Three-dimensional channel reconstruction enhancement method for moving target radar signal

Abstract

The invention belongs to the technical field of radars, and relates to a three-dimensional channel reconstruction enhancement method for a moving target radar signal, which comprises the following steps: 1) acquiring an original radar signal; 2) performing time-frequency conversion on the original radar signal; 3) performing moving target detection on the radar signal obtained in the step 2); 4) performing intra-frame zero mean value processing on the radar signal obtained in the step 3); 5) performing global frame standardization processing on the radar signal obtained in the step 4); 6) performing multi-channel signal filtering processing on the radar signal obtained in the step 5); and 7) a result obtained in the step 6) is injected into the neural network for deep learning, and enhancement processing of the radar signal is completed. The invention provides the three-dimensional channel reconstruction enhancement method for the moving target radar signal, which can effectively avoid noise interference.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar, and relates to a method for enhancing radar signals, and particularly to a three-dimensional channel reconstruction and enhancement method for moving target radar signals. Background Art

[0002] Radar signal enhancement technology is an important component of radar technology. Radar sensors have the advantages of protecting user privacy and being unaffected by light. However, the high abstraction of radar signals themselves and the sparsity of features make it difficult to directly characterize target features from reflected electromagnetic waves, which is not conducive to further processing of various subsequent algorithms.

[0003] Although many current studies perform enhancement preprocessing on the original radar signals, the effects are not good. For example, moving average filtering performs moving window averaging on time-domain signals to suppress high-frequency noise, but it will blur transient features (such as short-time pulses of targets) and reduce time-domain resolution. For example, Wiener filtering is based on the statistical characteristics of signals and noise to minimize the mean square error, but it requires a priori noise models and its performance degrades in scenarios of non-stationary noise (such as burst interference) of millimeter-wave radars. For example, median filtering replaces the current sampled value with the median in a moving window to suppress impulse noise, but it has poor effects on Gaussian noise and the fixed window size leads to insufficient adaptability. Another example is wavelet transform, which decomposes signals at multiple scales and denoises in different frequency bands. The selection of wavelet bases depends on experience, and the enhancement of medium and high-frequency components is insufficient, which is very disadvantageous in radar moving target detection. Another example is band-pass filtering, which can retain the target feature frequency band (such as the typical bands of 24 GHz / 77 GHz for millimeter-wave radars) and suppress out-of-band noise, but it requires precise predefined target frequency bands and may lose useful high-frequency details (such as micro-Doppler features). Although these traditional methods can well enhance the spectral information of radars, they are not aimed at the radar moving target detection task and have deficiencies in performance or computational efficiency. Summary of the Invention

[0004] In order to solve the above technical problems existing in the background art, the present invention provides a three-dimensional channel reconstruction and enhancement method for moving target radar signals that can effectively avoid being interfered by noise.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A three-dimensional channel reconstruction and enhancement method for moving target radar signals, characterized in that: the three-dimensional channel reconstruction and enhancement method for moving target radar signals includes the following steps:

[0007] 1) Obtain the original radar signal;

[0008] 2) Perform time-frequency transformation on the original radar signal obtained in step 1);

[0009] 3) Perform moving target detection on the radar signal obtained in step 2);

[0010] 4) Perform in-frame zero-mean processing on the radar signal obtained in step 3);

[0011] 5) Perform global frame normalization processing on the radar signal obtained in step 4);

[0012] 6) Perform multi-channel signal filtering processing on the radar signal obtained in step 5);

[0013] 7) Inject the result obtained in step 6) into a neural network for deep learning to complete the enhancement processing of the radar signal.

[0014] Preferably, the specific implementation manner of step 6) is:

[0015] 6.1) Copy the radar signal obtained in step 5) and divide it into three-way signals;

[0016] 6.2) Perform Log filtering processing on the first-way signal obtained in step 6.1) to obtain the signal after Log filtering processing; perform Sin filtering processing on the second-way signal obtained in step 6.1) to obtain the signal after Sin filtering processing; perform Gaussian filtering processing on the third-way signal obtained in step 6.1) to obtain the signal after Gaussian filtering processing.

[0017] Preferably, the specific implementation manner of performing Log filtering processing in step 6.2) is:

[0018]

[0019] Where:

[0020] I is the radar signal obtained in the input step 5);

[0021] x is the index of the Doppler dimension unit corresponding to I;

[0022] f is the mapping function;

[0023] N is the number of Doppler dimension units per frame;

[0024] A is the scaling coefficient;

[0025] 1 is a constant to avoid negative scaling;

[0026] I log is the output signal of I after being scaled by a specific frequency, that is, the signal after Log filtering processing;

[0027] The specific implementation manner of performing Sin filtering processing in step 6.2) is:

[0028]

[0029] Wherein:

[0030] I is the radar signal obtained in step 5) of the input;

[0031] x is the index of the Doppler dimension unit corresponding to I;

[0032] g is the mapping function;

