A gait feature construction method based on wireless signals

Abstract

The application discloses a gait feature construction method based on a wireless signal, and comprises the following steps: acquiring frequency-modulated continuous wave signals of multiple users collected by a radar transceiver device; obtaining corresponding filtered Doppler spectrograms of the frequency-modulated continuous wave signals based on high-pass filtering processing; performing gait discretization processing to obtain a gait signal scatter plot; sequentially performing regression processing and smoothing filtering processing on the gait signal scatter plot to obtain a gait feature curve; determining a periodic wave peak in the gait feature curve to obtain a segmentation starting point, and performing gait segmentation on the gait feature curve by using the segmentation starting point and a fixed time window to obtain multiple gait sub-spectrograms segmented from the frequency-modulated continuous wave signal; and the gait sub-spectrograms segmented from all the frequency-modulated continuous wave signals are used for training a preset model, and the preset model comprises an identity recognition model and a gait anomaly recognition model. The application can overcome the gait time-varying problem occurring in the construction process of the Doppler spectrogram, and can improve the accuracy and robustness of recognition.

Description

Technical Field

[0001] This invention belongs to the field of gait feature recognition, specifically relating to a method for constructing gait features based on wireless signals. Background Technology

[0002] Gait feature construction is a key technical issue in the field of gait recognition, as gait features can be used for identity recognition and abnormal gait diagnosis. Currently, identity recognition mainly adopts close-range active identification methods such as fingerprint recognition, facial recognition, and voice recognition; abnormal gait diagnosis usually requires data collection using wearable devices, which may not be suitable for prolonged use by patients. In contrast, gait Doppler spectra generated based on millimeter-wave signals can achieve long-range, contactless identity recognition without additional user cooperation. However, as the distance between the target and the transceiver increases, the target's RCS (Radar cross-section) decreases, leading to energy attenuation in the corresponding region of the Doppler spectrum, such as... Figure 1 As shown in Figure (a). Furthermore, in the process of gait feature construction, using the time-sliding window segmentation method can lead to differences in sample features. The results of the time-sliding window segmentation method are shown in Figure (a). Figure 1 As shown in Figure (b), the sliding window segmentation method disrupts the feature alignment of sample data, causing local features in the sub-spectral maps to no longer correspond to the same gait stage. This significantly impacts the sample distribution and reduces the model's classification performance. Based on the above, it can be seen that the original gait Doppler spectrogram cannot provide consistent recognition performance in all scenarios and cannot adequately meet the requirements of gait-based identity recognition.

[0003] Existing gait recognition methods improve the robustness of network models by training them with diverse, large datasets in different directions and positions. However, these methods require significant manpower for data collection. Furthermore, some methods segment Doppler spectrograms based on velocity in hopes of obtaining gait sub-spectral maps with uniform features; however, if the extracted velocity does not accurately reflect changes in leg speed, using this velocity for foot segmentation can lead to erroneous results. Summary of the Invention

[0004] To address the aforementioned problems in the existing technology, this invention provides a method for constructing gait features based on wireless signals, a method for training an identity recognition model, an identity recognition method, a method for training a gait anomaly recognition model, and a gait anomaly recognition method. The technical problems to be solved by this invention are achieved through the following technical solutions:

[0005] In a first aspect, embodiments of the present invention provide a gait feature construction method based on wireless signals, the gait feature construction method based on wireless signals comprising:

[0006] Acquire frequency-modulated continuous wave signals from multiple users collected by radar transceiver equipment;

[0007] For each frequency-modulated continuous wave signal, a filtered Doppler spectrum is obtained based on high-pass filtering.

[0008] The filtered Doppler spectrum is discretized to obtain a gait signal scatter plot.

[0009] The gait signal scatter plot is subjected to regression processing and smoothing filtering in sequence to obtain the gait feature curve;

[0010] The periodic peaks in the gait feature curve are determined to obtain the segmentation starting point. The gait feature curve is then segmented using the segmentation starting point and a fixed time window to obtain multiple gait sub-spectral maps segmented from the frequency-modulated continuous wave signal. All gait sub-spectral maps segmented from the frequency-modulated continuous wave signal are used to train a preset model, which includes an identity recognition model and a gait anomaly recognition model.

[0011] In one embodiment of the present invention, the step of obtaining the filtered Doppler spectrum corresponding to each frequency-modulated continuous wave signal based on high-pass filtering includes:

[0012] For each frequency-modulated continuous wave signal, perform a range-dimensional fast Fourier transform to obtain the range spectrum of the frequency-modulated continuous wave signal;

[0013] The distance spectrum is then subjected to high-pass filtering.

[0014] Perform a velocity-dimensional fast Fourier transform on the high-pass filtered distance spectrum to obtain the filtered Doppler spectrum corresponding to the frequency-modulated continuous wave signal.

[0015] In one embodiment of the present invention, the filtered Doppler spectrum is subjected to gait discretization processing to obtain a gait signal scatter plot, including:

[0016] The energy points below the preset energy threshold in the filtered Doppler spectrum are deleted to obtain a scatter plot of the gait signal.

