PaperlayerView exact revision on arXiv

Source-linked AI summary

WiSpeed: A Statistical Electromagnetic Approach for Device-Free Indoor Speed Estimation

Feng Zhang, Chen Chen, Beibei Wang, K. J. Ray Liu

arXiv:1712.00348v1eess.SP

TL;DR

Indoor device-free speed estimation is difficult under severe multipath and blocked direct paths. WiSpeed uses radio-signal statistics to estimate motion speed in device-free or device-based settings, with reported walking-speed accuracy and fall-detection results.

  • Problem

    Indoor device-free speed estimation is challenging because severe multipath and blocked paths impede measurement, while existing approaches have constraints such as specialized hardware or strong LOS requirements.

    Source: §II · p. 3

  • Method

    WiSpeed uses electromagnetic-wave statistics to relate received-field-power autocorrelation to motion speed, estimating it from the first local peak of the autocorrelation differential.

    Source: §I · p. 2

  • Results

    4.85% MAPE for device-free human walking speed monitoring and 4.62% MAPE for device-based speed estimation; fall detection averaged 95% with no false alarms.

    Source: §I · p. 2, §VIII · p. 14

  • Takeaways & Limitations

    WiSpeed also extracts stride length and step counts from speed patterns and is presented as a low-cost option for widely deployed indoor monitoring.

    Source: §I · p. 2

Abstract

from arXiv · show

Due to the severe multipath effect, no satisfactory device-free methods have ever been found for indoor speed estimation problem, especially in non-line-of-sight scenarios, where the direct path between the source and observer is blocked. In this paper, we present WiSpeed, a universal low-complexity indoor speed estimation system leveraging radio signals, such as commercial WiFi, LTE, 5G, etc., which can work in both device-free and device-based situations. By exploiting the statistical theory of electromagnetic waves, we establish a link between the autocorrelation function of the physical layer channel state information and the speed of a moving object, which lays the foundation of WiSpeed. WiSpeed differs from the other schemes requiring strong line-of-sight conditions between the source and observer in that it embraces the rich-scattering environment typical for indoors to facilitate highly accurate speed estimation. Moreover, as a calibration-free system, WiSpeed saves the users' efforts from large-scale training and fine-tuning of system parameters. In addition, WiSpeed could extract the stride length as well as detect abnormal activities such as falling down, a major threat to seniors that leads to a large number of fatalities every year. Extensive experiments show that WiSpeed achieves a mean absolute percentage error of 4.85% for device-free human walking speed estimation and 4.62% for device-based speed estimation, and a detection rate of 95% without false alarms for fall detection.

I. INTRODUCTION

WiSpeed estimates indoor motion speed from radio-signal statistics in device-free and device-based settings. It is designed for rich-scattering environments and addresses limitations of approaches that require direct paths, line of sight, or specialized hardware.

I. INTRODUCTION

  • Device-free speed estimation is challenging under severe indoor multipath and blocked paths, while earlier WiFi approaches require specialized hardware or strong line of sight.The cited WiGait and WiDar examples require, respectively, specialized FMCW hardware with 1.69 GHz bandwidth and strong LOS with dense WiFi deployment.Source: §I · p. 1
  • WiSpeed relates the autocorrelation of received electromagnetic-field power to motion speed and identifies the first local peak of its differential to estimate speed.The paper also describes step counts, stride length, and fall detection as outputs derived from speed patterns.Source: §I · pp. 1–2
  • 4.85% MAPE for device-free human walking speed monitoring and 4.62% MAPE for device-based speed estimation; fall detection averaged 95% with no false alarms.The experiments also report stride-length extraction and step-count estimation under the device-free setting.Source: §I · p. 2
  • The introduction presents indoor fitness tracking, navigation, fall detection, and home surveillance as potential applications of speed and step-pattern monitoring.It notes that WiFi infrastructure is readily available in most indoor spaces and frames WiSpeed as low-cost and widely deployable.Source: §I · p. 2