[0033] N is the number of Doppler dimension units per frame;

[0034] B is the scaling coefficient;

[0035] I sin is the output signal of I after being scaled by a specific frequency, that is, the signal after being processed by Sin filtering;

[0036] The specific implementation method of performing Gaussian filtering in step 6.2) is:

[0037]

[0038] Wherein:

[0039] I(x,y) is the radar signal obtained in step 5);

[0040] C is the scaling coefficient;

[0041] I Gauss (x,y) is the two-dimensional image after being processed by Gaussian filtering, that is, the signal after being processed by Gaussian filtering;

[0042] x is the index of the Doppler dimension unit corresponding to I;

[0043] y is the index of the range dimension unit corresponding to I;

[0044] k is 1 / 2 of the kernel size;

[0045] G(i,j) is the signal intensity of the Gaussian filter kernel at the coordinate (i,j), and the expression of G(i,j) is:

[0046] G(i,j) = G(i)·G(j)

[0047] Wherein:

[0048]

[0049] Wherein:

[0050] i and j respectively represent the abscissa and ordinate in the pixel image;

[0051] σ is the standard deviation parameter of the Gaussian kernel, which is selected according to the signal.

[0052] Preferably, the specific implementation of step 2) is as follows:

[0053] 2.1) Convert the original radar signal obtained in step 1) into a range-velocity two-dimensional spectrogram with clear physical meaning;

[0054] 2.2) Add time dimension information to the two-dimensional spectrogram obtained in step 2.1), convert the obtained radar electromagnetic wave signal into a three-dimensional representation of range-Doppler-time, and obtain a time-continuous micro-Doppler spectrum to complete the time-frequency conversion of the original radar signal.

[0055] Preferably, the moving target detection in step 3) is completed by using the adjacent frame difference method. Specifically, the adjacent frame difference method replaces the last frame of the continuous time series with its previous frame, keeping the length the same as the corresponding segment of the visual modality;

[0056] y n = x n - x n-1

[0057] Where:

[0058] x n is the radar signal of the nth frame;

[0059] x n-1 is the radar signal of the (n - 1)th frame;

[0060] y n is the radar signal generated after the difference.

[0061] Preferably, the specific implementation of step 4) is as follows:

[0062] 4.1) Obtain each fast-time sampling point and calculate the average value of each frame signal:

[0063]

[0064] Where:

[0065] N is the number of Doppler dimension units per frame;

[0066] M is the total number of range dimension units;

[0067] m is the range dimension unit index;

[0068] n is the Doppler dimension unit index;

[0069] 4.2) Subtract the average value of each frame data obtained in step 4.1) from each frame of data to obtain a zero-mean signal:

[0070]

[0071] Preferably, the specific implementation of step 5) is as follows:

[0072] 5.1) Obtain the radar signal processed in step 4), and the radar signal processed in step 4) is a multi-frame radar signal;

[0073] 5.2) Calculate the global mean and standard deviation of the radar signal obtained in step 5.1);

[0074] 5.3) Perform frame normalization on the radar signal after step 5.2) so that all processed frame radar signals follow a distribution with a mean of 0 and a standard deviation of 1;

[0075] Preferably, in step 5.2), the global mean is the average value of each sampling point of all frames, and the expression of the global mean is:

[0076]

[0077] Where:

[0078] K is the total number of frames;

[0079] N is the total number of Chirps per frame;

[0080] M is the number of range dimension cells per frame;

[0081] k is the frame index;

[0082] n is the Doppler dimension cell index;

[0083] m is the range dimension cell index;

[0084] X k,m,n is the signal intensity data index;

[0085] The global mean and standard deviation in step 5.2) reflect the degree of dispersion of all frame data, and the expression of the global standard deviation is:

[0086]

[0087] The expression of the global frame normalization operation in step 5.3) is:

[0088]

[0089] Preferably, the specific implementation of step 7) is: jointly inject the signal processed by Log filtering, the signal processed by Sin filtering, and the signal processed by Gaussian filtering obtained in step 6.2) into a neural network for deep learning to complete the enhancement processing of the radar signal.

[0090] Preferably, the original radar signal in step 1) is a millimeter-wave radar signal; the neural network in step 7) is R3D, I3D, ResNet or ResNet-LSTM.