[0017] In one embodiment of the present invention, determining the periodic peaks in the gait feature curve to obtain the segmentation starting point includes:

[0018] The peak position is located by differentiating the gait characteristic curve using derivative analysis.

[0019] The periodic peaks obtained from the peak position analysis are used as the starting point for segmentation.

[0020] In one embodiment of the present invention, a periodic peak is obtained based on the peak position analysis, which serves as the segmentation starting point, including:

[0021] Based on the peak position, periodic peaks are selected using the peak period comparison method and used as the starting point for segmentation.

[0022] In one embodiment of the present invention, a periodic peak is obtained based on the peak position analysis, which serves as the segmentation starting point, including:

[0023] Autocorrelation analysis is performed on the gait feature curve to determine the correct period. The correct period is then used to select periodic peaks from the peak positions as the segmentation starting point.

[0024] Secondly, embodiments of the present invention provide a training method for an identity recognition model, the training method comprising:

[0025] Obtain the training set of the identity recognition model; wherein the training set of the identity recognition model contains training data of multiple users, and the training data of each user is multiple gait sub-spectral maps segmented from the user's frequency-modulated continuous wave signal, which are obtained according to the gait feature construction method based on wireless signal described in the first aspect, and each gait sub-spectral map is labeled with user information as tag data.

[0026] A preset neural network model is trained using the training set of the identity recognition model to obtain the trained identity recognition model.

[0027] Thirdly, embodiments of the present invention provide an identity recognition method, the identity recognition method comprising:

[0028] Obtain the gait sub-spectrum of the user to be identified;

[0029] The gait sub-spectrum of the user to be identified is input into a pre-trained identity recognition model to obtain the corresponding user identity information; wherein the identity recognition model is obtained according to the training method of the identity recognition model described in the second aspect.

[0030] Fourthly, embodiments of the present invention provide a training method for a gait anomaly recognition model, the training method comprising:

[0031] Obtain the training set of the gait anomaly recognition model; wherein the training set of the gait anomaly recognition model contains training data of multiple users, and the training data of each user is multiple gait sub-spectral maps segmented from the user's frequency-modulated continuous wave signal, which are obtained according to the gait feature construction method based on wireless signal described in the first aspect, and each gait sub-spectral map is marked with label data to characterize whether the gait is abnormal.

[0032] A preset neural network model is trained using the training set of the gait anomaly recognition model to obtain the trained gait anomaly recognition model.

[0033] Fifthly, embodiments of the present invention provide a gait anomaly identification method, the gait anomaly identification method comprising:

[0034] Obtain the gait sub-spectrum map to be identified;

[0035] The gait sub-spectrum to be identified is input into a pre-trained gait anomaly identification model to obtain gait identification results, which include normal gait or abnormal gait; wherein the gait anomaly identification model is obtained according to the training method of the gait anomaly identification model described in the fourth aspect.

[0036] The beneficial effects of this invention are:

[0037] The gait feature construction method based on wireless signals provided in this invention first extracts and reconstructs gait information, and then effectively segments the Doppler spectrum using the regularity of gait to obtain gait sub-spectral maps with consistent gait features and similar energy intensities. This overcomes the gait time-varying problem that occurs during Doppler spectrum construction. These sub-spectral maps can be used to train identity recognition models and gait anomaly recognition models, improving recognition accuracy and robustness, resulting in higher recognition accuracy for the identity recognition models and gait anomaly recognition models implemented on this basis. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of arbitrary segmentation of gait Doppler spectra in the prior art, where (a) is the gait Doppler spectrum and (b) is the gait sub-spectrum;

[0039] Figure 2 This is a flowchart illustrating a gait feature construction method based on wireless signals provided in an embodiment of the present invention.

[0040] Figure 3 This is the unfiltered Doppler spectrum of the frequency-modulated continuous wave signal in this embodiment of the invention;

[0041] Figure 4 These are the results of high-pass filtering of the range spectrum in this embodiment of the invention;

[0042] Figure 5 This is an embodiment of the present invention. Figure 3 The corresponding filtered Doppler spectrum;

[0043] Figure 6 This is an embodiment of the present invention. Figure 5 The corresponding gait signal scatter plot;

[0044] Figure 7 This is the result of regression processing of the gait signal scatter plot in this embodiment of the invention;

[0045] Figure 8 This is a schematic diagram of the gait characteristic curve in an embodiment of the present invention;

[0046] Figure 9 This is a flowchart illustrating a training method for an identity recognition model provided in an embodiment of the present invention.

[0047] Figure 10 This is a schematic diagram illustrating the specific process of the training method for the identity recognition model provided in this embodiment of the invention;

[0048] Figure 11 This is a schematic diagram of the segmentation window and the segmented gait sub-spectrum in an embodiment of the present invention;

[0049] Figure 12 This is a schematic diagram of the structure of an identity recognition model in an embodiment of the present invention;

[0050] Figure 13 This is a flowchart illustrating an identity recognition method provided in an embodiment of the present invention;

[0051] Figure 14 This is a flowchart illustrating a training method for a gait anomaly recognition model provided in an embodiment of the present invention.