II. RELATED WORKS

  • Learning-based methods classify activities using speed-related features rather than measuring speed directly, and those features may vary with environment, subject, or device placement.Their offline-to-online feature reproducibility assumption may therefore be violated.Source: §II · p. 3
  • Ray-tracing methods either isolate human-reflected paths using fine temporal or spatial resolution, or attenuate multipath by placing devices near the subject; the latter require strong LOS.The path-isolation approach can require dedicated hardware unavailable on WiFi devices.Source: §II · p. 3
  • WiSpeed embraces indoor multipath and is described as working under severe NLOS conditions without training or calibration across environments and human subjects.Its stated basis is physical electromagnetic-wave features associated with motion speed.Source: §II · p. 3
  • A single pair of commercial WiFi devices makes WiSpeed a low-cost deployment, unlike cited approaches requiring specialized hardware or multiple device pairs.The paper also contrasts its computation with approaches using PCA, DWT, or STFT.Source: §II · pp. 3–4

III. STATISTICAL THEORY OF EM WAVES FOR WIRELESS MOTION SENSING

The section models indoor propagation as a sum of fields from static and moving scatterers, then uses statistical assumptions about rich scattering to derive how motion appears in the received-field autocorrelation function (ACF). The ACF combines moving-scatterer contributions weighted by their radiation power, with speed information embedded in it.

A. Decomposition of the Received Electric Field

  • In a rich-scattering model, each scatterer contributes to the received field, which is decomposed into static- and moving-scatterer components.A scatterer is treated as a virtual antenna diffusing waves in all directions; a moving transmitter is classified as a dynamic scatterer.Source: A. Decomposition of the Received… · p. 4

B. Statistical Behaviors of the Received Electric Field

  • The analysis treats each scatterer’s field as a superposition of plane waves integrated over arrival directions, using channel reciprocity to analyze the received field’s components.Each plane wave is described by its direction angle, vector wavenumber, and angular spectrum.Source: B. Statistical Behaviors of the… · p. 4, Figure 2 · p. 5
  • The statistical model assumes many plane waves with uniformly distributed arrival directions, polarizations, and phases, and uncorrelated angular components and distinct dynamic scatterers.The paper motivates these assumptions through random ray phases, differing multiple-scattering paths, and spatial decorrelation at separations of at least half a wavelength.Source: B. Statistical Behaviors of the… · p. 5
  • The received-field ACF combines moving-scatterer ACFs weighted by radiation power, and contains their speed information.The moving direction of each dynamic scatterer does not play a role in the ACF.Source: B. Statistical Behaviors of the… · p. 6

IV. THEORETICAL FOUNDATION OF WISPEED

WiSpeed links speed-dependent electromagnetic-wave autocorrelation to measurable CSI power response, then estimates speed from a feature of its sample ACF. The section also tests the model in blocked-LOS office scenarios involving a moving transmitter or a passing person.

Speed-dependent ACF

  • When dynamic scatterers share speed v, their electromagnetic-field ACF simplifies to a speed-dependent form, linking autocorrelation behavior to motion.The section treats received electric-field power as equivalent to measurable channel-response power and uses CSI power-response ACF for estimation.Source: §IV · p. 6

Noise model

  • Static-environment CSI measurements support modeling measurement noise as additive white Gaussian noise.The Q-Q plot is close to a normal distribution, and the sample ACF taps lie within ±2σ.Source: §IV · p. 6

Theoretical ACF

  • Theoretical spatial ACF magnitudes for all orthogonal components decay with oscillations as distance d increases.The section defines d as viτ.Source: §IV · p. 7

Speed feature

  • Speed is embedded in the CSI-estimated power-response ACF; WiSpeed localizes its first derivative peak to identify the speed-bearing component despite superposition.The first local peak of the squared-component derivative coincides with the aggregate ACF derivative's first local peak.Source: §IV · p. 8