[0091] The beneficial effects of the present invention are as follows:

[0092] The present invention provides a method for enhancing three-dimensional channel reconstruction of moving target radar signals. First, the present invention performs serial processing on the original radar signal, including time-frequency conversion, moving target detection, in-frame zero-mean normalization, and global frame normalization. Then, three-channel superposition based on log function for medium-high frequency enhancement, sin function for intermediate frequency enhancement, and Gaussian filter for micro-Doppler feature enhancement is completed, serving as the three input channels of the neural network. Different from the RGB three-channel input in the traditional image processing field, it has stronger physical significance and achieves the purpose of improving the moving target detection effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0093] Figure 1 is a schematic flowchart of the method for enhancing three-dimensional channel reconstruction of moving target radar signals provided by the present invention;

[0094] Figure 2 is a schematic diagram of time-frequency conversion adopted by the present invention;

[0095] Figure 3 is a comparison chart of radar moving target detection data, where (a) is the time-continuous micro-Doppler spectrum obtained through time-frequency conversion, and (b) is the result of moving target indication processing based on (a);

[0096] Figure 4 is a comparison chart of in-frame zero-mean processing, where (a) is the frequency-domain projection of the result of radar moving target detection processing, and (b) is the frequency-domain projection of the result of in-frame zero-mean normalization;

[0097] Figure 5 is a comparison chart of global frame normalization processing, where (a) is the frequency-domain projection of the result of in-frame zero-mean normalization, and (b) is the frequency-domain projection of the result of global frame normalization;

[0098] Figure 6 is a flowchart of the processing after log filtering;

[0099] Figure 7 is a flowchart of the processing after sin filtering;

[0100] Figure 8 is a flowchart of the processing after two-dimensional Gaussian filtering;

[0101] Figure 9 is a comparison chart of the processing results of the three channels, where (a) is the processing result of the global frame normalization data, and (b) is the processing result of the data after three-channel filtering and superposition;

[0102] Figure 10 It is a curve graph of the accuracy rate of the comparative experiment based on R3D;

[0103] Figure 11 It is a curve graph of the accuracy rate of the comparative experiment based on I3D;

[0104] Figure 12 It is a curve graph of the accuracy rate of the comparative experiment based on ResNet;

[0105] Figure 13 It is a curve graph of the accuracy rate of the comparative experiment based on ResNet-LSTM. Specific implementation manner

[0106] The present invention will be described in detail below.

[0107] Refer to Figure 1 , for the problem that the millimeter-wave radar signal itself has a high degree of abstraction and sparse features, making it difficult to directly characterize the target features of the reflected electromagnetic wave, which is not conducive to the further processing of various subsequent algorithms. Through the serial processing of time-frequency conversion, intra-frame zero-mean normalization, and global frame standardization of the radar original signal, and then through the three-channel weighting of high-frequency enhancement based on the log function, intermediate-frequency enhancement based on sin, and micro-Doppler feature enhancement based on Gaussian filtering, as the three input channels of the neural network, different from the RGB three-channel input in the traditional image processing field, it has stronger physical significance and achieves the purpose of improving the radar detection performance.

[0108] Specifically, the three-dimensional channel reconstruction and enhancement method for the moving target radar signal provided by the present invention specifically includes the following steps:

[0109] 1) Time-frequency conversion:

[0110] The millimeter-wave radar emits a frequency-modulated continuous wave (FMCW) or a pulse signal, and processes the reflected signal at the receiving end to extract the distance, speed, and angle information of the target. Among them, the FFT transformation of the fast-time and slow-time dimensions is the core signal processing step. By using the combination of two FFTs, the time-domain signal can be converted into a range-Doppler (RD) two-dimensional spectrogram with clear physical significance, thereby separating noise, suppressing interference, and enhancing the feature interpretability. Refer to Figure 2 , which is the time-frequency change process.

[0111] Adding time dimension information to the two-dimensional spectrogram with clear meaning, the obtained radar electromagnetic wave signal is converted into a three-dimensional representation of "range-Doppler-time", that is, a time-continuous micro-Doppler spectrum, effectively expressing the relationship between the distance and speed of the dynamic target in the time dimension.

[0112] By time-frequency conversion, the original data of the radar is processed into information convenient for subsequent neural network processing, adapting from traditional classifiers to deep learning, and improving the neural network's processing ability for radar signals.

[0113] 2) Moving target detection:

[0114] In moving target detection, in the application of millimeter-wave radar, in indoor or complex environments, the radar received signals contain a large amount of static clutter (such as signals generated by walls, furniture, and stationary targets) and low-frequency interference (such as interference caused by air-conditioning airflows and slowly moving objects). These signals will mask the dynamic characteristics of moving targets; by suppressing the zero-speed or low-speed signal components (i.e., signals generated by static backgrounds and slowly moving interferers, such as static backgrounds and slowly moving interferers), the effective signals of medium- and high-speed moving targets are retained, enhancing the detectability and distinguishability of dynamic characteristics.

[0115] Use the adjacent frame difference method to remove the static features in the micro-Doppler spectrum. To maintain the length of the data, the last frame of the continuous time series is replaced with its previous frame, keeping the length the same as the corresponding segment of the visual modality.

[0116] y n = x n - x n-1

[0117] where x n is the radar signal of the nth frame, x n-1 is the radar signal of the (n - 1)th frame, and y n is the radar signal generated after the difference.

[0118] See Figure 3 the micro-Doppler spectrogram of radar moving target detection and the moving target indication processing data shown in it. Among them, Figure 3 (a) in it is the micro-Doppler spectrogram obtained by time-frequency conversion of the original signal, with many static clutter and low-frequency noise interferences. Figure 3 (b) in it is the moving target indication processing based on (a), filtering out most of the static clutter and low-frequency noise, making the effective signals containing moving targets fully exposed.