[0052] Figure 15 This is a schematic diagram illustrating the specific process of training the gait anomaly recognition model provided in this embodiment of the invention;

[0053] Figure 16 This is a gait characteristic curve diagram of hemiplegic gait in an embodiment of the present invention;

[0054] Figure 17 This is a schematic diagram of gait periodicity analysis in an embodiment of the present invention, wherein (a) is the autocorrelation analysis result of normal gait and (b) is the autocorrelation analysis result of hemiplegic gait;

[0055] Figure 18These are comparison diagrams of gait segmentation results in embodiments of the present invention, wherein (a) is a normal gait spectrum, (b) is a hemiplegic gait spectrum, (c) is a sub-spectrum obtained after normal gait segmentation, and (d) is a gait sub-spectrum segmented after periodic analysis.

[0056] Figure 19 This is a schematic diagram of a gait anomaly recognition model in an embodiment of the present invention;

[0057] Figure 20 This is a flowchart illustrating a gait anomaly recognition method provided in an embodiment of the present invention. Detailed Implementation

[0058] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0059] To overcome the time-varying nature of gait during Doppler spectrum construction, this invention proposes a gait feature construction method based on wireless signals. Based on this method, it also provides a training method for an identity recognition model and an identity recognition method, which can improve the accuracy of user identification. Furthermore, based on the proposed gait feature construction method based on wireless signals, it provides a training method for a gait anomaly recognition model and a gait anomaly recognition method, which can achieve accurate diagnosis of abnormal gait.

[0060] In a first aspect, embodiments of the present invention provide a method for constructing gait features based on wireless signals, such as... Figure 2 As shown, the method may include the following steps:

[0061] S1, acquire frequency-modulated continuous wave signals from multiple users collected by the radar transceiver equipment;

[0062] The embodiments of the present invention can utilize any type of radar transceiver to receive frequency modulated continuous wave (FMCW) signals from users. For example, in one optional embodiment, the radar transceiver can be a Texas Instruments AWR1443 FMCW radar transceiver, which operates at 77 GHz and has a bandwidth of 3.4 GHz, providing a range resolution of 4.41 cm and a maximum detectable range of 16 m.

[0063] S2, For each frequency modulated continuous wave signal, based on high-pass filtering, the filtered Doppler spectrum corresponding to the frequency modulated continuous wave signal is obtained;

[0064] In the frequency-modulated continuous wave signal received by radar, in addition to information about human motion, there are also reflection signals from static objects in the environment, such as walls, the ground, and other obstacles. This static information appears in the Doppler spectrum as low-frequency components with zero velocity and extremely high energy (see [link to Doppler spectrum]). Figure 3 These low-frequency signals suppress the expression of dynamic components. Furthermore, as the target object moves away from the radar, the signal returned by the moving object becomes increasingly weak due to signal strength attenuation, making it difficult to effectively extract useful gait information.

[0065] To address the above issues, this invention employs a filter to suppress low-frequency signals and extracts key gait information, allowing focus to be placed solely on the dynamic components generated by gait movements.

[0066] Specifically, S2 may include the following steps:

[0067] S21. For each frequency-modulated continuous wave signal, perform a range-dimensional fast Fourier transform to obtain the range spectrum of the frequency-modulated continuous wave signal.

[0068] For details on performing the range-dimensional Fast Fourier Transform, please refer to the relevant technical explanations; it will not be explained in detail here. For the range spectrum diagram of the frequency-modulated continuous wave signal, please refer to [link to relevant documentation]. Figure 4 As shown.

[0069] S22, perform high-pass filtering on the distance spectrum;

[0070] Specifically, this step employs a signal processing method based on a Butterworth high-pass filter. The high-pass filter is designed based on a difference equation, and its specific operation is expressed as follows:

[0071]

[0072] Among them, b i and a i These are the Butterworth filter coefficients, where x[n] and y[n] are the input and output signals, respectively; m is the Butterworth filter order (i.e., m delayed signals); for example, if m = 2, the Butterworth filter will use the current input signal x[n], the input signals x[n-1] and x[n-2] from the previous two time points, and the output signals y[n-1] and y[n-2] from the previous two time points to calculate the current output y[n]. If the filtering result at the fourth frequency point is calculated, then:

[0073] y[4]=b0x[4]+b1x[3]+b2x[2]-a1y[3]-a2y[2].

[0074] Specifically, the input data for the Butterworth high-pass filter is the data from each row of the distance spectrum obtained after performing a distance-dimensional FFT (Fast Fourier Transform). The Butterworth filter is used to filter N time points (i.e., each column) at each distance point (i.e., each row), where n ranges from 1 to N-1 (this is to prevent out-of-bounds access). Through the difference equation, low-frequency signals, due to their greater dependence on the preceding data, are filtered out. High-frequency signals are less affected because their frequency variations are larger, making the difference equation less effective.

[0075] Please see the results before and after high-pass filtering. Figure 4 The left and right diagrams illustrate that in the range spectrum, each column contains frequency information corresponding to different range units, reflecting the relative position of the radar signal and the target object.

[0076] This invention effectively removes low-frequency components from the signal by performing a high-pass filter operation on the distance unit in the time series, thereby achieving suppression of static targets.

