A cross-domain adaptive multi-person gait identity authentication method based on commercial WiFi equipment

Abstract

The application discloses a cross-domain adaptive multi-person gait identity authentication method based on commercial WiFi equipment. The method first collects original channel state information (CSI) signals by using the WiFi equipment; sequentially carries out band-pass filtering, wavelet denoising and conjugate multiplication preprocessing on the CSI signals; carries out blind source separation on the mixed CSI signals of multiple persons by using a reconstructed independent component analysis (RICA) algorithm, extracts independent gait components belonging to target individuals, and converts the independent gait components into subcarrier frequency diagrams; inputs the subcarrier frequency diagrams into a hybrid deep learning network with a CNN feature extractor and a Transformer encoder as the backbone, extracts gait features, and deploys a weakly supervised adaptive training strategy on the upper part of the backbone network to complete cross-domain gait identity authentication. The application effectively solves the technical problems of low accuracy and poor environmental migration of existing WiFi gait recognition under the interference of multiple persons, and has the characteristics of non-sensing, privacy protection and strong robustness.

Description

Technical Field

[0001] This invention relates to the field of wireless sensing and biometrics, specifically to a technique for gait identity authentication in multi-person scenarios using channel state information. Background Technology

[0002] Modern ubiquitous computing systems deployed in homes, offices, and public buildings require continuous and unobtrusive authentication. However, in practice, unauthorized individuals can still follow people into office hallways, eavesdrop on confidential meetings, or enter living rooms when the homeowner is not home. Traditional authentication mechanisms, such as passwords, badges, or facial recognition unlocking, only function at specific checkpoints and are easily forgotten, shared, or forged. This necessitates an authentication signal that can be passively captured during people's daily activities.

[0003] Gait is a unique behavioral biometric that reflects a person's body shape, posture, and walking style. Unlike fingerprints or faces, gait can be sensed from a distance as people move naturally. This makes it well-suited for continuous or background authentication. Vision-based gait recognition technology has made significant progress with the help of deep learning. However, camera systems still require direct line of sight, perform poorly in low-light conditions, and raise privacy concerns in homes and workplaces. These limitations have prompted the search for non-visual, device-free gait sensing methods as a more practical, granular authentication solution for everyday use. WiFi Channel State Information (CSI) provides an attractive basis for device-free gait authentication.

[0004] Many commercial WiFi devices support Gait Intuition (CSI), which captures fine-grained multipath propagation information highly sensitive to human movement. Previous research has shown that CSI can support activity recognition, respiratory monitoring, and indoor positioning. Some systems further utilize gait-induced CSI variations for user identification or authentication. However, most of these systems are built and evaluated in controlled rooms, targeting a single walker along a predefined path, using a fixed transceiver deployment. Under these conditions, deep models trained on CSI time-frequency features can achieve high accuracy. However, these models cannot tell us how effective gait authentication is when the environment, path, or surrounding people change. They provide almost no information about how gait authentication performs in real-world multi-person and cross-room deployments.

[0005] WiFi-based gait authentication systems are more complex in real-world applications. For example, people may walk together in hallways, move around furniture in living rooms, and freely enter and exit meeting rooms. While some research has begun to consider such multi-person or complex layout scenarios, it is typically evaluated in limited environments and provides only partial evidence for cross-domain generalization. Existing designs struggle to reliably separate mixed gait signals when multiple people are walking simultaneously and often simplify the multi-subcarrier structure of CSI to a single spectrogram, thus failing to fully utilize its advantages. Consequently, the accuracy of gait authentication often drops sharply when the environment or crowd density changes. Summary of the Invention

[0006] Based on the above background technology, this invention proposes a cross-domain adaptive multi-person gait authentication method based on commercial WiFi devices. The specific implementation steps of this method include the following:

[0007] Step 1: Collect raw Channel State Information (CSI) signals within the monitoring area using WiFi transmitters and receivers;

[0008] Step 2: Perform bandpass filtering, wavelet denoising, and conjugate multiplication preprocessing on the original Channel State Information (CSI) signal in sequence to obtain the preprocessed multi-person mixed CSI signal;

[0009] Step 3: Use the Reconstructed Independent Component Analysis (RICA) algorithm to perform blind source separation on the preprocessed multi-person mixed CSI signal, extract the independent gait components belonging to the target individual, and convert the extracted independent gait components into a subcarrier frequency map;

[0010] Step 4: Input the subcarrier frequency map into a hybrid deep learning network with a CNN feature extractor and a Transformer encoder as the backbone to extract gait features, and deploy a weakly supervised adaptive training strategy on the backbone network to achieve complete cross-domain gait identity authentication.

[0011] Furthermore, in step 1, a WiFi device equipped with an Intel 5300 network card is used, configured with a single-antenna transmission and a three-antenna reception mode; 30 orthogonal frequency division multiplexing (OFDM) subcarriers are collected within each coherent bandwidth; the sampling frequency is set to 1000Hz to ensure complete coverage of the small Doppler frequency shift within the gait period.

[0012] Furthermore, step 2 specifically includes the following steps:

[0013] Step 2-1: Bandpass filtering: Perform Fourier transform on the CSI signal of each subcarrier, apply a bandpass filter to extract the gait-related 10–70 Hz frequency components, and then use inverse Fourier transform to recover the time-domain signal; the bandpass filter adopts an asymmetric cascaded filtering architecture, which is composed of a 6th-order low-pass filter and a 4th-order high-pass filter cascaded together, with cutoff frequencies of 70Hz and 10Hz, respectively. It provides a steep attenuation of 36dB / octave at the high-frequency end to suppress electronic noise, and provides 24dB / octave attenuation at the low-frequency end to remove static offset and slow environmental changes;

[0014] Step 2-2: Wavelet Denoising: First, the complex channel response of each subcarrier after bandpass filtering is subjected to multi-level recursive Discrete Wavelet Transform (DWT) using a Symlet-4 wavelet filter to decompose the signal into approximation coefficients and detail coefficients at each level.