[0119] 3) Intra-frame zero mean:

[0120] Intra-Frame Zero Averaging is a key preprocessing method for suppressing static clutter and offsets. Its core advantage is to provide a cleaner and more stable data basis for subsequent processing flows. Especially in refined tasks such as motion detection, it efficiently eliminates intra-frame offsets, improves the quality of signals, and thus enhances the performance of subsequent signal processing and classification models. The results are as Figure 4 shown, where Figure 4 (a) in is the frequency-domain projection of the result of the previous MTI processing. The middle part is the low-frequency part, and the frequency gradually increases towards both sides. It can be clearly seen that there is noise in each frequency before intra-frame zero averaging processing. Most of the noise is eliminated by the intra-frame zero averaging method. The specific results are as Figure 4 (b) in shown.

[0121] Advantages of Intra-Frame Zero Averaging: a) Eliminate hardware biases and restore real signal features. Radar hardware systems (such as receiver amplifiers and mixers) will introduce fixed DC offsets. Such static biases will mask weak motion signals. Zero averaging subtracts the mean value frame by frame, stripping the "false base" introduced by the hardware circuit, and highlighting the real signal fluctuations generated by the detected targets. This step is equivalent to pre-filtering the "electronic fingerprint" of the sensor itself before the signal enters the neural network, preventing the algorithm from being misled by hardware noise. b) Enhance dynamic sensitivity and amplify micro-motion details. Moving targets are manifested as weak time-varying modulation components in radar signals. Zero averaging focuses the processing on the intra-frame dynamic change part by eliminating the static background of each frame signal. For example, when a person changes from stationary to walking, the Doppler frequency shift caused by the periodic movement of the legs will be clearly presented in the signal after mean value zeroing, and the subtle phase changes originally suppressed by the strong DC component will also be amplified, providing a high-contrast input for subsequent time-frequency analysis and filtering enhancement. c) Be compatible with multi-stage processing and optimize feature collaboration. Zero averaging is a key link connecting the upstream and downstream. It plays a supplementary role in the upstream "static information filtering", further suppressing the remaining fixed background noise, and providing an unbiased data benchmark for the downstream global frame normalization (such as amplitude normalization), avoiding distortion of normalization parameters caused by DC offsets. Intra-frame zero averaging is not only a mathematical debiasing operation but also a bridge design between physical signals and algorithm logic. It preserves the electromagnetic "footprint" of moving targets in radar signals by eliminating the "masking effect" of system noise, providing a high-fidelity data starting point for subsequent feature decoupling and intelligent decision-making.

[0122] Exemplarily, the specific implementation method and core function of the intra-frame zero averaging adopted in the present invention.

[0123] Calculate the average value of each frame of data: Calculate the mean value for all sampling points within each frame:

[0124]

[0125] Where N is the number of Doppler dimension units in one frame, and M is the number of range dimension units.

[0126] Subtract this mean value from each frame of data to obtain the signal after zero-mean:

[0127]

[0128] For the case where negative values appear, take the absolute value for processing, and process all frame signals:

[0129] Completing the zero-mean processing of the in-frame data of all frames can well reduce the computational amount and improve the performance of subsequent classification algorithms.

[0130] 4) Global frame normalization:

[0131] In millimeter-wave radar signal processing, global frame normalization is a data preprocessing method that eliminates hardware differences and environmental interference by uniformly adjusting the statistical characteristics (mean and standard deviation) of multi-frame signals. Its core goal is to improve signal quality and model robustness through consistent scaling of cross-frame data; the core advantage is to establish a stable and comparable basis for subsequent analysis by unifying the data scale, especially for tasks such as radar signal processing. Through the processing of this part, the amplitude inconsistencies that may be introduced under different conditions (such as temperature drift, power supply fluctuations) can be well solved.

[0132] Comparison with traditional local normalization: Traditional methods often use in-frame normalization (such as scaling the amplitude of a single frame to [0, 1]). Although it can alleviate the in-frame dynamic range problem, it cannot solve the scale inconsistencies across frames and scenarios. Global frame normalization is based on the statistical laws of all training data (such as global maximum value, mean, and variance) to construct a unified physical quantity mapping relationship, enabling the model to still maintain stable feature parsing ability even when encountering extreme scenarios (such as very long-distance weak signals, strong electromagnetic interference environments) not covered by the training set during the inference stage. In the application of radar moving target detection, this global consistency is directly transformed into the physical interpretability of dynamic features: the normalized amplitude value corresponds to the kinetic energy intensity of the actual action, the phase change corresponds to the limb movement direction, and the time-frequency distribution corresponds to the action rhythm - this mapping relationship enables the algorithm not only to detect moving targets but also to quantitatively evaluate parameters such as speed.

