Toward Unobtrusive In-home Gait Analysis Based on Radar Micro-Doppler Signatures

I. INTRODUCTION

• Most earlier radar work discriminated between motion categories, leaving variations within a single motion class comparatively understudied.The paper identifies this within-class task as intra-motion-category classification.Source: §I · p. 1
• Gait monitoring can help assess injury recovery and support early disease diagnosis, while clinical gait studies can be time-consuming, costly, and lack reproducibility.The authors seek contact-less sensing to supplement, not necessarily replace, visual gait examination.Source: §I · p. 1
• The work targets lower-limb kinematic differences and considers two limping patterns plus two cane–leg synchronization styles.The authors position radar as a contact-less supplement or alternative to other sensors, including wearables.Source: §I · p. 2
• 94% of normal walking cases were correctly detected using PCA-based cadence-domain features, which outperformed kinematic modeling across five gait classes.PCA is described as requiring less computation than deep learning approaches and not requiring very large training sets.Source: §I · p. 2
• For four people with diagnosed gait disorders, the proposed method detected gait asymmetry with high sensitivity in three cases.The study used radar data from four individuals with different diagnosed gait disorders.Source: §I · p. 2

II. REPRESENTATIONS OF HUMAN RADAR SIGNATURES

• The radar return from a walking person contains multiple time-varying Doppler-shifted components, whose periodicity follows the gait cycle but is not directly accessible in the time domain.The observed signal is therefore periodically frequency-modulated.Source: §II · p. 2
• A spectrogram reveals local frequency behavior over time, including torso Doppler shifts and micro-Doppler components from swinging arms and legs.The Fourier transform alone does not show local frequency behavior.Source: §II · p. 2

A. Radar Signal Model

• The received signal in the monostatic radar model is a superposition of scatter components with Doppler shifts tied to target-component velocity and angle relative to the radar line of sight.The shifts generally vary over time as the target velocity changes.Source: A. Radar Signal Model · p. 2

C. Time-Frequency Representations

• Radar spectrograms reveal deviations from normal strides, including an impulsive-like spike for a normal stride away from the radar and signatures of knee bending or cane use.Normal strides toward and away from the radar also differ.Source: B. Time-Frequency Representations · p. 3
• The envelope signal tracks maximal Doppler shifts through a gait cycle as pulses over time, while short-time energy sums relevant micro-Doppler bins to represent signature patterns.The envelope can reflect either leg or cane motion; the short-time energy excludes the torso signature.Source: B. Time-Frequency Representations · p. 3
• Fig. 1 contrasts spectrogram examples for walking toward versus away from the radar; color encodes energy in dB.Source: Figure 1 · p. 3

C. Cadence-Velocity Diagram

• The CVD is formed by taking the Fourier transform over time for each Doppler-frequency bin, showing periodic patterns of micro-Doppler components.Velocity is proportional to the observed Doppler frequency, and the CVD does not depend on the gait cycle’s initial phase.Source: C. Cadence-Velocity Diagram · pp. 3–4
• The mean cadence spectrum sums CVD values across Doppler frequencies and indicates how often components occur, independent of their velocities.Its highest peak typically represents stride rate; the paper distinguishes cadence frequency from stride rate for pathological or assisted gait.Source: C. Cadence-Velocity Diagram · p. 4
• The mean Doppler spectrum sums across cadence frequencies and can provide features such as average walking speed and minimum or maximum Doppler shifts.Source: C. Cadence-Velocity Diagram · p. 4

A. Experimental Data and Setup

• Ten volunteers supplied 1,000 office measurements across five walking styles using 24-GHz continuous-wave radar, walking slowly without arm swing toward and away from the radar.The simulated limps involved one or both knees that could not bend.Source: A. Experimental Data and Setup · pp. 4–5
• A separate four-person dataset of participants with diagnosed gait disorders added 87 measurements, including 16 synchronized-cane walks for person C.Source: A. Experimental Data and Setup · p. 5

B. Physical Features based on Sum-of-Harmonics Analysis

• Cane use raises micro-Doppler repetition frequency, especially out of sync, while two-leg limping yields the lowest maximal Doppler shift and one-leg limping motivates a variation feature.The coefficient of variation of maximal Doppler shifts is expected to be particularly high for one-leg limping.Source: B. Physical Features based on… · p. 5
• The SOH algorithm refines an initial FF estimate fmD, then evaluates candidate FFs 1/3, 1/2, or 1 times fmD (if the initial estimate is correct) to limit computation and avoid overfitting.Using linear least squares, it computes harmonic amplitudes and phases in closed form as functions of f0 and q.Source: B. Physical Features based on… · p. 6
• The physical feature vector combines repetition frequency, maximal Doppler shift, shift variability, harmonic ratio, and harmonic amplitudes, omitting base velocity as nondiscriminative for these motions.The maximal harmonic order is set to 5.Source: B. Physical Features based on… · pp. 5–6
• 70%, 74%, and 90% of measurements showed the expected gait harmonic frequency ratios 1, 1/2, and 1/3, respectively.Table I presents confusion matrices for β-based classification of three gait patterns.Source: B. Physical Features based on… · p. 6, Table I · p. 7