[0077] S23. Perform a velocity-dimensional fast Fourier transform on the high-pass filtered distance spectrum to obtain the filtered Doppler spectrum corresponding to the frequency modulated continuous wave signal.

[0078] The process of performing the velocity-dimensional Fast Fourier Transform is explained in the relevant technical documentation and will not be detailed here. Figure 3 Please see the filtered Doppler spectrum of the frequency modulated continuous wave signal. Figure 5 As shown.

[0079] S3, perform gait discretization processing on the filtered Doppler spectrum to obtain a gait signal scatter plot;

[0080] Specifically, this step may include:

[0081] The energy points below the preset energy threshold in the filtered Doppler spectrum are deleted to obtain a scatter plot of the gait signal.

[0082] To extract key features of gait information, this invention employs a discretization processing method. Since the filtered gait information has high energy, gait features are extracted from the Doppler spectrum by setting an energy threshold. The preset energy threshold can be set as needed. This method, compared to… Figure 5 , generated as Figure 6 The scatter plot of gait signals shown intuitively illustrates the changing patterns of gait information over time.

[0083] S4, perform regression processing and smoothing filtering on the gait signal scatter plot in sequence to obtain the gait feature curve;

[0084] Because the gait signal scatter plot contains a large number of scatter points and rich outlier noise, this invention performs regression processing on the gait signal scatter plot to address this issue.

[0085] In one optional implementation, the regression processing can be implemented using the KNN (K-Nearest Neighbors) algorithm. For the gait signal scatter plot, regression prediction is performed by calculating the minimum Euclidean distance between the data point to be predicted and each point in the cluster, thus achieving a fit to the data. The corresponding calculation formula is as follows:

[0086]

[0087] Where x is the input point to be predicted, p j It is the j-th time dimension data point in the gait signal scatter plot, q j Is with p j The corresponding Doppler dimension target output, where K is the number of nearest neighbors selected. This is the predicted value corresponding to the input point x. M is the number of all samples in the dataset. The first formula in formula (2) means to iterate through all M data points (p j ,q j ), calculate each p j The first formula calculates the distance to the point x to be predicted and finds the point closest to it. The second formula, based on the distance calculation from the first step, selects the K nearest neighbor data points p to the point x to be predicted. j1 p j2 , ..., p jK Their corresponding target values ​​are q j1 q j2 , ..., q jK That is (p jk q jk To perform regression prediction on the target values ​​of these K nearest neighbors, it is necessary to calculate the average of the K targets. It refers to the final predicted value of the point x.

[0088] The results after regression processing are as follows Figure 7 As shown, the regression results exhibit some fluctuations and fail to accurately reflect the smooth trend of gait. To further optimize the smoothness of the KNN regression results, this invention performs a smoothing filter.

[0089] In one alternative implementation, the smoothing filtering can be achieved using a Savitzky-Golay filter, resulting in a smoother and more accurate gait characteristic curve. The relevant formulas for the smoothing filtering are as follows:

[0090]

[0091] i = 1, 2, ..., P

[0092] Among them, y i,smooth This is the data after smoothing and filtering, x i+j The data points on the KNN regression curve h j These are the coefficients of the Butterworth filter, and P is the total number of predicted points generated after KNN regression. This invention applies a smoothing filter with a window of size w to each point in P. For example, when m = 3, the window size for each smoothing process is x. i Centered on x i-1 ,x i ,x i+1 Smoothing is performed. This filter smooths the data within a window of size w, while effectively preserving key variation features in the signal. The gait characteristic curve is shown below. Figure 8 As shown.

[0093] S5, determine the periodic peaks in the gait characteristic curve, obtain the segmentation starting point, and use the segmentation starting point and a fixed time window to perform gait segmentation on the gait characteristic curve to obtain multiple gait sub-spectrums segmented from the frequency-modulated continuous wave signal;

[0094] Specifically, in order to obtain a stable gait spectrum, the present invention can adopt a preset method to first determine the periodic peaks in the gait feature curve, that is, the peaks in the gait feature curve that conform to the periodic characteristics, and use them as the segmentation starting point. Then, with each segmentation starting point as a starting point, the gait feature curve is segmented using a fixed time window, thereby obtaining multiple gait sub-spectrums segmented from the frequency modulated continuous wave signal.

[0095] This invention does not limit the preset method for determining periodic peaks; specific examples will be given later.

[0096] The above steps of this embodiment of the invention illustrate how to segment a gait sub-spectrum from a frequency-modulated continuous wave (FMLC) signal. Following this method, all FMLC signals can have their own gait sub-spectrums segmented. The gait sub-spectrums segmented from all FMLC signals are used to train a preset model, which includes an identity recognition model and a gait anomaly recognition model. Details will be provided later.