[0133] Exemplarily, the implementation method of the global frame normalization adopted by the present invention is:

[0134] 4.1) Input data: Multi-frame radar signals (each frame contains multiple Chirps).

[0135] 4.2) Calculate the global mean and standard deviation:

[0136] The global mean refers to the average value of each sampling point of all frames.

[0137]

[0138] Where:

[0139] K is the total number of frames, N is the number of Doppler dimension units per frame, M is the number of range dimension units per frame, m is the range dimension unit index, k is the frame index, n is the Doppler dimension index. Then, the global standard deviation is calculated, which can reflect the degree of dispersion of all frame data.

[0140]

[0141] 4.3) Frame normalization operation:

[0142] The original signals of all frames are adjusted according to the same set of parameters (μ global , σ global ) to ensure the comparability of data collected at different times and under different conditions. Through the processing of this part, the amplitude inconsistencies that may be introduced under different circumstances (such as temperature drift and power supply fluctuations) can be well solved.

[0143] For the specific processing effect, refer to Figure 5 , Figure 5 where (a) in it is the frequency-domain projection of the in-frame zero-mean result, Figure 5 and (b) in it is the frequency-domain projection of the global frame normalization result. After multi-frame global frame normalization, the influence of noise on the signal is further reduced, especially the noise in the high-frequency part is further weakened, improving the signal-to-noise ratio.

[0144]

[0145] 5) Three-channel signal input:

[0146] In this module, high-frequency signal extraction in log, intermediate-frequency signal extraction in Sin, and Gaussian filtering processing are three parallel independent dimensions. Finally, the information of the three dimensions is input into the subsequent classification model together.

[0147] Log high-frequency signal filtering:

[0148] In the range-Doppler map, the low frequency (close to the zero frequency) usually corresponds to strong static clutter (such as background and fixed objects), and its energy is much higher than that of the moving target signals in the middle and high frequencies. When directly linearly weighted, the middle and high-frequency signals may be submerged by the low-frequency energy. In order to highlight the characteristics of the radar signal, the normalized micro-Doppler spectrum is feature-enhanced. The log filter enhances both the intermediate-frequency and high-frequency micro-Doppler features through the inherent mathematical properties of the log function.

[0149]

[0150] I is the signal after zero-mean global frame normalization in the previous step, that is, the input signal of this module. x is the index of the Doppler dimension unit corresponding to I; f is the mapping function; N is the number of Doppler dimension units per frame; A is the scaling coefficient; 1 is a constant; to avoid negative scaling; I log is the output signal after I is scaled by a specific frequency, that is, the signal after Log filtering processing.

[0151] For the specific processing results, see Figure 6 In Figure 6 , after the left global frame-normalized radar signal is processed by log medium-high frequency filtering, the low-frequency information is further suppressed, and the static noise is almost reduced to 0, largely eliminating the useless information, excluding the zero low-frequency interference for the subsequent model, and saving a large amount of computing power; in the subsequent RGB three-channel model of neural network classification, the output result of this module is used to replace the input of the R channel.

[0152] Sin intermediate frequency signal filtering:

[0153] Because the intermediate frequency signal in the range-Doppler spectrogram often contains more dynamic information than the high frequency, in order to further enhance the dynamic information-rich in the range-Doppler spectrogram, a sin filter is designed to enhance the intermediate frequency features:

[0154]

[0155] Where:

[0156] I is the signal after zero-mean global frame normalization in the previous step, that is, the input signal of this module; x is the index of the Doppler dimension unit corresponding to I; N is the number of Doppler dimension units per frame; g is the mapping function; B is the scaling coefficient; I sin is the output signal after I is scaled by a specific frequency, that is, the signal after Sin filtering processing.

[0157] See Figure 7 , after the left global frame-normalized radar signal is processed by sin intermediate frequency filtering, the low-high frequency information is suppressed, and the intermediate frequency information is further amplified; the static noise is almost reduced to 0, largely eliminating the useless information, excluding the zero low-frequency interference for the subsequent model, and saving a large amount of computing power; in the subsequent RGB three-channel model of neural network classification, the output result of this module is used to replace the input of the G channel.

[0158] Gaussian filtering:

[0159] Gaussian filtering reduces the impact of random noise by weighted averaging of adjacent pixels, and has the best effect on suppressing noise that follows a normal distribution. At the same time, it can reduce the damage to the details of the radar spectrogram and is widely used in the thermal noise suppression of radar signals. Compared with methods such as mean filtering and median filtering, the weights of the Gaussian kernel gradually decrease from the center outwards, smoothing the area while preserving the edge transition.

[0160]

[0161] And in the process of algorithm implementation, the two-dimensional Gaussian kernel can be decomposed into two one-dimensional kernels, and the computational complexity is reduced from O(n 2 ) to O(2n), greatly simplifying the calculation.