[0015]

[0016] in, The first bandpass filter The complex channel response of each subcarrier; j is the decomposition level. Represents the approximation coefficient. For detail coefficients;

[0017] Next, the noise standard deviation is estimated using the median absolute deviation of the first-level detail coefficients:

[0018]

[0019] in Represent the median operator; subsequently, at each decomposition level Above, the threshold is adaptively determined according to the Bayes Shrink rule:

[0020]

[0021] in, This represents the variance of the detail coefficients for that layer. To avoid small constants with zero error;

[0022] Then, for each level of detail coefficient Apply a soft thresholding function for noise reduction:

[0023]

[0024] when If the value is less than the threshold, the value is set to zero; otherwise, the amplitude is contracted towards zero. The approximation coefficient is retained without processing.

[0025] Finally, the retained approximation coefficients and the processed detail coefficients are reconstructed by inverse wavelet transform IDWT to obtain the denoised CSI signal;

[0026] Step 2-3: Conjugate Multiplication: Select the Channel Frequency Response (CFR) of each of the two receiving antennas p and q on the same WiFi network card, and multiply the CFR of one antenna by the conjugate of the CFR of the other antenna to obtain a new signal sequence. In the new signal sequence, the amplitude part is the product of the amplitudes of the two antennas, retaining the amplitude change information caused by gait. In the phase part, the hardware random phase noise shared by the two antennas cancels each other out, and the gait signal is retained in differential form.

[0027] Furthermore, step 3 specifically includes the following steps:

[0028] Step 3-1: Model multi-person gait recognition as a blind source separation (BSS) problem, with I independent signal sources. L group observed mixed signal Establish a linear mixture model:

[0029]

[0030] in , representing the mixed signal vector, , representing the source signal vector, Represents an unknown mixture parameter matrix. Indicates additive channel noise;

[0031] Step 3-2: Solve the unmixing matrix W using the Reconstructed Independent Component Analysis (RICA) algorithm. By introducing reconstruction constraint terms to replace the orthogonal constraints of traditional ICA, a balance is achieved between separation accuracy and reconstruction fidelity, and multiple statistically independent gait source signals are recovered from the mixed CSI signal matrix X.

[0032] Step 3-3: Apply window slicing data augmentation to the separated gait signal. Slide a fixed-length window along the time axis with overlap between adjacent windows. Each window segment serves as an independent training sample and inherits the identity label of the original sequence.

[0033] Steps 3-4: Perform short-time Fourier transform (STFT) on the gait slices after window slicing data enhancement, calculate the time-frequency matrix for each subcarrier of each slice, and take the amplitude value to obtain the gait spectrogram;

[0034] Steps 3-5: Constructing a subcarrier frequency map: Stack the spectra of all subcarriers within each window, and construct a two-dimensional subcarrier frequency map with frequency as the vertical axis and subcarrier index as the horizontal axis.

[0035] Furthermore, the optimization objective of the RICA algorithm is:

[0036]

[0037] Where W is the unmixing matrix; g is the nonlinear non-Gaussianity metric function, and the reconstruction constraint is... Replacing the orthogonal constraints of traditional ICA with soft constraints This enables the algorithm to effectively separate gait components even in real-world environments where the source signals are only approximately independent; This is the balance coefficient; This represents the number of time sampling points.

[0038] Furthermore, to evaluate the quality of the gait components separated by RICA and to quantify the identifiability of each target user in the mixed signal, the SSINR metric is defined:

[0039]

[0040] in, This represents the SSINR of the nth user. Indicates the first Dynamic signal power contributed by each user, This indicates the background and total interference power.

[0041] Furthermore, in step 4, the CNN feature extractor is specifically a three-layer convolutional neural network, with each layer performing convolution, batch normalization, ReLU activation, and max pooling operations in sequence;

[0042] The first layer uses 16 3×3 filters to pool the feature map from 224×224 to 112×112; the second layer uses 32 filters to pool the feature map to 56×56; and the third layer uses 64 filters to pool the feature map to 28×28, outputting a 28×28×64 feature map, which is used to extract local spatial gait features from the subcarrier frequency map.

[0043] Furthermore, in step 4, the Transformer encoder specifically comprises:

[0044] The 28×28×64 feature map output by the CNN is divided into multiple non-overlapping 7×7 feature blocks, resulting in 16 feature blocks. Each feature block is flattened and mapped to a 128-dimensional token vector through a linear projection layer, and a learnable absolute position encoding is added to form a 16×128 token sequence.

[0045] Global interaction is achieved by stacking four layers of multi-head self-attention mechanism, with four attention heads per layer and 32 dimensions per head. The correlation between all tokens is calculated to capture the long-distance dependency between low-frequency trunk movement and high-frequency limb movement gait patterns.

[0046] The token information is aggregated by global average pooling to generate a 128-dimensional shared gait feature vector, which serves as the input to the weakly supervised adaptive module.

[0047] Furthermore, in step 4, the weakly supervised adaptive training strategy includes a source domain pre-training stage and a target domain adaptive stage.

[0048] The source domain pre-training stage includes: training a shared backbone network and three independent classifier heads F1, F2, and Ft using labeled single-user data. Each classifier head is a multi-layer MLP structure. The 128-dimensional shared gait feature vector is sequentially mapped to 64-dimensional hidden features. After batch normalization (BatchNorm), ReLU activation, and Dropout processing, these features are mapped to 2-dimensional classification logits. The predicted probabilities for two categories, legitimate users and illegitimate visitors, are output through Softmax. A diversity regularization term is introduced. This prompts F1 and F2 to learn different discriminative features, and the total training loss is a weighted sum of the classification cross-entropy loss and the diversity regularization loss; where, This represents the correlation matrix between the projection directions of the first fully connected layers of the two classifiers. Denotes the Frobenius norm;

[0049] The target domain adaptation stage includes: generating high-quality pseudo-labels through a triple screening mechanism, with the triple constraints being: F1 and F2 predicting the same category for the same component are consistent; the prediction confidence is higher than the dynamic threshold, with an initial threshold of 0.95 and a minimum threshold of 0.90, and linear relaxation during 30 rounds of adaptation; the pseudo-label set of all components is precisely matched with known user identity combinations; and the samples that pass the triple screening are added to the pseudo-label set to update the target domain classifier Ft, gradually achieving alignment of target feature distribution.

