Feature processing method for millimeter wave radar gesture recognition

Abstract

The invention belongs to the technical field of intelligent wireless sensing and radar signal processing, and particularly relates to a millimeter wave radar gesture recognition feature processing method, which is particularly suitable for scenes with environment interference (such as walking of others and static clutter), and specifically comprises the following steps: S1, preliminary filtering by a self-adaptive filter; s2, carrying out improved I CEEMDAN decomposition; s3, I MF component classification and processing; s4, signal reconstruction; the experimental result shows that the method provides a robust and efficient solution for gesture recognition, and can be widely applied to the fields of man-machine interaction, virtual reality, intelligent equipment control and the like.

Description

Technical Field

[0001] This invention belongs to the field of radar signal processing and pattern recognition technology, and specifically relates to a feature processing method for millimeter-wave radar gesture recognition. Background Technology

[0002] Gesture recognition, as a crucial component of human-computer interaction, has broad application prospects in fields such as gaming, healthcare, smart homes, and autonomous driving. Traditional gesture recognition technologies primarily rely on wearable devices (such as accelerometers and gyroscopes) or optical sensors (such as cameras). Wearable devices suffer from comfort and battery life issues, while optical sensors are severely affected by lighting conditions and pose privacy risks. Millimeter-wave radar, unaffected by lighting conditions, protects user privacy, possesses strong penetration capabilities, and can be installed internally in devices, making it an ideal choice for gesture recognition. Millimeter-wave radar can acquire multi-dimensional information such as target distance, velocity, angle, and micro-Doppler, exhibiting significant advantages, especially in dynamic gesture recognition. However, the echo signals acquired by millimeter-wave radar are susceptible to environmental noise and clutter interference, affecting the accuracy of feature extraction. Traditional denoising methods have limitations when processing nonlinear and non-stationary radar signals; therefore, more advanced signal processing techniques are needed to improve the accuracy and robustness of gesture recognition. In recent years, Empirical Mode Decomposition (EMD) and its improved algorithms (such as CEEMDAN and I CEEMDAN) combined with wavelet thresholding (WT) denoising have shown good performance in the field of signal processing. These methods can effectively process non-stationary signals and adapt to different noise characteristics. Summary of the Invention

[0003] (a) Technical problems to be solved

[0004] To address the shortcomings of existing technologies, this invention provides a feature processing method for millimeter-wave radar gesture recognition, which solves the problems mentioned in the background section.

[0005] (II) Technical Solution

[0006] To achieve the above objectives, the present invention specifically adopts the following technical solution:

[0007] A feature processing method for millimeter-wave radar gesture recognition specifically includes the following steps;

[0008] S1: Obtain the original echo signal through millimeter-wave radar, and use an adaptive filter to perform preliminary filtering on the original echo signal to obtain a preliminary filtered signal;

[0009] S2: The improved adaptive noise complete set empirical mode decomposition (ICEEMDAN) algorithm is used to decompose the preliminary filtered signal into multiple intrinsic mode function (IMF) components; the sample entropy of each IMF component and its correlation coefficient with the original echo signal are calculated, and the IMF components are classified into noise-dominant components and signal-dominant components based on the calculation results.

[0010] S3: Perform wavelet thresholding (WT) denoising on the noise-dominant IMF components after classification, and retain the signal-dominant IMF components;

[0011] S4: Reconstruct the noise-dominant IMF component and the retained signal-dominant IMF component after denoising to obtain the denoised radar signal;

[0012] S5: Perform multi-dimensional feature extraction on the denoised radar signal to obtain feature data for gesture recognition;

[0013] Furthermore, in S1, the original echo signal is acquired from the millimeter-wave radar, and environmental noise and interference are initially suppressed: the filter calculates the current error e(n) and, in conjunction with the current input signal x(n), adjusts its own weight w(n) according to the learning rate determined by μ, with the goal of making the error smaller in the next iteration.

[0014] w(n+1) = w(n) + μ·e(n)·x(n)

[0015] Where: w(n) and w(n+1) represent the filter weight vectors at the current time and the next time, respectively; μ determines the magnitude of each weight adjustment; and e(n) is the error signal, which indicates the difference between the current filter performance and the ideal situation.