[0162] G(i,j) = G(i)·G(j)

[0163] Where:

[0164]

[0165] Where:

[0166] i and j represent the abscissa and ordinate in the pixel image respectively, G(i,j) is the signal intensity of the Gaussian filtering kernel at the coordinate (i, j), σ is the Gaussian kernel standard deviation parameter, which can be selected according to the signal. The larger σ is, the wider the kernel is, the stronger the smoothing effect is, but the more details are lost.

[0167]

[0168] Among them, the original signal two-dimensional image is I(x,y), and the two-dimensional image after Gaussian filtering is I Gauss (x,y), k = [kernel size / 2] (for example, a 3×3 kernel corresponds to k = 1), G(i,j) is the two-dimensional Gaussian kernel, C is the scaling coefficient, and boundary processing usually uses symmetric padding or zero padding.

[0169] See Figure 8 , where the radar signal of the left global frame standardized is processed by Gaussian filtering. Through Gaussian filtering, random noise is greatly suppressed, random error is reduced, the impact of random noise is reduced, and the effect of suppressing noise that follows a normal distribution is the best. In the RGB three-channel model of subsequent neural network classification, the output result of this module is used to replace the input of the B channel.

[0170] The core advantage of the multi-channel filtering and weighting adopted in the present invention lies in constructing a signal representation system for radar moving target detection through three feature channels with clear physical meanings. In the traditional process, the original time-frequency signal or the data after single processing is often directly input into the neural network, which essentially relies on the algorithm to mine features by itself. However, in this solution, through three physically interpretable signal processing modules, namely Log filtering, Sin filtering, and Gaussian filtering, a three-channel input structure with clear information division of labor is formed, realizing the decoupling and enhancement of the inherent features of radar signals. In the radar moving target detection scenario, the three channels respectively undertake different-dimensional feature enhancement functions: the Log filtering channel compresses the dynamic range of radar signals spanning several orders of magnitude into a stable interval by taking the logarithm operation of the signal amplitude, enabling the weak signals generated by target micro-movements and the strong reflection background to be presented evenly, avoiding the weak features being submerged by strong signals, and highlighting the medium and high-frequency information well; the Sin filtering channel focuses on the extraction of intermediate-frequency information, uses the sine function to perform non-linear mapping on the signal, and strengthens the unique phase modulation characteristics in moving targets, enabling the neural network to directly capture the time-varying laws related to moving targets; the Gaussian filtering channel, through the characteristics of smoothing noise reduction and edge preservation, suppresses the interference of environmental noise while highlighting the contour features (such as the transient jumps at the start / end) of moving targets in the time-frequency domain, providing clear structured information input for the neural network.

[0171] Finally, the output results of the three filters, namely Log filtering, Sin filtering, and Gaussian filtering, are used as the three input channels of the neural network. Different from the RGB three-channel input in the traditional image processing field, it has stronger physical meanings and achieves the purpose of enhancing radar signals. The three-channel weighted fusion mechanism adopted in the present invention is essentially different from the traditional RGB image input: the traditional RGB channels represent the color space information captured by optical sensors, while the three channels in the present invention are derived from the physical property decomposition of the interaction between radar electromagnetic waves and moving targets, forming a unique "feature spectrum" of radar signals. This design enables the neural network to directly utilize physical prior knowledge to accelerate the feature learning process.

[0172] See Figure 9 , the three-channel processing (a) original data (b) processed data. Among them, the original data is respectively passed through three parallel processes of Log filtering, Sin filtering, and Gaussian filtering to obtain the output results in three dimensions, and the results of these three dimensions are visualized as Figure 9 (b) shown.

[0173] Experimental verification example:

[0174] A) Experimental comparison under the R3D neural network:

[0175] Table 1 Comparison experimental results based on the R3D neural network

[0176]

[0177] See Figure 10 The accuracy of the feature-enhanced data is 8.93% higher than that of the heatmap data and 3.47% higher than that of the original data, verifying the strengthening effect of feature enhancement on spatio-temporal feature extraction. Feature-enhanced data: The convergence speed is the fastest. After 50 Epochs, the accuracy breaks through 68%, and after 100 Epochs, it stabilizes at about 70%, with strong stability and fluctuations less than ±1% between 150 - 250 Epochs. Original data: There are obvious oscillations, with a sharp fluctuation of 5% between 100 - 250 Epochs. After about 100 Epochs, it enters the model bottleneck: the final accuracy is close to 67%, but it lags behind the feature-enhanced data by 3.47%. Heatmap data: The convergence is also relatively stable, but the final accuracy is only 61.55%, significantly lower than the other two types of data.