[0097] To extract gait information from gait Doppler spectra so that segmentation tasks can focus on gait information, this invention employs a high-pass filter to remove static components from stationary objects in the Doppler spectrum, retaining only dynamic components from moving objects. The gait information with the highest energy is discretized, followed by KNN regression analysis and finally SG smoothing filtering, resulting in gait feature curves that characterize key features of gait changes. Furthermore, periodic analysis is used to obtain the first significant peak point of the gait, thus constructing a gait sample dataset for the recognition task. After obtaining the peak point on the gait feature curve, it is anchored and traced back to the scatter plot position to determine the location of the segmentation point in the Doppler spectrum. A fixed time window length is used for gait spectrum segmentation.

[0098] As can be seen, the gait feature construction method based on wireless signals provided in this invention first extracts and reconstructs gait information, and then effectively segments the Doppler spectrum using the regularity of gait, thereby obtaining gait sub-spectral maps with consistent gait features and similar energy intensities. This overcomes the gait time-varying problem that occurs during Doppler spectrum construction. These sub-spectral maps can be used to train identity recognition models and gait anomaly recognition models, improving the accuracy and robustness of recognition.

[0099] Secondly, embodiments of the present invention also provide a training method for an identity recognition model, such as... Figure 9 As shown, the training method for this identity recognition model may include the following steps:

[0100] Step A1: Obtain the training set for the identity recognition model;

[0101] The training set of the identity recognition model contains training data of multiple users. The training data of each user is multiple gait sub-spectrums segmented from the user's frequency-modulated continuous wave signal. These are obtained according to the gait feature construction method based on wireless signals described in the first aspect. Each gait sub-spectrum is labeled with user information as tag data.

[0102] Step A2: Use the training set of the identity recognition model to train a preset neural network model to obtain the trained identity recognition model.

[0103] For details on the training process of this identity recognition model, please refer to [link / reference]. Figure 10Understanding. The workflow is generally divided into a data preprocessing stage and a gait segmentation stage. In the data preprocessing stage, static objects typically appear as static components with zero velocity and extremely high energy in Doppler spectra, which may suppress the expression of gait signals. As mentioned above, this invention proposes a method to effectively remove static information based on the difference between dynamic and static information. To address the problem of restoring dynamic information, this invention first discretizes the gait features to extract gait information from the spectra. Then, the KNN algorithm and SG filter are used to smooth the discrete information, thereby restoring the gait information. In the gait segmentation stage, based on the extracted gait feature curves, potential gait periods are first found through gait periodicity analysis. Finally, using sliding window techniques and periodic stability rules, the gait Doppler spectrum is segmented to obtain feature-stable gait Doppler sub-spectrums for use in training the identity recognition model.

[0104] Specifically, for step A1, multiple gait sub-spectral maps are obtained from the segmentation of all frequency-modulated continuous wave signals according to steps S1 to S5 of the first aspect. The multiple gait sub-spectral maps segmented from each frequency-modulated continuous wave signal serve as training data for the corresponding user. Each gait sub-spectral map is labeled with user information as tag data. The user information can be the user's name, ID number, or other information that can identify the user. The training data of all users constitute the training set of the identity recognition model.

[0105] Specifically, for the identity recognition model, S5 determines the periodic peaks in the gait feature curve to obtain the segmentation starting point, which may include:

[0106] Step 1), use derivative analysis to differentiate the gait characteristic curve to locate the peak position;

[0107] Step 2): Based on the analysis of the peak position, periodic peaks are obtained as the starting point for segmentation.

[0108] Step 2) may include:

[0109] Based on the peak position, periodic peaks are selected using the peak period comparison method and used as the starting point for segmentation.

[0110] Specifically, in order to obtain a stable gait sub-spectrum, this invention combines derivative analysis and peak period comparison to determine the segmentation starting point, and segments the gait feature curve through a fixed time window.

[0111] Specifically, for step 1), the derivative of the gait characteristic curve is obtained using derivative analysis to initially identify the positions of peaks and troughs. For details on using derivative analysis to obtain the curve's derivative and distinguish peak and trough positions, please refer to relevant technical explanations; they will not be elaborated upon here.

[0112] Regarding step 2), considering the relatively stable periodicity of normal gait, this invention employs a peak period comparison method to eliminate abnormal peaks, thereby determining the segmentation starting point of normal gait. Specifically, each peak is taken as the starting point of a period, and the length consistency of adjacent periods is compared. Only when the current peak's length is similar to the length of several subsequent periods is the peak determined to be the correct segmentation starting point, i.e., a periodic peak. Furthermore, by setting a protection time band and avoiding cross-segmentation, the consistency of gait segmentation and the purity of the data are ensured. The segmentation window is as follows... Figure 11 As shown in the upper middle section of the figure, the starting and ending column indices of the subspectral map are determined, thus completing the gait segmentation. The gait subspectral map obtained after the segmentation operation is shown below. Figure 11 The lower part of the image is shown.

[0113] For step A2, the structure of the preset neural network model can be any existing classification network.

[0114] To implement the gait-based user recognition method, the preset neural network model of this invention can be implemented based on convolutional neural network (CNN) layers, such as... Figure 12 As shown. The segmented gait sub-spectral maps are first processed by convolutional layers for feature extraction, then further compressed by pooling layers, and finally processed by the ReLU activation function before being input into a fully connected layer to obtain a higher-level feature representation. Finally, the data enters the fully connected layer for feature extraction, and the Softmax function is used to output the identity recognition result. For simplicity, Figure 12 The ReLU activation function and the Softmax function are not shown in the diagram.

