GPU-based millimeter wave radar signal gesture micro-Doppler feature extraction method

Abstract

The invention discloses a millimeter wave radar signal gesture micro-Doppler feature extraction method based on a GPU, and belongs to the technical field of radar signal processing. The invention aims to solve the problem that the processing speed and the extraction precision are difficult to consider at the same time due to large operand and poor real-time performance when the gesture micro-Doppler signal is processed by a traditional CPU (Central Processing Unit). According to the technical scheme, the method comprises the following steps: performing phase intervention and high-resolution two-dimensional Fourier transform on multi-channel original echoes through a GPU (Graphics Processing Unit) to generate a distance-Doppler spectrum of each channel; performing amplitude accumulation fusion on the distance-Doppler spectrum of each channel to enhance a signal, and discriminating a hand target area by adopting a two-dimensional self-adaptive operator; and finally, integrating continuous multi-frame detection results, constructing a micro-Doppler characteristic spectrogram representing fine gesture motion through time-frequency analysis, and extracting quantized micro-motion characteristics from the micro-Doppler characteristic spectrogram. The method can be used for accurately extracting the micro-Doppler features of the gestures at a high speed, and is mainly applied to the fields of real-time human-computer interaction, intelligent control and the like.

Description

Technical Field

[0001] The present invention belongs to the technical field of radar signal processing, and in particular relates to a method for extracting micro-Doppler features of millimeter-wave radar signal gestures based on a GPU. Background Art

[0002] Millimeter-wave radar, due to its high operating frequency band, short wavelength, compact antenna size, and high detection accuracy, has been widely used in many fields. These applications include environmental perception in assisted driving and autonomous driving systems, human presence detection and gesture recognition in smart homes, contactless vital sign monitoring in healthcare, and security monitoring of specific areas. In these application scenarios, minute movements of a target, such as fine hand movements, produce a subtle frequency modulation in the radar echo, in addition to the main Doppler shift. This phenomenon is known as the micro-Doppler effect. Micro-Doppler signatures can precisely characterize the dynamic behavior and structural properties of a target, providing a rich information dimension for precise target detection, especially the recognition, accurate classification, and state assessment of fine movements such as gestures. Therefore, they have crucial research and application value in fields such as human-computer interaction and intelligent control.

[0003] To achieve effective gesture micro-Doppler feature extraction, the signal processing process usually needs to include: precise interferometric processing of the radar's multi-channel raw data to ensure signal fidelity and phase consistency; effective dynamic and static clutter filtering and suppression to enhance weak hand motion signals; generating a range-Doppler spectrum through high-resolution fast Fourier transform (FFT); precise adaptive target detection on the spectrum to accurately delineate the hand target area; and integrating multi-frame detection results to reconstruct and accumulate the target's final micro-Doppler feature spectrum, which can clearly show the time-frequency spectrum characteristics of gesture motion.

[0004] Millimeter-wave radars typically have high operating bandwidths and sampling rates, resulting in a very large amount of raw data. Especially in applications requiring high refresh rates and high resolutions, the aforementioned signal processing flow, which includes precise corrections, multiple high-resolution FFT operations, complex two-dimensional adaptive operator processing, and cross-frame data analysis, places extremely high demands on computing resources. Traditional central processing unit (CPU)-based serial or limited parallel processing methods often face bottlenecks in insufficient computing power when processing such large-scale and highly complex data. This makes it difficult to meet the stringent real-time signal processing requirements of many application scenarios, potentially leading to significant processing delays. This limits the effectiveness of micro-Doppler technology in dynamic, real-time monitoring, and rapid response scenarios.

[0005] To address these challenges, existing technologies have explored ways to improve processing efficiency. Some approaches may simplify signal processing algorithms or reduce the dimensionality of data processing, but this often comes at the expense of micro-Doppler feature extraction accuracy, resolution, or signal-to-noise ratio, potentially leading to loss or blurring of key target micro-motion information. On the other hand, pursuing high-precision gesture micro-Doppler feature extraction, such as employing more sophisticated spectral analysis techniques, advanced clutter suppression algorithms, or multi-dimensional joint processing algorithms, inherently increases computational complexity, further exacerbating the difficulty of real-time processing on traditional computing platforms. Specifically, high-standard two-dimensional adaptive operator detection requires selecting reference and protection windows around each range-Doppler unit and performing complex statistical calculations, resulting in a high computational load. Generating high-quality micro-Doppler spectra requires precise accumulation and transformation of multi-frame processing results. Without the support of an efficient computing architecture, these steps can easily become performance bottlenecks in the entire signal processing chain. Therefore, developing a millimeter-wave radar gesture micro-Doppler feature extraction method that can balance processing speed and feature extraction accuracy has important practical significance and application prospects for fully leveraging the performance advantages of millimeter-wave radar in close-range and fine motion perception and expanding its application in fields such as intelligent human-computer interaction. Summary of the Invention

[0006] In view of the shortcomings of the above-mentioned prior art, the purpose of the present invention is to propose a GPU-based millimeter-wave radar signal gesture micro-Doppler feature extraction method. To achieve the above technical objectives, the present invention adopts the following technical solutions to achieve it.

[0007] The present invention provides a method for extracting micro-Doppler features of millimeter-wave radar signal gestures based on a GPU, the method comprising:

[0008] Step 1: The multi-channel raw echoes of the millimeter-wave radar are collected and interpolated through GPU optimization to obtain a high-fidelity and phase-consistent multi-channel digital signal stream of gesture data.