[0050] Furthermore, in the aforementioned triple screening mechanism, assuming there are L users walking simultaneously in the scenario, their true identities are set as follows: RICA separates L gait components, and the set of prediction results P of the classifier for the L components must precisely satisfy P=A. Through the triple constraint, the permutation mapping is converged from L! candidate relations to a unique correspondence, thereby completing the permutation ambiguity resolution in multi-user CSI recognition without knowing in advance the specific identity corresponding to each separated component.

[0051] The beneficial effects of this invention are:

[0052] 1. Overcame the signal interference bottleneck in multi-person environments.

[0053] Traditional WiFi gait recognition technology is typically only applicable to single-person scenarios. When multiple people are present in the monitoring area simultaneously, the signals undergo complex superposition. This invention utilizes a robust independent component analysis (ICA) algorithm to model the mixed CSI signals from multiple individuals as a blind source separation problem, successfully achieving high-fidelity extraction of single-person gait features. The experimental verification of this invention recruited 15 volunteers. The experimental equipment and acquisition parameters are as follows: a WiFi device equipped with 3 transmitting antennas and 3 receiving antennas was used to collect CSI data from 30 subcarriers at a sampling frequency of 1kHz. The distance between the transmitting and receiving ends was 1 to 7 meters, and the volunteers' walking speed ranged from 1.1 to 1.7 m / s. Experiments demonstrate that in scenarios with 1-2 people walking, the system maintains an accuracy rate of over 90% in corridors, living rooms, and meeting rooms. In complex scenarios with 3-4 people walking, the system still maintains an accuracy rate of over 80%. Figure 3 As shown.

[0054] 2. Possesses excellent adaptability across environmental domains.

[0055] WiFi signals are extremely sensitive to the physical environment (such as room layout and walking paths), a phenomenon known as "domain drift." This invention employs a weakly supervised adaptive framework and compares it with representative WiFi gait authentication systems (MUID, Wi-Diag, IMep, and WiFiU). Figure 4 As shown in the plotted ROC curve, this invention achieves the best ROC characteristic curve, with the largest area under the curve, indicating its superior ability to distinguish between legitimate and illegitimate users. In a two-person scenario, this invention achieves an authentication accuracy of 93.58%, higher than all benchmark systems. This framework effectively eliminates feature distribution shifts caused by environmental changes by guiding classifier updates with pseudo-labels generated in the target domain. When users change rooms or alter their walking habits, the system does not require time-consuming retraining with re-labeled data.

[0056] 3. High-precision spatiotemporal feature representation was achieved.

[0057] Gait encompasses both instantaneous posture features and long-term rhythmic features. The system employs a hybrid deep learning architecture that cascades a CNN feature extractor and a Transformer encoder. For example... Figure 5As shown, in three environments—corridor, living room, and conference room—the accuracy of using a CNN feature extractor alone is only around 60%, the accuracy of using a CNN feature extractor with a Transformer encoder is around 85%, and the accuracy of using a CNN feature extractor, a Transformer encoder, and the aforementioned weakly supervised adaptive framework reaches over 90%. The CNN layer focuses on extracting local spatial correlations in the subcarrier frequency map. The Transformer layer utilizes a self-attention mechanism to capture the long-range temporal dependencies of gait sequences. This dual modeling approach enables the system to accurately distinguish between extremely similar gaits.

[0058] 4. Extremely low deployment cost and excellent hardware compatibility

[0059] This invention is based on commercial WiFi devices and does not require expensive specialized radar or sensor arrays. Existing office and home WiFi infrastructure can be directly converted into a sensing network, greatly reducing the economic barrier to large-scale deployment. Attached Figure Description

[0060] Figure 1 This is a flowchart of the cross-domain adaptive multi-user authentication method based on commercial WiFi devices in an embodiment of the present invention;

[0061] Figure 2 This is a flowchart of the CNN feature extractor, Transformer encoder, and weakly supervised adaptive gait authentication module in this embodiment of the invention.

[0062] Figure 3 This is a bar chart showing the recognition accuracy of the present invention under different environments and with different numbers of people;

[0063] Figure 4 This is a comparison chart of the ROC curves of the present invention and some representative algorithms in cross-domain scenarios;

[0064] Figure 5 This is a bar chart showing the recognition accuracy of different learning architectures of this invention;

[0065] Figure 6 This is a comparison chart of the average SSINR of the RICA method of this invention and existing blind source separation methods under whitening and non-whitening conditions;

[0066] Figure 7 This is a comparison chart of SSINR between the RICA method of this invention and existing blind source separation methods under different pedestrian number groups. Detailed Implementation

[0067] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings.

[0068] This invention discloses a cross-domain adaptive multi-person gait authentication method based on commercial WiFi devices. The method first collects raw CSI signals containing human gait information using commercial WiFi devices; then stabilizes the raw channel state information through bandpass filtering, wavelet denoising, and conjugate multiplication; using a reconstructed independent component analysis algorithm, it extracts independent individual gait components from the mixed CSI signals of multiple users, and introduces a source signal interference-to-noise ratio metric to measure the quality of the separated individual signals; next, it performs window-slicing-based data augmentation on the separated signals, then converts them into subcarrier frequency maps, and uses a hybrid architecture combining a CNN feature extractor and a Transformer encoder to extract deep gait features; finally, a weakly supervised adaptive mechanism achieves cross-environment and cross-path adaptation to complete identity authentication. This invention effectively solves the technical problems of low accuracy and poor environmental mobility in existing WiFi gait recognition under multi-person interference, and features seamless operation, privacy protection, and strong robustness.

[0069] A cross-domain adaptive multi-user authentication method based on commercial WiFi devices, process Figure 1 Specifically, it includes the following steps:

[0070] Step 1: Collect the raw CSI signal within the monitoring area using the WiFi transmitter and receiver, specifically:

[0071] Using a commercial WiFi device equipped with an Intel 5300 network card, configured with single-antenna transmit and three-antenna receive modes; 30 orthogonal frequency division multiplexing (OFDM) subcarriers were collected within each coherent bandwidth; the sampling frequency was set to 1000Hz to ensure complete coverage of the small Doppler shift within the gait cycle.