[0178] B) Experimental comparison under the I3D neural network:

[0179] Table 2 Comparative experimental results based on the I3D neural network

[0180]

[0181] See Figure 11 Through experimental comparison, the accuracy of the feature-enhanced data is 23.19% higher than that of the original data and 5.38% higher than that of the heatmap data, verifying the core value of data preprocessing and feature engineering. Feature-enhanced data: The convergence speed is the fastest. After about 70 Epochs, it enters the stable period (accuracy > 65%). The final accuracy is the highest and the fluctuations are small (stable at 70%+ after 200 Epochs). Heatmap data: The convergence speed is the second. After about 100 Epochs, there is a small oscillatory increase (such as a fluctuation increase of about ±3% between 150 - 200 Epochs). Original data: It is difficult to converge and is still not stable until 250 Epochs. The final accuracy is less than 50%, indicating that the model cannot effectively capture the discriminative features in the original signal.

[0182] C) Experimental comparison based on ResNet:

[0183] Table 3 Comparative experimental results based on the ResNet neural network

[0184]

[0185]

[0186] See Figure 12, Feature-enhanced data (63.25%) > Original data (61.2%) > Heatmap data (60.7%). The accuracy of the feature-enhanced data is only 2.05% higher than that of the original data and 2.55% higher than that of the heatmap data, indicating that ResNet is less sensitive to data forms than 3D CNN (such as I3D / R3D), but still verifies the universal advantage of feature enhancement. Feature-enhanced data: The convergence speed is the fastest. After 10 Epochs, the accuracy breaks through 55%, and stabilizes at around 63% after about 25 Epochs and then slowly decreases. Original data: Slowly converges until it approaches 61% after about 75 Epochs; slightly oscillates and then slowly decreases. Heatmap data: The worst performance, with a final accuracy of only 60.7% and the most serious overfitting phenomenon.

[0187] D) Experimental comparison under ResNet-LSTM neural network:

[0188] Table 4 Comparison test results based on ResNet-LSTM neural network

[0189]

[0190] See Figure 13 , The feature-enhanced data is 3.81% higher than the original data and 3.2% higher than the heatmap data, verifying the strong adaptability of spatio-temporal joint modeling (ResNet-LSTM) to feature-enhanced data. Feature-enhanced data: Fast convergence. After 50 Epochs, the accuracy breaks through 67%, and stabilizes at 70%+ after 75 Epochs; high stability, with fluctuations of about ±0.5% between 200 - 250 Epochs. Original data: Convergence lag: The slowest to converge among the three groups of data, with an accuracy of 69.16%. Heatmap data: Medium performance, with a final accuracy of 69.77%.

[0191] Comprehensive analysis:

[0192] Based on the comparative experiments of different model architectures (I3D / R3D, ResNet, ResNet-LSTM) and input data forms (original data, heatmap data, feature-enhanced data), the core advantages of data augmentation preprocessing can be summarized in the following dimensions:

[0193] 1) Universality and significance of performance improvement: Data augmentation has a significant improvement effect on spatio-temporal joint modeling models (3D CNN / LSTM) because it can optimize spatio-temporal features simultaneously; even for pure spatial models (ResNet), enhancement can still bring stable gains, verifying its cross-architecture universality.

[0194] 2) Dynamic feature capture, dynamic background elimination reduces the influence of environmental clutter, making the model more focused on effective motion features.

[0195] 3) Noise suppression and signal-to-noise ratio improvement: By operations such as amplitude normalization and filtering, the ADC quantization error and multipath reflection interference in the radar signal are suppressed.

[0196] 4) Filtering input in three-channel dimension: Through log filtering, sin filtering and Gaussian filtering, signals with different emphases on physical meanings in three dimensions are output, converting the original signal into highly discriminative features and reducing the implicit learning difficulty of the model.

[0197] 5) Optimization of training efficiency and stability: The convergence speed is increased, fluctuations are suppressed, and the risk of overfitting can be reduced.

[0198] Data augmentation preprocessing shows significant advantages in enhancing radar moving target signals through triple mechanisms of dynamic feature capture, noise suppression, and optimization of physical meanings in three channels.

Claims

1. A three-dimensional channel reconstruction enhancement method for moving target radar signals, characterized in that: The three-dimensional channel reconstruction enhancement method for moving target radar signals includes the following steps: 1) Obtain the original radar signal; 2) Perform time-frequency transformation on the original radar signal obtained in step 1); 3) Perform moving target detection on the radar signal obtained after step 2); 4) Perform in-frame zero-mean processing on the radar signal obtained after step 3); 5) Perform global frame normalization processing on the radar signal obtained after step 4); 6) Perform multi-channel signal filtering processing on the radar signal obtained after step 5); 7) Inject the result obtained in step 6) into a neural network for deep learning to complete the enhancement processing of the radar signal.

2. The three-dimensional channel reconstruction enhancement method for moving target radar signals according to claim 1, characterized in that: The specific implementation method of step 6) is as follows: 6.1) Duplicate the radar signal obtained in step 5) into three signals; 6.2) Perform Log filtering on the first signal obtained in step 6.1) to obtain the signal after Log filtering; perform Sin filtering on the second signal obtained in step 6.1) to obtain the signal after Sin filtering; perform Gaussian filtering on the third signal obtained in step 6.1) to obtain the signal after Gaussian filtering.