C. Subspace Features

• PCA learns intrinsic features of walking styles that need not correspond one-to-one with human motion kinematics.The data matrix is decomposed using SVD; its left eigenvectors with the largest eigenvalues define the leading principal components.Source: C. Subspace Features · p. 7
• Each signal is vectorized and stacked as a column of the data matrix used to derive the principal components.The matrix has one column per training sample.Source: C. Subspace Features · p. 7
• The resulting projections onto the first λ eigenvectors form the PCA feature vector used for classification.The feature vector contains λ projected values.Source: C. Subspace Features · p. 7

D. Radar Data Representations

• The spectrogram is formed from 6-second measurements and reduced to 101 × 192 pixels after time subsampling and image binning.It uses a Hamming window of about 0.1 seconds, 2048 frequency points, and Doppler components below 500 Hz.Source: D. Radar Data Representations · p. 7
• The paper lists the radar data representations and their dimensions in Table II.The time-domain alternative removes lower Doppler components attributed to torso motion before taking the Fourier transform.Source: Table II · p. 8, D. Radar Data Representations · p. 8
• For toward/away radar-motion measurements, CVDs include cadences up to 5 Hz and Doppler ranges 0 to +500 Hz/0 to −500 Hz, then are downsampled in Doppler to 101 × 129 pixels and normalized to [0,1].Pre-processing warps CVDs along the cadence and Doppler axes to compensate for differing stride rates and maximal Doppler shifts; pre-processed CVDs assume fmD = 1 Hz and fDmax = 500 Hz.Source: D. Radar Data Representations · p. 8

E. Methodology and Parameter Optimization

• The evaluation uses a nearest-neighbor classifier and 10-fold stratified cross-validation, and also assesses performance with leave-one-subject-out cross-validation.Accuracy, false positive rate, false negative rate, and true positive rate are used as performance measures.Source: E. Methodology and Parameter… · p. 8
• The pre-processed CVD gives the highest classification rates among the considered representations, while accuracy does not significantly increase beyond 20 principal components.The spectrogram performs worse than the other representations; the pre-processed CVD is selected for the subsequent comparison.Source: E. Methodology and Parameter… · p. 8, Figure 4 · p. 8
• With pre-processed CVDs, the highest accuracy is achieved by the nearest-neighbor classifier (κ = 1), and using more than 22 principal components does not significantly increase accuracy.Larger values of κ or λ increase computation time; Fig. 5 shows average correct classification rate across the parameter choices.Source: E. Methodology and Parameter… · p. 8, Figure 5 · p. 9

A. Physical Features

• 86.6% overall accuracy was achieved with the physical feature vector, with an 18.0% false positive rate and a 4.4% false negative rate.Class accuracies were 82.0% for NW, 88.0% for L1, 95.0% for L2, 80.0% for CW, and 88.0% for CW/oos; Table III presents the confusion matrices.Source: A. Physical Features · p. 9, Table III · p. 9
• 79.4% overall correct classification was achieved with Björklund et al.’s zB1 features, falling to 40.4% when velocity profiles were removed to form zB2.The results indicate that cadence frequencies f1, f2, and f3 together with base velocity v0 are not key for discriminating these gait classes.Source: A. Physical Features · p. 9
• 60.7% overall accuracy with zR1 and 75.0% with zR2 were reported for Ricci and Balleri’s feature sets.The corresponding FPR/FNR values were 56.6%/13.6% for zR1 and 21.5%/8.5% for zR2.Source: A. Physical Features · p. 9
• Base velocity and micro-Doppler repetition frequency are not suited to distinguish the considered gait classes, whereas CVD-derived signals contain discriminative characteristics.The cited comparisons also show lower classification rates when velocity profiles or selected cadence and velocity features are used.Source: A. Physical Features · p. 9

• 93.8% overall accuracy, with FPR 6.5% and FNR 2.1%, was achieved using PCA-based CVD features with the NN classifier.Source: B. Subspace Features · p. 9
• The CW class had an 88.5% classification rate, with errors arising from confusion with normal walking in both directions.Source: B. Subspace Features · p. 10
• 80.4 ± 4.9% accuracy was achieved in leave-one-subject-out cross-validation, with FPR 23.5 ± 15.5% and FNR 6.6 ± 3.4%.Most confusion occurred between normal and cane-assisted walks.Source: B. Subspace Features · p. 10

C. Discussion

• 93.8% correct classification was achieved by PCA-based CVD subspace features, which outperformed physical features and kept FPR at 6.5% and FNR at 2.1%.The method was reported to work across gait classes despite 1,000 measurements from ten test subjects.Source: C. Discussion · p. 10
• Spectrograms indicated asymmetry through alternating lower maximal Doppler shifts for Persons A, B, and D, and higher energy at alternating stride onsets for Person C.Person C’s cane signature overlapped with every other stride signature.Source: C. Discussion · p. 10
• Abnormal gait was correctly identified in 92% (12/13), 100% (20/20), 75% (9/12), and 100% (26/26) for Persons A–D, respectively; Person C’s cane was detected in 81% (13/16).The classifier was trained on ten healthy individuals and evaluated on four people with pathological gait; each measurement lasted six seconds.Source: C. Discussion · pp. 10–11

V. CONCLUSION

• The conclusion describes recognition across five simulated gait classes and recommends future studies with a wider group of patients with pathological gait.Source: §V · p. 11