[0072] Step 2: Preprocess the raw CSI signal by performing bandpass filtering, wavelet denoising, and conjugate multiplication in sequence. This provides clean and stable data for subsequent signal separation and gait authentication. Specifically:

[0073] Step 2-1: Bandpass filtering: Perform Fourier transform on the CSI of each subcarrier, apply a filter to extract the gait-related 10–70 Hz frequency components, and then use inverse Fourier transform to recover the time domain signal. This step removes low-frequency slow environmental drift and high-frequency electronic noise.

[0074] Studies have shown that the frequency variation caused by effective walking ranges from 10 to 70 Hz. Therefore, bandpass filtering is applied to the data to eliminate out-of-band interference. The filtering process can be expressed as:

[0075]

[0076] in, Indicates the first Complex channel response of each subcarrier The first bandpass filter Complex channel response of each subcarrier and They represent time respectively Fourier transform and inverse transform This represents the time frequency obtained after performing a Fourier transform on the CSI time series. This represents the frequency response of the bandpass filter. Considering the difference in noise characteristics between low and high frequencies in CSI data, this system employs an asymmetric cascaded filter architecture:

[0077]

[0078] in It is a 6th-order low-pass filter (cutoff frequency 70 Hz). It is a 4th-order high-pass filter (cutoff frequency 10 Hz). This design provides a steep attenuation of 36 dB / octave at the high-frequency end, effectively suppressing high-frequency noise generated by electronic devices; and provides 24 dB / octave attenuation at the low-frequency end, removing static offset and slow environmental changes.

[0079] Step 2-2: Wavelet Denoising: Multi-level Discrete Wavelet Transform (DWT), noise estimation, adaptive threshold calculation, soft thresholding, and signal reconstruction (IDWT). This step removes impulse noise and jitter from CSI, preserves gait characteristics, and improves signal quality.

[0080] Wavelet denoising is used to further eliminate noise interference in CSI measurement data. The discrete wavelet transform (DWT) method can simultaneously analyze time-frequency domain signal characteristics without requiring a pre-defined signal model. DWT decomposes the signal into approximation coefficients and detail coefficients: approximation coefficients describe the signal trend that preserves the macroscopic features of the CSI pattern, while detail coefficients contain low-scale components characterizing high-frequency signal noise and subtle CSI features. Specifically, wavelet-based denoising involves three steps: decomposition, thresholding, and reconstruction. First, the bandpass-filtered... Complex channel response of subcarriers Multi-stage recursive DWT signal decomposition using Symlet-4 wavelet filters:

[0081]

[0082] in, The number of decomposition layers, Represents the approximation coefficient. These are the detail coefficients. The Symlet-4 wavelet possesses good symmetry and smoothness, effectively suppressing noise while preserving signal edge features. In the thresholding stage, a Bayesian denoising method is used for adaptive thresholding of the detail coefficients. The threshold is set using the median thresholding rule, that is, the noise standard deviation is estimated using the median absolute deviation of the first-level detail coefficients.

[0083]

[0084] in The median operator is used to calculate the median of the absolute values ​​of the input vector; the constant 0.6745 comes from the median of the absolute values ​​of the standard normal distribution to achieve unbiased estimation. Subsequently, at each decomposition level... Above, the threshold is adaptively determined according to the Bayes Shrink rule:

[0085]

[0086] in, This represents the variance of the detail coefficients for that layer. To avoid small constants with zero error, for each level of detail coefficients... Apply a soft threshold function:

[0087]

[0088] when The value is set to zero when it is less than the threshold; otherwise, the amplitude is contracted to suppress noise while preserving significant features. Approximation coefficients. This primarily represents the low-frequency trend and overall structural information of the CSI signal, without any thresholding processing, retaining the original values. Finally, the approximation coefficients are directly retained, while the detail coefficients undergo soft thresholding before being used together for inverse wavelet reconstruction. The retained approximation coefficients and detail coefficients are then combined using inverse wavelet transform (IDWT) to obtain the wavelet-denoised CSI signal.

[0089]

[0090] Steps 2-3: Conjugate Multiplication: Since multiple receiving antennas on the same WiFi card share the same local oscillator, the phase shifts of the signals they receive are almost identical at a given moment. Select two antennas (p and q), and multiply the CFR of one antenna by the conjugate of the CFR of the other antenna to obtain a new sequence. The amplitude component becomes the product of the amplitudes of the two antennas, still containing amplitude variation information caused by gait. The phase component's hardware noise is completely canceled out after subtraction, while the gait signal is retained in differential form. This step solves the problem of random phase noise introduced by the WiFi hardware, while preserving the amplitude and phase channel variations caused by human movement.

[0091] Step 3: Use the reconstructed independent component analysis algorithm to perform blind source separation on the preprocessed multi-person mixed CSI signal, extract the independent gait components belonging to the target individual, and convert the extracted independent gait components into a subcarrier frequency map, specifically:

[0092] Step 3-1: Blind Source Separation: In multi-person gait scenarios, each antenna-subcarrier channel records a mixed signal containing multiple walkers, static clutter, and hardware noise. Multi-person gait recognition is formalized as a BSS problem, and individual gait extraction is achieved through a linear separation model.

[0093] Blind source separation model refers to the model that separates data from a given set of observations. Reconstructing the unknown source These observations are from It was formed through mixing via unknown transmission channels. Essentially, and These are all forms of CSI expression. It is a CSI that includes multiple people with mixed gait. This is CSI that only includes the gait of one person. Assuming the number of active source signals is a known quantity, separation is achieved by determining disjoint basic sets for different sources. Let... An independent signal source The signal caused by multiple people walking is Group linear signals Composed of. At a given time The combined signal is represented as follows:

[0094]

[0095] in express The mixing parameters are represented as a vector as follows:

[0096]

[0097] in , representing the mixed signal vector, , representing the source signal vector, Represents an unknown mixture parameter matrix. This represents the additive channel noise that can be removed from the signal. Specifically:

[0098]

[0099]

[0100] The fundamental problem of blind source separation is to determine a corresponding solution mixing matrix. And to separate the mixed signals as completely as possible. These are estimated independent components, through The corresponding solution mixing vector is obtained:

[0101]

[0102] in Represents the estimated vector of the source signal. ;matrix and They are inverse matrices. Therefore, the unmixing matrix is ​​obtained. Blind source separation can be used to deduce the dynamic changes in CSI caused by the movement of each subject.