[0016] Furthermore, S2, ICEEMDAN, is an improvement on the CEEMDAN algorithm. Its core purpose is to decompose the signal more effectively, reduce mode aliasing, and improve the signal-to-noise ratio of the final signal. A shortcoming of traditional CEEMDAN is that it directly adds Gaussian white noise in each decomposition process. While effective, the added noise has uniform energy across all frequency bands, which may introduce unwanted interference or lead to residual noise in some cases.

[0017] ICEEMDAN's improvement strategy: The key improvement of ICEEMDAN is that it no longer adds the original Gaussian white noise, but adds a specific IMF component of the white noise after EMD decomposition (usually the Kth order IMF, where K is a preset order, such as the 1st or 2nd order).

[0018] Mathematical expression: Let the Gaussian white noise be n(t). After EMD decomposition, a series of MF components are obtained. ICEEMDAN selects the Kth order component As added noise, adding IMF components of a specific order is equivalent to adding noise with a specific center frequency and bandwidth. This makes the resulting IMF more robust and better able to separate modes at different scales, thereby significantly reducing mode aliasing.

[0019] For each IMF component generated by the decomposition: imf_i(t), it is necessary to determine whether it is noise-dominated or signal-dominated. Sample entropy is used to measure the complexity and irregularity of a time series. The higher the entropy value, the more complex and unpredictable the sequence.

[0020]

[0021] Where: m is the embedding dimension, i.e. the length of the constructed template vector; r is the similarity tolerance, i.e. the distance threshold for judging whether two vectors are "similar"; N is the length of the data sequence; B is the number of template matches in the sequence where the distance between vectors of length m is less than r; and A is the number of template matches in the same sequence where the distance between vectors of length m+1 is less than r.

[0022] The correlation coefficient is used to measure the correlation between two signals (here referring to the IMF component). i The degree of linear correlation between (t) and the original signal x(t):

[0023]

[0024] Where: x(t) is the original signal, imf i (t) is the i-th IMF component; and These are the mean values ​​of the original signal and the IMF component, respectively; N is the signal length.

[0025] Furthermore, in step S3, the noise-dominant IMF component is refined to remove noise while retaining the signal-dominant component. The noise IMF is decomposed into approximation coefficients (low frequency) and detail coefficients (high frequency). Thresholding is performed using an adaptive threshold.

[0026]

[0027] Where σ is the noise standard deviation (estimated by: σ = median(|c|) / 0.6745, where c is the detail coefficient), and N is the signal length;

[0028] Furthermore, in step S4, the denoised noise IMF is superimposed with the retained signal-dominant IMF to restore the clean signal.

[0029]

[0030] Where: c i (t) represents the signal-dominant IMF, T(c) j )(t) represents the noise IMF after denoising, and r(t) represents the residual component;

[0031] Further, in step S5, multi-dimensional features are extracted from the denoised signal and input into a deep learning model for classification, generating time-frequency feature maps: Distance-Time Map (RTM): extracts the time-varying features of the signal in the distance dimension through Short Time Fourier Transform (STFT) or Continuous Wavelet Transform (CWT); Doppler-Time Map (DTM): calculates the speed change over time based on the Doppler effect; Angle-Time Map (ATM / ETM): estimates the horizontal azimuth and pitch angles through phase difference.

[0032]

[0033] Where: f d Let f be the Doppler frequency shift, v be the target velocity, and f be the frequency shift. c Where c is the radar carrier frequency and c is the speed of light;

[0034] Concatenate the individual feature maps:

[0035]

[0036] Where: RTM, DTM, ATM, and ETM are four feature tensors to be fused, typically representing features from different modalities or sources; N is the batch size, the number of data samples processed at one time. All input tensors must have the same N, C r C d C a C e These correspond to the number of channels, height, and width of the feature maps for the four feature tensors: RTM, DTM, ATM, and ETM. H and W represent the input tensors, and Concat(·) is the concatenation operation. F... fused For the fused hybrid feature tensor, C total This represents the total number of channels in the merged tensor.

[0037] The obtained features are then used for a classification task.

[0038] (III) Beneficial Effects

[0039] Compared with existing technologies, the present invention provides a feature processing method for millimeter-wave radar gesture recognition, which has the following beneficial effects:

[0040] This invention proposes a feature processing method for millimeter-wave radar gesture recognition. The improved ICEEMDAN algorithm generates cleaner IMF components with less modal aliasing. The proposed IMF classification method based on sample entropy and correlation coefficients can more accurately separate noise-dominated IMFs. Subsequent processing (such as wavelet thresholding) on ​​these filtered noisy IMFs can significantly reduce the influence of noise components while preserving useful signal features to the greatest extent. The resulting reconstructed signal therefore has a higher signal-to-noise ratio and is closer to the real signal.