[0115] in, Figure 12 In the diagram, the convolutional layer (64) indicates that it has 64 output channels, and the fully connected layer (500) indicates that it has 500 neurons. The output of this layer is a one-dimensional vector of shape 500. The following layers follow the same pattern. The pooling layer (2×2) indicates that a window of size (2×2) is used, and each pooling operation downsamples the (2×2) region. The number 10 in the last fully connected layer indicates that the number of categories is 10, that is, the number of user identity categories is 10.

[0116] The preset neural network model is trained on an RTX 4090 GPU and an Intel i9-14900K device. For details on training the preset neural network model using the training set of the aforementioned identity recognition model, please refer to the existing neural network model training process; it will not be described in detail here.

[0117] The training method for the identity recognition model provided in this invention first constructs a training set for the identity recognition model using the proposed gait feature construction method based on wireless signals. Then, a preset neural network model is trained using the training set to obtain the trained identity recognition model. In this process, the invention analyzes the extracted gait feature curves to identify their gait patterns and further determines the potential gait cycle. Then, the Doppler spectrum is segmented using a sliding window technique to construct sample data with stable features and strong consistency, which is ultimately used for model training to achieve user identity recognition. This ensures that the trained identity recognition model, during the identity recognition process, only uses the dynamic components caused by the target gait movement as the intrinsic information of the sample data, thereby improving the accuracy of recognition.

[0118] Thirdly, embodiments of the present invention also provide an identity recognition method, such as... Figure 13 As shown, the identity recognition method includes:

[0119] Step B1: Obtain the gait sub-spectrum of the user to be identified;

[0120] The gait sub-spectrum of the user to be identified can also be obtained in the manner described above, and will not be repeated here.

[0121] Step B2: Input the gait sub-spectrum of the user identity to be identified into the pre-trained identity recognition model to obtain the corresponding user identity information;

[0122] The identity recognition model is obtained according to the training method of the identity recognition model described in the second aspect. For details, please refer to the relevant content above, which will not be repeated here. User identity information can be user name and ID, etc., corresponding to the label data used in training. Obtaining user identity information enables user category identification.

[0123] The identity recognition method provided in this embodiment of the invention is based on a model obtained by training the provided identity recognition model, which can identify user identities and improve the accuracy of recognition.

[0124] Fourthly, embodiments of the present invention provide a training method for a gait anomaly recognition model, such as... Figure 14 As shown, the training method for this gait anomaly recognition model may include the following steps:

[0125] Step C1: Obtain the training set for the gait anomaly recognition model;

[0126] The training set of the gait anomaly recognition model contains training data from multiple users. The training data of each user consists of multiple gait sub-spectral maps segmented from the user's frequency-modulated continuous wave signal. These are obtained according to the gait feature construction method based on wireless signals described in the first aspect. Each gait sub-spectral map is labeled with tag data to characterize whether the gait is abnormal.

[0127] Step C2: Train a preset neural network model using the training set of the gait anomaly recognition model to obtain the trained gait anomaly recognition model.

[0128] For details on the training process of this gait anomaly recognition model, please refer to [link / reference]. Figure 15 Understanding. The workflow is generally divided into a data preprocessing stage and a gait segmentation stage. In the data preprocessing stage, static objects typically appear as static components with zero velocity and extremely high energy in Doppler spectra, which may suppress the expression of gait signals. As mentioned above, this invention proposes a method to effectively remove static information based on the difference between dynamic and static information. To address the problem of restoring dynamic information, this invention first discretizes the gait features to extract gait information from the spectrum. Then, the KNN algorithm and SG filter are used to smooth the discrete information, thereby restoring the gait information. In the gait segmentation stage, periodic peaks are first found based on the extracted gait feature curves. Finally, using the sliding window technique and periodic stability rules, the gait Doppler spectrum is segmented to obtain a feature-stable gait Doppler sub-spectrum for training the gait anomaly recognition model.

[0129] Specifically, for step C1, multiple gait sub-spectral maps are obtained from all frequency-modulated continuous wave signals according to steps S1 to S5 of the first aspect. The multiple gait sub-spectral maps from each frequency-modulated continuous wave signal are used as training data for the corresponding user. Each gait sub-spectral map is labeled with tag data indicating whether it is a normal gait or an abnormal gait. The training data of all users constitute the training set of the gait anomaly recognition model.

[0130] Specifically, abnormal or irregular walking patterns in humans caused by various factors such as neurological disorders, musculoskeletal disorders, post-traumatic recovery, or post-operative rehabilitation are generally referred to as abnormal gait. Types of abnormal gait include spastic gait, scissor gait, straddling gait, staggering gait, and propulsive gait. These abnormal gaits are typically characterized by a walking speed slightly slower than normal gait, and the gait cycle also differs from that of normal gait.

[0131] Considering the slow walking speed of abnormal gait, this invention employs a lower high-pass filter cutoff frequency to retain more gait details. Subsequently, gait discretization, regression, and smoothing filtering are used to further preserve key gait features, resulting in gait feature curves.