[0103] Step 3-2: Using the Reconstructed Independent Component Analysis (RICA) algorithm: Multiple statistically independent gait source signals are recovered from the mixed Wi-Fi channel state information signal matrix X. RICA introduces a reconstruction term, transforming hard orthogonal constraints into soft constraints, achieving a balance between separation accuracy and reconstruction fidelity.

[0104] Reconstruction of independent component analysis has advantages over traditional ICA, which is defined as the following optimization problem:

[0105]

[0106] in It is a nonlinear convex function. , This is the unmixing matrix, which serves as the weight matrix in this context. The non-Gaussianity metric in the objective function The core of ICA (Independent Coding) is its encouragement of independence, which approximates statistical independence by maximizing the non-Gaussianity of the projection. However, to avoid uncertainty in the solution, traditional ICA introduces orthogonal constraints. This approach typically requires the input data to be pre-whitened. While this ensures mathematical solvability, it also introduces two limitations: firstly, the orthogonal constraints are too strict, making it difficult to fully satisfy the independence assumption in real data; secondly, high-dimensional data is difficult to whiten accurately, easily introducing noise and other factors that affect the accuracy of the algorithm.

[0107] In contrast, RICA addresses both core issues by introducing reconstruction costs instead of orthogonal constraints:

[0108]

[0109] in, This is a balancing coefficient used to control the relative weights between the reconstruction term and the non-Gaussian term. This represents the number of time sampling points.

[0110] RICA relaxes the independence requirements by replacing rigid orthogonal constraints with flexible reconstruction constraints. Specifically:

[0111] Traditional ICA orthogonal constraints Strictly restricting the solution to an orthogonal matrix manifold implicitly requires that the source signals be sufficiently independent to guarantee that the optimal solution lies within that manifold. When the real data deviates from this assumption, the algorithm can only search for suboptimal solutions within the restricted space. RICA's reconstruction constraints... This provides a continuous, trade-off constraint mechanism: it does not presuppose the geometric structure of the solution, but rather prevents solution degradation by penalizing reconstruction errors. When the source signals are completely independent, the non-Gaussian term... It can guide the algorithm to find independent solutions; when the source signals are only approximately independent, the algorithm can dynamically balance between the "degree of independence" and the "reconstruction quality", so that it can still learn effective feature representations under more relaxed conditions.

[0112] Step 3-3: Constructing SSINR: To evaluate the quality of the gait components separated by RICA and quantify the identifiability of each target user in the mixed signal, this invention defines the SSINR metric. This metric is only used to verify and compare the separation effects of different blind source separation methods, does not participate in the RICA separation process, and does not provide feedback adjustment to the separation process. For the first... For each user, SSINR is defined as:

[0113]

[0114] in, Indicates the first Dynamic signal power contributed by each user, The SSINR represents the total power of background and interference, including line-of-sight path components, static environmental reflections, hardware noise, and dynamic gait components from all other users. Intuitively, the numerator of the SSINR measures the intensity of the target user's reflected signal, while the denominator measures the degree to which the signal is masked by static clutter and other pedestrians. A higher SSINR value indicates that the user's gait components are more prominent in the mixed signal, and therefore easier to separate and identify.

[0115] To verify the advantages of RICA over other blind source separation methods, this invention compares the SSINR performance of various strategies under different pedestrian numbers. Figure 6This chart compares the average SSINR of the RICA method of this invention with existing blind source separation methods under whitening and non-whitening conditions. Whitening consistently improves the SSINR of all methods because it decorrelateds the mixed channels and balances their energy. Among all the methods compared, RICA achieves the highest SSINR under both whitening and non-whitening conditions, with a small performance difference between the two settings, indicating its low sensitivity to whitening and its ability to maintain stable gait component separation capability even under interference and noise environments.

[0116] Figure 7 This chart compares the SSINR of RICA (Invention of the Technique for the Creation of the Pedestrians) with existing blind source separation methods under different pedestrian number groups. Groups 1-3 represent two-person walking scenarios, groups 4-6 represent three-person walking scenarios, and groups 7-9 represent four-person walking scenarios. As the number of pedestrians increases, the SSINR of all methods decreases due to enhanced interference between users and more complex multipath aliasing. Nevertheless, RICA consistently outperforms tensor decomposition and the classic ICA baseline method in all groups. This stable advantage, exhibited in both low-density and high-density crowd scenarios, is the reason why RICA was chosen as the core blind source separation method in this invention.

[0117] The single-person gait signal obtained after RICA separation differs in distribution from the directly acquired single-person gait CSI signal. The single-person CSI directly reflects the user's original gait characteristics, while the separated signal undergoes a hybrid separation reconstruction process, introducing additional signal processing paths. This domain shift directly affects the generalization performance of downstream recognition models; if the model is trained only on single-person data, performance may degrade when applied to separated multi-person data.

[0118] Steps 3-4: To mitigate the aforementioned domain shift, a direct approach is to collect more labeled trials covering diverse environments and crowd scenarios. However, in actual indoor deployments, such data collection is often costly and limited, making it difficult to obtain large-scale labeled data. This invention employs a simple time-series data augmentation technique using window slicing, effectively expanding the training dataset, alleviating the small sample size problem, enabling the model to learn the local dynamic features of gait signals, and enhancing the model's robustness to time shifts, thereby indirectly mitigating the domain shift impact between single-person data and separated data. On the separated gait signals, a fixed-length window slides along the time axis, with overlap between adjacent windows; each extracted window segment is considered an independent training sample and inherits the identity label of the original complete sequence; by selecting a window length shorter than the complete sequence and a smaller sliding step size, a large number of overlapping slices can be generated from a gait trajectory.

[0119] Steps 3-5: Short-Time Fourier Transform (STFT): The gait slices after windowed data enhancement are converted into a two-dimensional time-frequency representation that reflects the time and frequency variation characteristics, so as to be used for subsequent construction of subcarrier-frequency maps for gait recognition. The pedestrian's gait data is divided into multiple slices, each slice being a discrete-time signal of length N; each slice is processed using a Short-Time Fourier Transform (STFT) to obtain the time-frequency matrix of each slice; the amplitude value of each STFT result is taken to obtain the gait spectrogram, preserving the time-varying characteristics and frequency structure of the signal.