[0041] This invention proposes a feature processing method for millimeter-wave radar gesture recognition. It dynamically determines the number of decomposition layers based on the signal's local features (such as instantaneous frequency complexity), representing a significant improvement. Traditional EMD and its variants typically require pre-setting the number of decomposition layers or iterations. For complex and variable real-world signals, a fixed number of decomposition layers may lead to insufficient decomposition of certain details (under-decomposition) or the generation of numerous meaningless spurious components (over-decomposition). The adaptive decomposition strategy enables the algorithm to "adapt to the signal," automatically adjusting to the most suitable decomposition depth, thereby better adapting to various complex signals and improving the algorithm's generalization ability. Attached Figure Description

[0042] Figure 1 This is a flowchart of the present invention;

[0043] Figure 2 This is a diagram illustrating the result of feature fusion in this invention; Detailed Implementation

[0044] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Example

[0046] like Figure 1 As shown, an embodiment of the present invention proposes a feature processing method for millimeter-wave radar gesture recognition, which includes the following steps:

[0047] S1: Obtain the raw echo signal from the millimeter-wave radar and initially suppress environmental noise and interference: The filter calculates the current error e(n) and combines it with the current input signal x(n) according to the learning speed determined by μ, and adjusts its own weight w(n) to make the error smaller in the next iteration.

[0048] w(n+1) = w(n) + μ·e(n)·x(n)

[0049] Where: w(n) and w(n+1) represent the filter weight vectors at the current time and the next time, respectively; μ determines the magnitude of each weight adjustment; and e(n) is the error signal, which indicates the difference between the current filter performance and the ideal situation.

[0050] S2: ICEEMDAN is an improvement on the CEEMDAN algorithm. Its core purpose is to decompose signals more effectively, reduce mode aliasing, and improve the signal-to-noise ratio of the final signal. The shortcomings of traditional CEEMDAN: CEEMDAN directly adds Gaussian white noise in each decomposition process. While effective, the added noise has uniform energy across all frequency bands, which may introduce unwanted interference or lead to residual noise in some cases.

[0051] ICEEMDAN's improvement strategy: The key improvement of ICEEMDAN is that it no longer adds the original Gaussian white noise, but adds a specific IMF component of the white noise after EMD decomposition (usually the Kth order IMF, where K is a preset order, such as the 1st or 2nd order).

[0052] Mathematical expression: Let the Gaussian white noise be n(t). After EMD decomposition, a series of MF components are obtained. ICEEMDAN selects the Kth order component As added noise, adding IMF components of a specific order is equivalent to adding noise with a specific center frequency and bandwidth. This makes the resulting IMF more robust and better able to separate modes at different scales, thereby significantly reducing mode aliasing.

[0053] For each IMF component generated by the decomposition:imf i (t), it is necessary to determine whether it is noise-driven or signal-driven: Sample entropy is used to measure the complexity and irregularity of a time series. The higher the entropy value, the more complex and unpredictable the sequence.

[0054]

[0055] Where: m is the embedding dimension, i.e. the length of the constructed template vector; r is the similarity tolerance, i.e. the distance threshold for judging whether two vectors are "similar"; N is the length of the data sequence; B is the number of template matches in the sequence where the distance between vectors of length m is less than r; and A is the number of template matches in the same sequence where the distance between vectors of length m+1 is less than r.

[0056] The correlation coefficient is used to measure the correlation between two signals (here referring to the IMF component). i The degree of linear correlation between (t) and the original signal x(t):

[0057]

[0058] Where: x(t) is the original signal, imf i (t) is the i-th IMF component; and These are the mean values ​​of the original signal and the IMF component, respectively; N is the signal length.

[0059] S3: Refine the noise-dominant IMF component to retain the signal-dominant component, perform multi-scale decomposition on the noise IMF to obtain approximation coefficients (low frequency) and detail coefficients (high frequency); threshold processing is performed using an adaptive threshold.

[0060]

[0061] Where σ is the noise standard deviation (estimated by: σ = median(|c|) / 0.6745, where c is the detail coefficient), and N is the signal length;

[0062] S4: Superimpose the denoised noise IMF with the preserved signal-dominant IMF to restore the clean signal.

