effects on WiFi CSI, addressing constraints in existing identification methods. It focuses on LOS-crossing waveforms and combines PCA and DWT for feature extraction.

I. INTRODUCTION

  • Existing approaches can require close sensor proximity, line-of-sight and rich lighting, or wearable sensors; vision methods also raise privacy concerns.The paper motivates WiFi sensing as a non-intrusive, privacy-preserving alternative for domestic environments.Source: §I · pp. 1–2
  • Person-specific body shapes and motion patterns are proposed to leave distinctive effects in WiFi CSI time series during indoor movement.The paper targets home environments, where recognizing family members could support personalized services.Source: §I · p. 1
  • FreeSense treats LOS-path crossings as a functional window because arbitrary walking paths can produce less distinguishable signals for identification.The proposed segmentation algorithm extracts LOS waveforms from CSI time series using their characteristic changes in rates of change.Source: §I · pp. 1–2
  • Ordinary rate-of-change, signal-energy, and peak-power features are described as ineffective because their values are similar across people.The paper identifies extracting useful CSI features as a challenge for exact identification.Source: §I · pp. 1–2
  • The proposed feature model applies PCA to CSI time series and DWT to compress LOS waveforms and extract their shape features.Recognition is based on differences in how people’s movement influences WiFi signals.Source: §I · p. 2
  • 88.9% to 94.5% identification accuracy was reported as the candidate user set changed from 6 to 2.The prototype was evaluated in a 6m*5m smart home with 9 users.Source: §I · p. 2

II. RELATED WORK

Related work spans physiological and behavioral biometrics, alongside CSI-based motion sensing. The paper distinguishes its focus on identifying people from CSI studies of activities.

A. HUMAN IDENTIFICATION

  • Fingerprint, iris, vein, hand-shape [1], and face recognition are physiological identification approaches, but they require proximity to the sensor for accuracy.The passage groups these methods as physiological-feature based.Source: A. HUMAN IDENTIFICATION · p. 2
  • Camera-based gait recognition needs line of sight and sufficient lighting and raises privacy issues, while accelerometer-based gait recognition [7] requires users to wear sensors.Both examples are behavioral biometrics focused on gait.Source: A. HUMAN IDENTIFICATION · p. 2

B. CSI-BASED MOTION DETECTION

The paper frames human identification as a new application of CSI sensing and develops a pipeline around LOS-crossing waveforms. It selects, segments, compresses, and compares those waveforms to account for path interference and variation in length.

B. CSI-BASED MOTION DETECTION

  • Existing CSI motion studies cover activity recognition, while this work focuses on human identification.Examples include fall detection [15], household activity recognition, gestures, mouth shapes [17], and keystrokes.Source: B. CSI-BASED MOTION DETECTION · p. 2

A. PROBLEM ANALYSIS

  • LOS crossings are selected as the functional window because WiFi signals fluctuate more notably there than on NLOS paths.The paper calls the waveform produced when a person walks across the LOS path the LOS waveform.Source: A. PROBLEM ANALYSIS · p. 2, Figure 1 · p. 2
  • The segmentation method uses the typical increasing and decreasing trend in CSI change rates and adds qualifications to reduce interference errors from walking on other paths.It differs from the cited WiKey MAD approach by adopting a sliding-window search with additional qualifications.Source: A. PROBLEM ANALYSIS · pp. 2–3
  • LOS waveform shapes are used as features to retain time- and frequency-domain information, with PCA reducing CSI dimensionality and DWT compressing the waveform.The input comprises 30 * MTx * MRx CSI streams, and the passage says direct use would have high computational cost.Source: A. PROBLEM ANALYSIS · p. 3
  • Dynamic Time Warping compares waveforms of different lengths when extracted start and end times are imprecise and midpoints rarely align.The method finds a minimum-distance alignment between the waveforms.Source: A. PROBLEM ANALYSIS · p. 3

B. SYSTEM FRAMEWORK

The FreeSense framework comprises data collection and noise removal, feature extraction, and human identification. The system records CSI during LOS crossings, extracts LOS waveform shape features, and classifies identities.

B. SYSTEM FRAMEWORK

  • The framework has three parts: data collection and noise removal, feature extraction, and human identification.The framework is summarized in Figure 2.Source: B. SYSTEM FRAMEWORK · p. 3, Figure 2 · p. 3
  • CSI is recorded with COTS WiFi network interface cards while people walk across the LOS path, then filtered to remove noise.Source: B. SYSTEM FRAMEWORK · p. 3
  • The system extracts LOS waveform shape features and uses a k-nearest neighbor classifier for human identification.Source: B. SYSTEM FRAMEWORK · p. 3

IV. DETAILED DESIGN OF THE SYSTEM

The system filters noisy commodity-WiFi CSI with a Butterworth IIR filter before later processing, targeting the frequency range associated with human walking.

A. NOISE REMOVING

  • A Butterworth IIR filter removes inherent noise in commodity-WiFi CSI while preserving phase information and providing a maximally flat passband response.The reported human-walking variations are around 10 Hz, with CSI sampled at 1000 Hz; the filter cutoff is calculated from these values.Source: A. NOISE REMOVING · p. 3
  • Figure 3 shows the original and filtered CSI time series, with the filtering result displayed in panel (b).The supplied text does not report a quantitative comparison between the traces.Source: Figure 3 · p. 3, A. NOISE REMOVING · p. 3

B. FEATURE EXTRACTION

Feature extraction uses PCA to reduce CSI dimensionality, segments line-of-sight waveforms, and compresses each waveform with DWT while retaining time- and frequency-domain information.