[0120] Due to radio frequency signal reflection, different parts of the human body contain multiple frequency components. After preprocessing and signal separation, the CSI measurement values ​​are converted to the frequency domain using a short-time Fourier transform (STFT). The basic idea of ​​STFT is to divide the time-domain signal into multiple segments using a window, and then calculate the Fourier transform of each segment.

[0121] First, let's introduce the Fast Fourier Transform (FFT), which is widely used to transform time-domain signals. Convert to its frequency domain representation :

[0122]

[0123] in Indicates the number of sample points. Indicates frequency, The imaginary unit, This represents the discrete-time sample index. (Based on the signal...) Add windows Then, frame-by-frame FFT processing is performed to achieve short-time Fourier transform:

[0124]

[0125] Steps 3-6: Constructing a subcarrier frequency graph: Windowing: Divide the CSI amplitude signal into multiple overlapping time windows; Subcarrier-level spectrum extraction: Calculate the spectrum of each subcarrier in each window; Construct a two-dimensional graph: Stack the spectra of all subcarriers in each window to construct a subcarrier-frequency graph, with the vertical axis representing frequency and the horizontal axis representing subcarrier index.

[0126] Step 4: As Figure 2 The flowchart describes how the subcarrier frequency map is input into a CNN feature extractor, and then a Transformer encoder captures long-term temporal relationships to fully extract time-frequency domain gait features. On this backbone network, a weakly supervised adaptive training strategy is deployed to adapt to cross-domain transformations, specifically:

[0127] Step 4-1: CNN Feature Extractor: Employs a three-layer convolutional neural network to filter gait features layer by layer.

[0128] The first convolutional layer uses 16 3×3 filters to find the most basic local patterns. After pooling compression, the image size is reduced from 224×224 to 112×112, retaining only the strongest signals.

[0129] The second convolutional layer increases to 32 filters, combining the scattered features found in the first layer to identify slightly larger patterns. After pooling again, the image is reduced to 56×56, and the features are more refined.

[0130] The third convolutional layer uses 64 filters to integrate information from the first two layers and understand the overall structure and layout of the image. After the final pooling, the image is finally reduced to 28×28.

[0131] Each layer strictly follows the process of convolution (feature extraction), batch normalization (stabilizing numerical range), ReLU activation (filtering negative values), and max pooling (compressing size) to ensure stable training and effective feature extraction. Gait features that can distinguish different people are automatically extracted from the subcarrier-frequency map generated by CSI.

[0132] Step 4-2: Transformer encoder: captures long-distance dependencies between distant regions in the subcarrier-frequency map, associates gait patterns of low-frequency trunk movements with high-frequency limb movements, and thus constructs gait features with global perception capabilities.

[0133] Blocking and Embedding: The 28×28×64 feature map output by the CNN is divided into multiple non-overlapping 7×7 feature blocks, resulting in 16 feature blocks. For each feature block, it is flattened into a 3136-dimensional vector and mapped to a 128-dimensional token vector through a linear projection layer, resulting in the input token sequence of the Transformer encoder with a size of 16×128. Learnable absolute position encoding is added to the input sequence to preserve the spatial location information of each feature block in the original feature map.

[0134] Global interaction: By stacking 4 layers of multi-head self-attention mechanism, with 4 multi-head attentions per layer and 32 dimensions per attention head, the correlation between all tokens is calculated, and gait feature segments distributed on various frequencies and subcarriers are associated.

[0135] Feature output: After processing by the Transformer layer, global average pooling is used to aggregate token information to generate a compact shared gait feature vector containing global context.

[0136] The output contains a unified feature representation that includes both local spectral details and global structural patterns. This feature fully models the collaborative relationships between different body parts during human walking, providing high-quality input for subsequent weakly supervised classification and identity authentication.

[0137] Step 4-3: Weakly Supervised Adaptation: In a weakly supervised adaptive domain scenario with only set-level labels, this step addresses the permutation ambiguity problem in multi-user CSI recognition, achieving high-precision individual identification. Set-level labels refer to the labels of all pedestrians in a multi-user experiment, rather than the specific labels of each separated component. For example, in an experiment where two people walk simultaneously, after data collection, signal separation techniques are used to separate the mixed signal into two independent sources; the set of labels for these two independent sources is the set-level label.

[0138] Assume the scene has If multiple users walk simultaneously, their set of real identities is denoted as […]. RICA separated Each gait component is denoted as .

[0139] Since the separation algorithm does not guarantee that the order of the output components matches the order of the user identities, there exists an unknown permutation mapping. This makes the separated components The corresponding real user identity is The objective of this invention is to... In this case, correctly identify the user identity corresponding to each component.

[0140] The set-level labels provide a collection of all real identities in this experiment. Let the classifier be... The set of prediction results for each component is The constraints provided by the set-level labels are:

[0141]

[0142] That is, classifier pairs The set of prediction results for each component must precisely match the set-level label set of the experiment.

[0143] Without a triple screening mechanism, the number of correspondences satisfying the above set constraints is [number missing]. There are ! types (e.g., 2!=2 types for a two-player game and 3!=6 types for a three-player game). The triple-screening mechanism of this invention uses consistency constraints and dynamic thresholds to progressively filter high-quality pseudo-labels, reducing the number of possible correspondences from... The number of possibilities converges to one, thus uniquely determining the user identity corresponding to each component.

[0144] A weakly supervised adaptive framework based on multi-classifier collaboration and combinatorial consistency constraints includes two training phases and a triple selection mechanism:

[0145] Source Domain Pre-training: A shared backbone network and three classifier heads are trained using labeled single-user data. After processing by a Transformer encoder, the output shared gait feature vector is 128-dimensional. This shared gait feature vector is then input into three independent classifier heads, F1, F2, and Ft. F1 and F2 are used to generate pseudo-labels, while Ft is used for target domain classification and final authentication output. Each classifier head employs a multi-layer MLP structure: the 128-dimensional shared gait features are first mapped to 64-dimensional hidden features, then processed by batch normalization (BatchNorm), ReLU activation, and Dropout, before being mapped to 2-dimensional classification logits. Softmax is then used to output the predicted probabilities for both legitimate users and illegitimate visitors; the Dropout rate is 0.5.