[0132] This invention uses hemiplegic gait as an example of abnormal gait. Due to insufficient flexor muscle strength, the lower limb is swung forward by leaning the body forward and using the pelvis, as if drawing circles on the ground. Simultaneously, due to extensor muscle spasms and other reasons, the affected lower limb cannot properly support the weight, resulting in a dragging effect. The gait characteristic curve of hemiplegic gait is shown in the figure below. Figure 16 As shown. Due to the large number of noise peaks in abnormal gait, using the peak period comparison method to find the segmentation starting point will result in segmentation errors.

[0133] To analyze the potential gait cycles of abnormal gait and achieve accurate gait segmentation, this invention utilizes autocorrelation functions to analyze gait cycles. Specifically, for the gait anomaly identification model, step S5 determines the periodic peaks in the gait feature curve to obtain the segmentation starting point, which may include:

[0134] Step 1), use derivative analysis to differentiate the gait characteristic curve to locate the peak position;

[0135] Step 2): Based on the analysis of the peak position, periodic peaks are obtained as the starting point for segmentation.

[0136] Step 2) may include:

[0137] Autocorrelation analysis is performed on the gait feature curve to determine the correct period. The correct period is then used to select periodic peaks from the peak positions as the segmentation starting point.

[0138] Specifically, the periodicity of the gait characteristic curve is analyzed using the autocorrelation method with unbiased estimation, employing the following formula:

[0139]

[0140] Where R(τ) represents the autocorrelation value of the signal x(t) at a lag time τ, and T is the total number of sample points on the gait characteristic curve, where x(t) is the number of sample points on the gait characteristic curve (after filtering). i,smooth The data is analyzed using the autocorrelation result as input. This ensures that the results are not affected by the signal length. Then, the positive lag component is extracted from the autocorrelation result, as this part of the data best reflects the regularity of the signal over time.

[0141] Autocorrelation analysis of gait signals, such as Figure 17 As shown in the figure, Figure a shows the autocorrelation analysis of normal gait, which shows that the gait cycle has a strong autocorrelation and a short average cycle length; Figure b shows the autocorrelation analysis of hemiplegic gait, which shows a low gait autocorrelation and a long cycle length.

[0142] The peak value of the autocorrelation function is an important indicator of periodicity, and the lag time of the peak value corresponds to the potential period of the signal. To locate the segmentation point of the gait characteristic curve of abnormal gait, this invention first directly uses autocorrelation analysis to analyze the gait characteristic curve to find the correct period in the abnormal gait. The correct period is then used to identify periodic peaks, which serve as the starting point for segmentation.

[0143] Specifically, the lag time from autocorrelation analysis directly indicates the periodicity of the signal, helping to determine the gait cycle length and the location of peak points. By calculating the time difference between the lag time and nearby peak points, the peak point with the smallest time difference is set as the peak point of the correct cycle. This invention then marks the peak points of the correct cycle on the gait characteristic curve as the starting points for gait segmentation. A gait segmentation algorithm is then used to segment the Doppler spectrum of abnormal gaits, thereby obtaining gait sub-spectrums for model training. A comparison of the gait segmentation results is shown in the figure below. Figure 18 As shown, (a) is the normal gait spectrum, (b) is the hemiplegic gait spectrum, (c) is the sub-spectrum obtained after segmentation of the normal gait, and (d) is the gait sub-spectrum segmented after periodic analysis. The segmented gait sub-spectrums are used to construct the training set for the gait anomaly recognition model.

[0144] Because the potential gait cycle lengths differ between normal and abnormal gait, the cycle length of each step in normal gait is usually quite similar; however, the cycle of abnormal gait is unstable, or its cycle length differs from that of normal gait. This invention uses an autocorrelation function to perform autocorrelation analysis on gait characteristic curves, calculating the cycle length between adjacent peaks to evaluate the stability of the signal cycle. By using the autocorrelation analysis function for gait cycle analysis and employing a gait segmentation algorithm to segment the abnormal gait spectrum, abnormal gait sample data can be constructed for model recognition.

[0145] For step B2, the structure of the preset neural network model can be any existing classification network. This invention can achieve the classification task of abnormal gait using a CNN network. The structure of the CNN network is as follows... Figure 19 As shown. Figure 19 and Figure 12 The difference lies in the fact that the last fully connected layer has 2 classifications, which means binary classification, either normal gait or abnormal gait.

[0146] The network model was trained using an RTX 4090 GPU and an Intel i9-14900K device. For details on training the preset neural network model using the training set of the gait anomaly recognition model, please refer to the existing neural network model training process; it will not be described in detail here.

[0147] The training method for the gait anomaly recognition model provided in this invention first constructs a training set for the gait anomaly recognition model using the proposed gait feature construction method based on wireless signals. Then, a preset neural network model is trained using the training set to obtain the trained gait anomaly recognition model. In this process, the invention analyzes the extracted gait feature curves to identify gait patterns and further determines potential gait cycles. Then, the Doppler spectrum is segmented using a sliding window technique to construct stable and consistent sample data, which is ultimately used for model training to achieve gait anomaly recognition. This ensures that the trained gait anomaly recognition model, during gait recognition, only uses the dynamic components caused by the target gait movement as intrinsic information of the sample data, thereby improving the accuracy of recognition.