[0063]

[0064] Where: c i (t) represents the signal-dominant IMF, T(c) j )(t) represents the noise IMF after denoising, and r(t) represents the residual component;

[0065] S5: Extract multi-dimensional features from the denoised signal and input them into a deep learning model for classification, generating time-frequency feature maps: Distance-Time Map (RTM): Extract time-varying features of the signal in the distance dimension through Short Time Fourier Transform (STFT) or Continuous Wavelet Transform (CWT); Doppler-Time Map (DTM): Based on the Doppler effect, the calculation speed varies with time; Angle-Time Map (ATM / ETM): Estimate the horizontal azimuth and pitch angles through phase difference.

[0066]

[0067] Where: f d Let f be the Doppler frequency shift, v be the target velocity, and f be the frequency shift. c Where c is the radar carrier frequency and c is the speed of light;

[0068] Concatenate the feature maps:

[0069]

[0070] Where: RTM, DTM, ATM, and ETM are four feature tensors to be fused, typically representing features from different modalities or sources; N is the batch size, the number of data samples processed at one time. All input tensors must have the same N, C r C d C a C e These correspond to the number of channels, height, and width of the feature maps for the four feature tensors: RTM, DTM, ATM, and ETM. H and W represent the input tensors, and Concat(·) is the concatenation operation. F... fused For the fused hybrid feature tensor, C total This represents the total number of channels in the merged tensor.

[0071] The obtained features are then used for a classification task.

[0072] like Figure 1 The diagram shown is a flowchart of the present invention. Through the processes S1-S5, the signal is denoised and then the features are fused.

[0073] like Figure 2 The image shown is a result of feature fusion according to the present invention, which combines four features, with color depth representing energy.

[0074] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A feature processing method for millimeter-wave radar gesture recognition, characterized in that, Includes the following steps: S1: Obtain the original echo signal through millimeter-wave radar, and use an adaptive filter to perform preliminary filtering on the original echo signal to obtain a preliminary filtered signal; S2: The improved adaptive noise complete set empirical mode decomposition (ICEEMDAN) algorithm is used to decompose the preliminary filtered signal into multiple intrinsic mode function (IMF) components; the sample entropy of each IMF component and its correlation coefficient with the original echo signal are calculated, and the IMF components are classified into noise-dominant components and signal-dominant components based on the calculation results. S3: Perform wavelet thresholding (WT) denoising on the noise-dominant IMF components after classification, and retain the signal-dominant IMF components; S4: Reconstruct the noise-dominant IMF component and the retained signal-dominant IMF component after denoising to obtain the denoised radar signal; S5: Perform multi-dimensional feature extraction on the denoised radar signal to obtain feature data for gesture recognition; The feature data is input into the gesture recognition model, and the gesture recognition result is output.

2. The method according to claim 1, characterized in that, In step S1, the adaptive filter is a minimum mean square (LMS) adaptive filter. The raw echo signal is acquired from millimeter-wave radar, and environmental noise and interference are initially suppressed: the filter calculates the current error e(n) and, in conjunction with the current input signal x(n), adjusts its own weight w(n) according to the learning rate determined by μ, with the goal of minimizing the error in the next iteration. w(n+1) = w(n) + μ·e(n)·x(n) Where w(n) and w(n+1) represent the filter weight vectors at the current and next time steps, respectively, μ determines the magnitude of each weight adjustment, and e(n) is the error signal, which indicates the difference between the current filter performance and the ideal situation.