[0146] The introduction of diversity regularization encourages F1 and F2 to learn different discriminative features, reducing the correlation between their errors and laying the foundation for subsequent consensus screening. The specific form of this diversity regularization is as follows:

[0147] Let the shared gait feature vector output by the CNN feature extractor and the Transformer encoder be...

[0148]

[0149] in, This represents the shared gait feature vector input to the weakly supervised adaptive module.

[0150] Set classifier head and The first fully connected layer of both maps the 128-dimensional shared gait features to 64-dimensional hidden features. The weight matrices of the first fully connected layer of both are expressed as follows:

[0151]

[0152] in, Represents the classifier head The first fully connected layer weight matrix, Represents the classifier head The weight matrix of the first fully connected layer.

[0153] The corresponding linear output of the first fully connected layer can be expressed as:

[0154]

[0155]

[0156] in, These represent the classifier heads respectively. and The bias term of the first fully connected layer These represent the classifier heads respectively. and The first fully connected layer has a linear output.

[0157] To prompt the classifier head and By learning different discriminative features, a diversity regularization term is introduced:

[0158]

[0159] in, This represents the correlation matrix between the projection directions of the first fully connected layers of the two classifiers. This represents the Frobenius norm.

[0160] The classification cross-entropy loss can be expressed as

[0161] ,

[0162] in, represents the target domain classifier header, and y represents the sample label. This represents the cross-entropy loss.

[0163] The training loss after adding diversity regularization is

[0164]

[0165] in, Represents the classifier head , and The sum of the classification cross-entropy loss, This represents the diversity regularization weight coefficient. Represents the classifier head and The weights are used to remove the relevant losses.

[0166] Target domain adaptation: Pseudo-label filtering must simultaneously satisfy the following three constraints:

[0167] F1 and F2 predict the same category for the same component;

[0168] If the prediction confidence is higher than the current dynamic threshold, a linear relaxation strategy is used for the threshold. The initial threshold is 0.95, the minimum threshold is 0.90, and the adaptive process is set to 30 rounds. In each round, the threshold is uniformly reduced by approximately 0.00167 until it reaches 0.90 and then remains constant. Pseudo-labels are only added to the pseudo-label training set if the F1 and F2 predictions are consistent, the confidence exceeds the current threshold, and the pseudo-label combination matches the known user identity combination.

[0169] In the same experiment, the pseudo-label set for all components must precisely match known combinations of user identities. Through this triple-screening mechanism, the system can filter out high-quality pseudo-labels without prior knowledge of the specific user identity corresponding to each separated component. The correspondence between components and identities is implicitly and uniquely determined during training through statistical accumulation of a large number of samples. Training parameters: In this invention, the CNN + Transformer encoder is trained end-to-end, and the entire weakly supervised adaptive process is a two-stage training; the loss function is primarily binary cross-entropy with a classifier diversity regularization term added; the optimizer is Adam with an initial learning rate of 0.001. The target domain adaptive process is set to 30 rounds of outer loop iteration. Each round uses a mini-batch method to update network parameters, with a batch size of 64. The overall model is trained offline, with a total of 200 epochs.

[0170] The selected samples are added to a pseudo-label set to update the target domain classifier Ft, achieving a gradual alignment of the target feature distribution. The selection threshold is dynamically relaxed during the adaptive process to balance sample quality and quantity, ultimately outputting the authentication result.

Claims

1. A cross-domain adaptive multi-person gait authentication method based on commercial WiFi devices, characterized in that, Includes the following steps: Step 1: Collect raw Channel State Information (CSI) signals within the monitoring area using WiFi transmitters and receivers; Step 2: Perform bandpass filtering, wavelet denoising, and conjugate multiplication preprocessing on the original Channel State Information (CSI) signal in sequence to obtain the preprocessed multi-person mixed CSI signal; Step 3: Use the Reconstructed Independent Component Analysis (RICA) algorithm to perform blind source separation on the preprocessed multi-person mixed CSI signal, extract the independent gait components belonging to the target individual, and convert the extracted independent gait components into a subcarrier frequency map; Step 4: Input the subcarrier frequency map into a hybrid deep learning network with a CNN feature extractor and a Transformer encoder as the backbone to extract gait features, and deploy a weakly supervised adaptive training strategy on the backbone network to achieve complete cross-domain gait identity authentication.

2. The method according to claim 1, characterized in that, In step 1, a WiFi device equipped with an Intel 5300 network card is used, configured with a single-antenna transmit mode and a three-antenna receive mode; 30 orthogonal frequency division multiplexing (OFDM) subcarriers are collected within each coherent bandwidth; the sampling frequency is set to 1000Hz to ensure complete coverage of the small Doppler frequency shift within the gait period.

3. The method according to claim 1, characterized in that, Step 2 specifically includes the following steps: Step 2-1: Bandpass filtering: Perform Fourier transform on the CSI signal of each subcarrier, apply a bandpass filter to extract the gait-related 10–70 Hz frequency components, and then use inverse Fourier transform to recover the time-domain signal; the bandpass filter adopts an asymmetric cascaded filtering architecture, which is composed of a 6th-order low-pass filter and a 4th-order high-pass filter cascaded together, with cutoff frequencies of 70Hz and 10Hz, respectively. It provides a steep attenuation of 36dB / octave at the high-frequency end to suppress electronic noise, and provides 24dB / octave attenuation at the low-frequency end to remove static offset and slow environmental changes; Step 2-2: Wavelet Denoising: First, the complex channel response of each subcarrier after bandpass filtering is subjected to multi-level recursive Discrete Wavelet Transform (DWT) using a Symlet-4 wavelet filter to decompose the signal into approximation coefficients and detail coefficients at each level. ； in, The first bandpass filter The complex channel response of each subcarrier; j is the decomposition level. Represents the approximation coefficient. For detail coefficients; Next, the noise standard deviation is estimated using the median absolute deviation of the first-level detail coefficients: ； in Represent the median operator; subsequently, at each decomposition level Above, the threshold is adaptively determined according to the Bayes Shrink rule: ； in, This represents the variance of the detail coefficients for that layer. To avoid small constants with zero error; Then, for each level of detail coefficient Apply a soft thresholding function for noise reduction: ； when If the value is less than the threshold, the value is set to zero; otherwise, the amplitude is contracted towards zero. The approximation coefficient is retained without processing. Finally, the retained approximation coefficients and the processed detail coefficients are reconstructed by inverse wavelet transform IDWT to obtain the denoised CSI signal; Step 2-3: Conjugate Multiplication: Select the Channel Frequency Response (CFR) of each of the two receiving antennas p and q on the same WiFi network card, and multiply the CFR of one antenna by the conjugate of the CFR of the other antenna to obtain a new signal sequence. In the new signal sequence, the amplitude part is the product of the amplitudes of the two antennas, retaining the amplitude change information caused by gait. In the phase part, the hardware random phase noise shared by the two antennas cancels each other out, and the gait signal is retained in differential form.