B. FEATURE EXTRACTION

  • The feature pipeline applies PCA to each antenna pair, segments the CSI series, and then compresses the extracted waveforms with DWT.Source: B. FEATURE EXTRACTION · p. 3
  • PCA reduces each TX-RX antenna pair to four principal components, which are ordered by descending peak-to-peak value to reduce reordering across walks.The paper reports that the top four components contain significant CSI variation caused by human walking.Source: B. FEATURE EXTRACTION · pp. 3–4
  • The segmentation algorithm locates line-of-sight waveform boundaries using the characteristic increasing and decreasing rates of change in CSI.It computes mean absolute deviations in windows before and after points, then sums MAD across principal components and applies thresholds T1 and T2, with T1>T2.Source: B. FEATURE EXTRACTION · p. 4
  • Boundary rules record an end point only after a start point and constrain the interval using empirically determined timelen1 and timelen2 thresholds.The resulting start-end point pair defines the line-of-sight waveform.Source: B. FEATURE EXTRACTION · p. 4
  • DWT compresses each line-of-sight waveform using Daubechies D4 approximation coefficients as shape features.The authors state that this preserves most time- and frequency-domain information.Source: B. FEATURE EXTRACTION · p. 4

C. HUMAN IDENTIFICATION

Human identification uses a KNN classifier that compares line-of-sight waveform shape features with DTW distance and assigns the majority label among the nearest neighbors.

C. HUMAN IDENTIFICATION

  • The KNN classifier assigns the majority class among the k nearest waveform-shape feature neighbors ranked by DTW distance.The distance is computed between two line-of-sight waveform series.Source: C. HUMAN IDENTIFICATION · p. 4

V. EVALUATION AND DISCUSSIONS

The evaluation section includes a figure labeled as the experimental hardware setup.

V. EVALUATION AND DISCUSSIONS

  • Figure 5 is identified as the experimental scene showing the hardware set.Source: Figure 5 · p. 4

A. IMPLEMENTATION

The prototype used a six-antenna-pair WiFi setup in a home environment and collected data from nine volunteers under a prescribed walking protocol. For LOS-waveform segmentation, both algorithms had high detection ratios, while FreeSense had a lower error ratio.

A. IMPLEMENTATION

  • The prototype used a Lenovo X200 receiver with an Intel Link 5300 WiFi NIC and a TP-Link router operating at 2.4 GHz.The transmitter had two antennas and the receiver had three; they were placed 250 cm apart in the entry hall, with CSI measured from ICMP pings at about 1,000 packets per second.Source: A. IMPLEMENTATION · pp. 4–5
  • Nine university-student volunteers aged 20–26 each provided 40 samples, with walking direction and foot placement constrained for comparability.Participants crossed the LOS path from left to right, placing the left foot in area A and the right foot in area B.Source: A. IMPLEMENTATION · p. 5

B. LINE-OF-SIGHT WAVEFORM DETECTION ACCURACY

  • Both segmentation algorithms achieved high detection ratios, while FreeSense had a lower error ratio.The comparison used Detection Ratio and Error Ratio metrics, with WiKey as the baseline.Source: B. LINE-OF-SIGHT WAVEFORM DETECTION… · p. 5

C. CLASSIFICATION ACCURACY

Identification accuracy declined as the candidate set grew from 2 to 9 people, while a prescribed round trip and larger training sets were also evaluated. The reported accuracy was nearly 81.5% on average for six people in the round-trip condition, and rose from 75.0% to 91.7% as training samples increased from 10 to 30.

Accuracy for different number of human

  • 75.5% to 94.5% accuracy was achieved as the candidate set varied from 2 to 9 people, with accuracy declining as more users were included.For candidate sets of 2 to 6, accuracy reached up to 88.9%.Source: C. CLASSIFICATION ACCURACY · p. 5, Figure 6 · p. 5

Performance for more walking conditions

  • Nearly 81.5% average accuracy was achieved for six people walking a prescribed round trip across the LOS path in both directions.Participants followed specified foot-placement conditions for the two crossing directions.Source: C. CLASSIFICATION ACCURACY · p. 5, Figure 7 · p. 5

Effect of the size of training sets

D. DISCUSSIONS

FreeSense reports identification accuracy from 94.5% to 88.9% as the candidate set grows from 2 to 6, while the discussion notes factors that may affect accuracy and limits current testing mainly to one-person environments.

D. DISCUSSIONS

  • Accuracy may be affected by subject-to-transmitter distance and LOS-path crossing manner, whose detailed impacts remain for future study.The authors characterize the experiments as restricted and report more than 88% accuracy with 6 subjects as family members.Source: D. DISCUSSIONS · p. 6
  • Current tests are mainly in one-person environments; the influence of surrounding people and performance with multiple people remain to be tested.The authors also plan to study FreeSense's effective bounds and improve identification performance.Source: D. DISCUSSIONS · p. 6

VI. CONCLUSION

  • 94.5% to 88.9% identification accuracy as the candidate user set changes from 2 to 6 in domestic environments.The authors describe FreeSense as an early effort and plan to measure the impact of different variables and improve performance.Source: §VI · p. 6
Loading 1608.03430v1…