Experimental validation

  • In blocked-LOS tests, the averaged ACF matched theory for a moving transmitter, while a passing person produced a less pronounced squared-component pattern.The moving transmitter dominated the other scatterers; the person's moving body parts had lower radiation power.Source: §IV · pp. 8–9

CSI processing

  • Averaging sample ACFs across subcarriers produces a less noisy estimate than using individual subcarrier ACFs.The observed averaged ACF in the moving-transmitter case matches the theoretical analysis.Source: §IV · p. 8

V. KEY COMPONENTS OF WISPEED

WiSpeed combines moving-speed, acceleration, and gait-cycle estimators; the latter two extract features from speed estimates for fall detection and gait-cycle estimation.

V. KEY COMPONENTS OF WISPEED

  • WiSpeed integrates moving-speed, acceleration, and gait-cycle estimators, with the latter two extracting features from speed estimates for fall detection and gait-cycle estimation.Source: §V · p. 9

A. Moving Speed Estimator

WiSpeed estimates speed from the first local peak in a CSI-derived sample ACF, using robust local peak detection to handle noise.

A. Moving Speed Estimator

  • WiSpeed maps the first local peak of the sample ACF to speed and median-filters estimates to remove outliers.The peak location is mapped using the distance 0.54λ and the estimated peak lag.Source: A. Moving Speed Estimator · pp. 9–10
  • A moving window tests for peaks through linear and quadratic fits, retaining windows whose regression-error ratio exceeds a threshold and neighboring values.The threshold is set by the desired probability of finding a false peak; a lone peak in a sufficiently small window is located from the fitted quadratic curve.Source: A. Moving Speed Estimator · pp. 9–10
  • Without noise, peaks at 0.331/0.760 s are estimated as 0.327/0.763 s; with σ=0.2, estimates are 0.336/0.762 s.The example uses a signal sampled at 100 Hz.Source: Figure 7 · p. 10, A. Moving Speed Estimator · p. 10

B. Acceleration Estimator

WiSpeed estimates acceleration by smoothing speed estimates with an ℓ1 trend filter, then differentiating the smoothed trend.

B. Acceleration Estimator

  • Directly differencing adjacent speed estimates can magnify noise, so WiSpeed instead models acceleration as piecewise linear over short durations with enough speed estimates.The ℓ1 trend filter extracts the piecewise linear trend before acceleration is estimated by differentiation.Source: B. Acceleration Estimator · p. 10
  • The ℓ1 trend filter produces a piecewise-linear smoothed speed trend, from which acceleration is estimated by differentiation.The method uses the filter because it produces smooth, piecewise-linear trend estimates.Source: B. Acceleration Estimator · p. 10
  • The smoothing objective trades off residual size against smoothness, and the ℓ1 filter's complexity grows linearly with data length.The regularization parameter controls this trade-off; the filter can be calculated in real time on most platforms.Source: B. Acceleration Estimator · p. 10

C. Gait Cycle Estimator

WiSpeed estimates gait cycles from periodic changes in speed and uses adjacent peak intervals and distances to estimate cycle time and stride length.

C. Gait Cycle Estimator

  • WiSpeed begins gait-cycle estimation when estimated speed is in a selected range, such as 1–2 m/s, and acceleration estimates are small.The cited example range is 1 m/s to 2 m/s.Source: C. Gait Cycle Estimator · p. 10
  • A walking step is decomposed into lifting a leg, contacting the ground and pushing forward, then briefly keeping still before the next step.Source: C. Gait Cycle Estimator · p. 10
  • WiSpeed treats each walking cycle as acceleration followed by deceleration, uses local speed maxima for gait-cycle timing, and estimates stride length from peak-to-peak distance.Adjacent peak intervals estimate gait-cycle duration; peak locations are identified with the persistence-based scheme.Source: C. Gait Cycle Estimator · pp. 10–11

VI. EXPERIMENTAL RESULTS

The experiments introduce the indoor environment and system setups, then evaluate WiSpeed in human walking monitoring and fall detection.