[0148] Fifthly, embodiments of the present invention also provide a method for gait anomaly recognition, such as... Figure 20 As shown, the gait anomaly identification method includes:

[0149] Step D1: Obtain the gait sub-spectrum map of the gait to be identified;

[0150] The gait sub-spectrum map to be identified can also be obtained in the manner described above, and will not be repeated here.

[0151] Step D2: Input the gait sub-spectrum map to be identified into the pre-trained gait anomaly identification model to obtain the gait identification result, which includes normal gait or abnormal gait.

[0152] The gait anomaly recognition model is obtained according to the training method of the gait anomaly recognition model described in the fourth aspect. For details, please refer to the relevant content above, which will not be repeated here.

[0153] The gait anomaly recognition method provided in this embodiment of the invention, based on the model implementation obtained by the training method of the provided gait anomaly recognition model, can identify whether a user's gait is abnormal, thereby improving the accuracy of recognition.

[0154] It should be noted that, in the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0155] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A method for constructing gait features based on wireless signals, characterized in that, include: Acquire frequency-modulated continuous wave signals from multiple users collected by radar transceiver equipment; For each frequency-modulated continuous wave signal, a filtered Doppler spectrum is obtained based on high-pass filtering. The energy points below the preset energy threshold in the filtered Doppler spectrum are deleted to remove the static components generated by stationary objects in the filtered Doppler spectrum, and only the dynamic components generated by moving objects are retained to obtain a gait signal scatter plot. The gait signal scatter plot is subjected to regression processing and smoothing filtering in sequence to obtain gait feature curves that characterize the key features of gait changes. The peak position is located by differentiating the gait characteristic curve using derivative analysis. Based on the analysis of the peak positions, periodic peaks are obtained as segmentation starting points. Each segmentation starting point is used as a starting point, and the gait characteristic curve is segmented using a fixed time window to obtain multiple gait sub-spectral maps segmented from the frequency-modulated continuous wave signal. All gait sub-spectral maps segmented from the frequency-modulated continuous wave signal are used to train a preset model, which includes an identity recognition model and a gait anomaly recognition model. When the preset model is the identity recognition model, obtaining periodic peaks as segmentation starting points based on the peak positions includes: using a peak-period comparison method to select periodic peaks as segmentation starting points based on the peak positions. When the preset model is the gait anomaly recognition model, obtaining periodic peaks as segmentation starting points based on the peak positions includes: performing autocorrelation analysis on the gait characteristic curve to determine the correct period, and using the correct period to select periodic peaks from the peak positions as segmentation starting points.

2. The gait feature construction method based on wireless signals according to claim 1, characterized in that, For each frequency-modulated continuous wave signal, based on high-pass filtering, the filtered Doppler spectrum corresponding to that frequency-modulated continuous wave signal is obtained, including: For each frequency-modulated continuous wave signal, perform a range-dimensional fast Fourier transform to obtain the range spectrum of the frequency-modulated continuous wave signal; The distance spectrum is then subjected to high-pass filtering. The high-pass filtered distance spectrum is subjected to a velocity-dimensional fast Fourier transform to obtain the filtered Doppler spectrum corresponding to the frequency-modulated continuous wave signal.

3. A training method for an identity recognition model, characterized in that, include: Obtain the training set of the identity recognition model; wherein the training set of the identity recognition model contains training data of multiple users, and the training data of each user is multiple gait sub-spectral maps segmented from the frequency-modulated continuous wave signal of that user, which are obtained by the gait feature construction method based on wireless signal according to claim 1 or 2, and each gait sub-spectral map is marked with user information as label data. A preset neural network model is trained using the training set of the identity recognition model to obtain the trained identity recognition model.

4. An identity recognition method, characterized in that, include: Obtain the gait sub-spectrum of the user to be identified; The gait sub-spectrum of the user to be identified is input into a pre-trained identity recognition model to obtain the corresponding user identity information; wherein, the identity recognition model is obtained by the training method of the identity recognition model according to claim 3.

5. A training method for a gait anomaly recognition model, characterized in that, include: Obtain the training set of the gait anomaly recognition model; wherein the training set of the gait anomaly recognition model contains training data of multiple users, and the training data of each user is multiple gait sub-spectral maps segmented from the frequency-modulated continuous wave signal of that user, which are obtained by the gait feature construction method based on wireless signal according to claim 1 or 2, and each gait sub-spectral map is marked with label data to characterize whether the gait is abnormal. A preset neural network model is trained using the training set of the gait anomaly recognition model to obtain the trained gait anomaly recognition model.

6. A method for gait anomaly identification, characterized in that, include: Obtain the gait sub-spectrum map to be identified; The gait sub-spectrum to be identified is input into a pre-trained gait anomaly identification model to obtain gait identification results, which include normal gait or abnormal gait; wherein the gait anomaly identification model is obtained by the training method of the gait anomaly identification model according to claim 5.

Images