3. The three-dimensional channel reconstruction enhancement method for moving target radar signals according to claim 2, wherein: The specific implementation method of performing Log filtering in step 6.2) is as follows: Where: I is the radar signal obtained by inputting step 5); x is the index of the Doppler dimension unit corresponding to I; N is the number of Doppler dimension units per frame; f is the mapping function; A is the scaling coefficient; 1 is a constant to avoid negative scaling; I log is the output signal after I is scaled by a specific frequency, that is, the signal after Log filtering; The specific implementation method of performing Sin filtering in step 6.2) is as follows: Where: I is the radar signal obtained by inputting step 5); x is the index of the Doppler dimension unit corresponding to I; g is the mapping function; N is the number of Doppler dimension units per frame; B is the scaling coefficient; I sin is the output signal after I is scaled by a specific frequency, that is, the signal after Sin filtering; The specific implementation method of performing Gaussian filtering in step 6.2) is as follows: Where: I(x, y) is the radar signal obtained in step 5); I Gauss (x, y) is a two-dimensional image after Gaussian filtering, that is, a signal after Gaussian filtering; C is the scaling coefficient; x is the index of the Doppler dimension unit corresponding to I; y is the index of the range dimension unit corresponding to I; k is 1 / 2 of the kernel size; G(i, j) is the signal intensity of the Gaussian filter kernel at coordinates (i, j), and the expression of G(i, j) is: G(i, j) = G(i) · G(j) Where: Where: i and j respectively represent the abscissa and ordinate in the pixel image; σ is the Gaussian kernel standard deviation, which is selected according to the signal specifically.

4. The three-dimensional channel reconstruction enhancement method for moving target radar signals according to claim 3, characterized in that: The specific implementation method of step 2) is as follows: 2.1) Convert the original radar signal obtained in step 1) into a range-velocity two-dimensional spectrogram with clear physical meaning; 2.2) Add time dimension information to the two-dimensional spectrogram obtained in step 2.1), convert the obtained radar electromagnetic wave signal into a three-dimensional representation of range-Doppler-time, and obtain a time-continuous micro-Doppler spectrogram to complete the time-frequency transformation of the original radar signal.

5. The three-dimensional channel reconstruction enhancement method for moving target radar signals according to claim 4, wherein: The moving target detection in step 3) is completed by using the adjacent frame difference method. The adjacent frame difference method specifically replaces the last frame of the continuous time series with its previous frame, keeping the length the same as the corresponding segment of the visual modality; y n = x n - x n-1 Where: x n is the radar signal of the nth frame; x n-1 is the radar signal of the (n - 1)-th frame; y n is the radar signal generated after differentiation.

6. The three-dimensional channel reconstruction enhancement method for moving target radar signals according to claim 5, characterized in that: The specific implementation method of step 4) is: 4.1) Obtain each fast time sampling point and calculate the average value of each frame signal: Where: N is the number of Doppler dimension units per frame; M is the number of range dimension units per frame; m is the range dimension unit index; n is the Doppler dimension unit index; 4.2) Subtract the average value of each frame data obtained in step 4.1) from each frame of data to obtain a signal with zero mean:

7. The three-dimensional channel reconstruction enhancement method for moving target radar signals according to claim 6, wherein: The specific implementation method of step 5) is: 5.1) Obtain the radar signal processed in step 4), and the radar signal processed in step 4) is a multi-frame radar signal; 5.2) Calculate the global mean and standard deviation of the radar signal obtained in step 5.1); 5.3) Perform frame normalization operation on the radar signal after step 5.2) to make all processed frame radar signals follow a distribution with a mean of 0 and a standard deviation of 1.

8. The three-dimensional channel reconstruction enhancement method for moving target radar signals according to claim 7, wherein: In step 5.2), the global mean is the average value of each sampling point of all frames, and the expression of the global mean is: Where: K is the total number of frames; N is the total number of Chirps per frame; M is the number of range dimension units per frame; K is the frame index; n is the Doppler dimension index; m is the range dimension unit index; X k,m,n is the signal strength data index; In step 5.2), the global mean and standard deviation reflect the degree of dispersion of all frame data, and the expression of the global standard deviation is: The expression of the frame normalization operation in step 5.3) is: In step 5.4), the multi-frame unified scaling process scales the original signals of all frames according to the same set of parameters (μ global , σ global ).

9. The three-dimensional channel reconstruction enhancement method for moving target radar signals according to claim 8, wherein: The specific implementation method of step 7) is: jointly inject the signal processed by Log filtering, the signal processed by Sin filtering, and the signal processed by Gaussian filtering obtained in step 6.2) into a neural network for deep learning to complete the enhancement processing of the radar signal.

10. The three-dimensional channel reconstruction enhancement method for moving target radar signals according to any one of claims 1-9, characterized in that: The original radar signal in step 1) is a millimeter-wave radar signal; the neural network in step 7) is R3D, I3D, ResNet or ResNet-LSTM.

Images