4. The method according to claim 1, characterized in that, Step 3 specifically includes the following steps: Step 3-1: Model multi-person gait recognition as a blind source separation (BSS) problem, with I independent signal sources. L group observed mixed signal Establish a linear mixture model: ； in , representing the mixed signal vector, , representing the source signal vector, Represents an unknown mixture parameter matrix. Indicates additive channel noise; Step 3-2: Solve the unmixing matrix W using the Reconstructed Independent Component Analysis (RICA) algorithm. By introducing reconstruction constraint terms to replace the orthogonal constraints of traditional ICA, a balance is achieved between separation accuracy and reconstruction fidelity, and multiple statistically independent gait source signals are recovered from the mixed CSI signal matrix X. Step 3-3: Apply window slicing data augmentation to the separated gait signal. Slide a fixed-length window along the time axis with overlap between adjacent windows. Each window segment serves as an independent training sample and inherits the identity label of the original sequence. Steps 3-4: Perform short-time Fourier transform (STFT) on the gait slices after window slicing data enhancement, calculate the time-frequency matrix for each subcarrier of each slice, and take the amplitude value to obtain the gait spectrogram; Steps 3-5: Constructing a subcarrier frequency map: Stack the spectra of all subcarriers within each window, and construct a two-dimensional subcarrier frequency map with frequency as the vertical axis and subcarrier index as the horizontal axis.

5. The method according to claim 3, characterized in that, The optimization objective of the RICA algorithm is: ； Where W is the unmixing matrix; g is the nonlinear non-Gaussianity metric function, and the reconstruction constraint is... Replacing the orthogonal constraints of traditional ICA with soft constraints This enables the algorithm to effectively separate gait components even in real-world environments where the source signals are only approximately independent; This is the balance coefficient; This represents the number of time sampling points.

6. The method according to claim 3, characterized in that, To evaluate the quality of the gait components separated by RICA and to quantify the identifiability of each target user in the mixed signal, the SSINR metric is defined: ； in, This represents the SSINR of the nth user. Indicates the first Dynamic signal power contributed by each user, This indicates the background and total interference power.

7. The method according to claim 1, characterized in that, In step 4, the CNN feature extractor is specifically a three-layer convolutional neural network, with each layer performing convolution, batch normalization, ReLU activation and max pooling operations in sequence. The first layer uses 16 3×3 filters to pool the feature map from 224×224 to 112×112; the second layer uses 32 filters to pool the feature map to 56×56; and the third layer uses 64 filters to pool the feature map to 28×28, outputting a 28×28×64 feature map, which is used to extract local spatial gait features from the subcarrier frequency map.

8. The method according to claim 1, characterized in that, In step 4, the Transformer encoder specifically includes: The 28×28×64 feature map output by the CNN is divided into multiple non-overlapping 7×7 feature blocks, resulting in 16 feature blocks. Each feature block is flattened and mapped to a 128-dimensional token vector through a linear projection layer, and a learnable absolute position encoding is added to form a 16×128 token sequence. Global interaction is achieved by stacking four layers of multi-head self-attention mechanism, with four attention heads per layer and 32 dimensions per head. The correlation between all tokens is calculated to capture the long-distance dependency between low-frequency trunk movement and high-frequency limb movement gait patterns. The token information is aggregated by global average pooling to generate a 128-dimensional shared gait feature vector, which serves as the input to the weakly supervised adaptive module.

9. The method according to claim 1, characterized in that, In step 4, the weakly supervised adaptive training strategy includes a source domain pre-training stage and a target domain adaptive stage. The source domain pre-training stage includes: training a shared backbone network and three independent classifier heads F1, F2, and Ft using labeled single-user data. Each classifier head is a multi-layer MLP structure. The 128-dimensional shared gait feature vector is sequentially mapped to 64-dimensional hidden features. After batch normalization (BatchNorm), ReLU activation, and Dropout processing, these features are mapped to 2-dimensional classification logits. The predicted probabilities for two categories, legitimate users and illegitimate visitors, are output through Softmax. A diversity regularization term is introduced. This prompts F1 and F2 to learn different discriminative features, and the total training loss is a weighted sum of the classification cross-entropy loss and the diversity regularization loss; where, This represents the correlation matrix between the projection directions of the first fully connected layers of the two classifiers. Denotes the Frobenius norm; The target domain adaptation stage includes: generating high-quality pseudo-labels through a triple screening mechanism, with the triple constraints being: F1 and F2 predicting the same category for the same component are consistent; the prediction confidence is higher than the dynamic threshold, with an initial threshold of 0.95 and a minimum threshold of 0.90, and linear relaxation during 30 rounds of adaptation; the pseudo-label set of all components is precisely matched with known user identity combinations; and the samples that pass the triple screening are added to the pseudo-label set to update the target domain classifier Ft, gradually achieving alignment of target feature distribution.

10. The method according to claim 9, characterized in that, In the aforementioned triple screening mechanism, assuming there are L users walking simultaneously in the scenario, their true identities are set as follows: RICA separates L gait components, and the set of prediction results P of the classifier for the L components must precisely satisfy P=A. Through the above triple constraints, the permutation mapping is converged from L! candidate relations to a unique correspondence, thereby completing the permutation ambiguity resolution in multi-user CSI recognition without knowing in advance the specific identity corresponding to each separated component.

Images