VI. EXPERIMENTAL RESULTS

  • WiSpeed is evaluated in two applications: human walking monitoring and human fall detection.The experiments first introduce the indoor environment and system setups.Source: §VI · p. 11

A. Environment

WiSpeed was evaluated in a typical office across device-free and device-based walking scenarios and a fall-detection task. Walking estimates were accurate across varied conditions, while fall detection separated falls from other activities with one outlier.

A. Environment

  • The experiments used a typical office environment and tested walking-distance estimation against ground truth, plus activities including falling, walking, and everyday movements.The device-free walking experiments varied routes and device locations; device-based settings were also tested.Source: A. Environment · p. 11, B. Experimental Settings · p. 11, Table I · p. 11

C. Human Walking Monitoring

  • In one static-Tx/Rx walking trial, estimated distance was 8.46 m versus an 8 m ground truth, and average stride length was 0.7 m.The first local peak of the estimated ACF difference remained consistent across different ACFs calculated at similar walking speeds.Source: C. Human Walking Monitoring · p. 11
  • Across 200 walking experiments, estimates were consistent across device locations, routes, subjects, and speeds, but device-free distance estimates tended to be high.The stated bias arises because route distance baselines omit displacement in the gravity direction, whereas WiSpeed measures absolute distance in its coverage area.Source: C. Human Walking Monitoring · p. 12
  • MAPE was 4.85% for device-free human walking speed estimation and 4.62% for device-based speed estimation.The paper reports consistent performance across Tx/Rx locations, routes, subjects, and walking speeds, including severe NLOS conditions.Source: C. Human Walking Monitoring · p. 12

D. Human Fall Detection

VII. DISCUSSION

WiSpeed’s tracking window averages motion over time, so rapidly changing speeds can reduce performance; higher sampling rates or carrier frequencies can shorten the window. The conclusion reports walking MAPE values of 4.85% and 4.62%, plus 95% fall detection with no false alarms.

A. Tracking a Fast Moving Object

  • WiSpeed captures average speed over a window of duration T0, so substantial speed changes within that interval can degrade performance.A smaller T0 is desirable for fast-varying motion and can be achieved by increasing channel sampling rate or carrier frequency.Source: A. Tracking a Fast Moving Object · p. 13

B. Computational Complexity

  • For walking and standing-up motions, a lower channel sampling rate suffices and can reduce computational complexity.At Fs = 1500 Hz, fc = 5.805 GHz, F = 180, and M = 100, one output required around 3 million multiplications and took 80.4 ms on the reported desktop.Source: B. Computational Complexity · p. 13

VIII. CONCLUSIONS

  • WiSpeed achieved MAPE values of 4.85% for device-free human walking speed monitoring and 4.62% for device-based speed estimation, with a 95% fall-detection rate and no false alarms.The conclusion describes the system as low-cost and low-complexity, and a candidate for indoor passive activity monitoring.Source: §VIII · p. 14

A. Derivation of (10)

The derivation rewrites the channel response and develops covariance expressions for received-field components and channel power. It invokes Gaussian-variable expectations and stated assumptions to obtain the result in equation (10).

A. Derivation of (10)

  • The derivation rewrites G(t, f) and expresses its covariance using a term for the received-field component covariance.The passage then focuses on γGu(τ, f) under stated assumptions.Source: A. Derivation of (10) · p. 14
  • After defining real and imaginary field components, the derivation applies assumptions about their statistical behavior and obtains the result in equation (10).The intermediate steps derive and combine additional terms before stating the result.Source: A. Derivation of (10) · pp. 15–16
  • The derivation proceeds by expressing the covariance between Eiu(t, f) and |Eiu(t −τ, f)|2, then evaluating terms using zero-mean Gaussian expectations.It cites the Gaussian expectation relationship.Source: A. Derivation of (10) · p. 15

Related papers

Loading 1712.00348v1…