3. The method according to claim 1, characterized in that, In step S2, the improvement of the improved CEEMDAN algorithm lies in: adaptively determining the number of decomposition layers based on the local characteristics of the preliminary filtered signal, and using a local mean operator to calculate the upper and lower envelope mean values ​​of the signal. CEEMDAN is an improvement on the CEEMDAN algorithm, and its core purpose is to decompose the signal more effectively, reduce mode aliasing, and improve the signal-to-noise ratio of the final signal. A shortcoming of the traditional CEEMDAN algorithm is that it directly adds Gaussian white noise in each decomposition process. While effective, the added noise has uniform energy across all frequency bands, which may introduce unnecessary interference or lead to residual noise in some cases. ICEEMDAN's improvement strategy: The key improvement of ICEEMDAN is that it no longer adds the original Gaussian white noise, but adds a specific IMF component of the white noise after EMD decomposition (usually the Kth order IMF, where K is a preset order, such as the 1st or 2nd order). Mathematical expression: Let the Gaussian white noise be n(t). After EMD decomposition, a series of MF components are obtained. ICEEMDAN selects the Kth order component As added noise, adding IMF components of a specific order is equivalent to adding noise with a specific center frequency and bandwidth. This makes the resulting IMF more robust and better able to separate modes at different scales, thereby significantly reducing mode aliasing. For each IMF component generated by the decomposition: imf_i(t), it is necessary to determine whether it is noise-dominated or signal-dominated. Sample entropy is used to measure the complexity and irregularity of a time series. The higher the entropy value, the more complex and unpredictable the sequence. Where: m is the embedding dimension, i.e. the length of the constructed template vector; r is the similarity tolerance, i.e. the distance threshold for judging whether two vectors are "similar"; N is the length of the data sequence; B is the number of template matches in the sequence where the distance between vectors of length m is less than r; and A is the number of template matches in the same sequence where the distance between vectors of length m+1 is less than r. The correlation coefficient is used to measure the correlation between two signals (here referring to the IMF component). i The degree of linear correlation between (t) and the original signal x(t): Where: x(t) is the original signal, imf i (t) is the i-th IMF component; and These are the mean values ​​of the original signal and the IMF component, respectively; N is the signal length.

4. The method according to claim 1, characterized in that, In step S3, the specific criteria for classifying IMF components into noise-dominant components and signal-dominant components based on the calculation results are as follows: IMF components with sample entropy values ​​higher than a set first threshold and correlation coefficients with the original echo signal lower than a set second threshold are classified as noise-dominant components; IMF components with sample entropy values ​​lower than a set first threshold and correlation coefficients with the original echo signal higher than a set second threshold are classified as signal-dominant components. The noise-dominant IMF components undergo refined denoising, while the signal-dominant components are retained. The noise IMFs are then decomposed into approximate coefficients (low-frequency) and detail coefficients (high-frequency); threshold processing is performed using an adaptive threshold. Where σ is the noise standard deviation (estimated by: σ = median(|c|) / 0.6745, where c is the detail coefficient), and N is the signal length.

5. The method according to claim 1, characterized in that, In step S4, the wavelet basis used in the wavelet thresholding denoising process is the Symlet wavelet, and the threshold function used is a modified semi-soft threshold function. The denoised noise IMF is superimposed with the retained signal dominant IMF to restore the clean signal. Where: c i (t) represents the signal-dominant IMF, T(c) j )(t) represents the noise IMF after denoising, and r(t) represents the residual component.

6. The method according to claim 1, characterized in that, In step S5, the multi-dimensional feature extraction includes generating one or more of the following: distance-time map (RTM), Doppler-time map (DTM), horizontal angle-time map (ATM), and pitch angle-time map (ETM). The generated feature maps are then fused to extract multi-dimensional features from the denoised signal. These features are then input into a deep learning model for classification, generating time-frequency feature maps: Distance-time map (RTM): extracts the time-varying features of the signal in the distance dimension using short-time Fourier transform (STFT) or continuous wavelet transform (CWT); Doppler-time map (DTM): calculates the speed change over time based on the Doppler effect; Angle-time map (ATM / ETM): estimates the horizontal azimuth and pitch angles using phase difference. Where: f d Let f be the Doppler frequency shift, v be the target velocity, and f be the frequency shift. c Where c is the radar carrier frequency and c is the speed of light; Concatenate the individual feature maps: Where: RTM, DTM, ATM, and ETM are four feature tensors to be fused, typically representing features from different modalities or sources; N is the batch size, the number of data samples processed at one time. All input tensors must have the same N, C r C d C a C e These correspond to the number of channels, height, and width of the feature maps for the four feature tensors: RTM, DTM, ATM, and ETM. H and W represent the input tensors, and Concat(·) is the concatenation operation. F... fused For the fused hybrid feature tensor, C total This represents the total number of channels in the merged tensor. The obtained features are then used for a classification task.

7. The method according to claim 1, characterized in that, The step S5 is followed by: standardizing the fused features to eliminate the influence of dimensions.

8. The method according to claim 1, characterized in that, In step S5, the gesture recognition model is a CNN-LSTM-Attention hybrid model; the gesture recognition model processes and classifies the input feature data through an attention mechanism.

Images