[0009] Step 2: Performing GPU parallel transformation and dynamic enhancement processing on the multi-channel digital signal stream of the gesture data to obtain the range-Doppler spectrum of the gesture data of each channel;

[0010] Step 3: For the gesture data range-Doppler spectrum of each channel, GPU-accelerated multi-channel information fusion is performed to obtain an enhanced gesture data range-Doppler amplitude image;

[0011] Step 4: The enhanced gesture data range-Doppler amplitude image is subjected to GPU parallel two-dimensional adaptive operator discrimination to obtain gesture target detection results that can be used for micro-motion analysis;

[0012] Step 5: For the gesture target detection results, GPU information fusion and time-frequency analysis are performed to obtain a gesture target micro-Doppler feature spectrum that characterizes the fine motion characteristics of the gesture target.

[0013] Step 6: By performing feature analysis on the micro-Doppler characteristic spectrum of the gesture target, a quantitative micro-motion feature that can support fine recognition of the gesture target is obtained.

[0014] Compared with the prior art, the present invention has the following beneficial effects:

[0015] 1. High-speed processing of the entire micro-Doppler analysis process for gesture recognition is achieved, meeting the real-time requirements of scenarios such as human-computer interaction. This invention addresses the computational efficiency bottlenecks of traditional methods when processing the high data rates and large computational loads of millimeter-wave radars. It deeply integrates the complete signal processing chain for high-quality gesture micro-Doppler analysis, including precise interferometric processing, high-resolution multi-dimensional Fourier transforms, multi-channel information fusion, advanced two-dimensional adaptive operator detection, and multi-frame micro-Doppler feature accumulation and spectrum construction, into the GPU parallel computing architecture. This significantly improves the computing speed and data throughput of the overall algorithm, providing powerful high-performance computing support for the real-time extraction and analysis of gesture micro-Doppler features, making applications in scenarios such as dynamic gesture tracking and instant human-computer interaction possible.

[0016] 2. GPU parallel optimization ensures precise and efficient execution of advanced processing algorithms for gesture signals, laying a high-quality data foundation for refined gesture feature extraction. This invention leverages the massively parallel processing capabilities of GPUs to efficiently and precisely parallelize the computationally intensive steps in radar signal processing that are crucial for extracting refined gesture micro-Doppler signatures. The efficient GPU implementation of these advanced processing steps not only significantly reduces the time required to acquire high-quality intermediate data, but also ensures rapid processing without sacrificing signal analysis accuracy, laying a solid data foundation for subsequent, precise gesture micro-Doppler signature extraction.

[0017] 3. The rapid generation and output of high-resolution gesture micro-Doppler feature spectra are achieved, significantly improving the reliability and timeliness of subsequent applications such as gesture recognition. By efficiently executing gesture micro-Doppler feature accumulation and spectrum construction algorithms specifically for multi-frame data on the GPU, the present invention can quickly generate micro-Doppler feature spectra with high temporal resolution and Doppler resolution, clearly showing the fine movement patterns of various parts of the hand. This real-time and fine micro-Doppler feature output effectively overcomes the problems of outdated target motion information and loss or distortion of gesture feature details caused by calculation delays in traditional methods. This provides a more reliable, timely and detailed dynamic feature basis for subsequent complex gesture recognition based on micro-Doppler information, specific hand movement classification, and other advanced applications that rely on fine human body part movement analysis, significantly enhancing the application efficiency of millimeter-wave radar in the fields of intelligent human-computer interaction and close-range fine perception.

[0018] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flowchart of a method for extracting micro-Doppler features of millimeter-wave radar signal gestures based on a GPU, provided by the present invention;

[0020] Figure 2 This is a schematic diagram of the optimization of the millimeter-wave radar multi-channel raw sampling data structure after the GPU-based millimeter-wave radar signal gesture micro-Doppler feature extraction method provided by the present invention;

[0021] Figure 3 This is a schematic diagram of the range-Doppler spectrum of gesture data of a single receiving channel of a GPU-based millimeter-wave radar signal gesture micro-Doppler feature extraction method provided by the present invention;

[0022] Figure 4 This invention provides a GPU-based millimeter-wave radar signal gesture micro-Doppler feature extraction method, a single-frame target echo signal energy-enhanced gesture data range-Doppler amplitude diagram;

[0023] Figure 5 This is a gesture target detection result diagram generated by a GPU-based millimeter-wave radar signal gesture micro-Doppler feature extraction method provided by the embodiment of the present invention;

[0024] Figure 6 This is a schematic diagram of gesture target micro-Doppler features generated by a GPU-based millimeter-wave radar signal gesture micro-Doppler feature extraction method provided by the present invention;

[0025] Figure 7The present invention provides a method for extracting micro-Doppler features of millimeter-wave radar signal gestures based on GPU, and generates a micro-Doppler feature spectrum after extracting features of the waving gesture.

[0026] Figure 8 The present invention provides a method for extracting micro-Doppler features of millimeter-wave radar signal gestures based on GPU, and generates a micro-Doppler feature spectrum after extracting features of the downward-swinging hand gesture.

[0027] Figure 9 The present invention provides a method for extracting micro-Doppler features of millimeter-wave radar signal gestures based on GPU, and generates a micro-Doppler feature spectrum after extracting features of the check gesture.

[0028] Figure 10 A micro-Doppler feature spectrum is generated after extracting the features of an open fist gesture using a GPU-based millimeter-wave radar signal gesture micro-Doppler feature extraction method provided by the present invention;

[0029] Figure 11 The present invention provides a method for extracting micro-Doppler features of millimeter-wave radar signal gestures based on GPU, which generates a micro-Doppler feature spectrum after extracting features of left and right waving gestures. DETAILED DESCRIPTION

[0030] In order to further illustrate the technical means and effects adopted by the present invention to achieve the predetermined purpose of the invention, the following is a detailed description of a GPU-based millimeter-wave radar signal gesture micro-Doppler feature extraction method proposed in accordance with the present invention, in conjunction with the accompanying drawings and specific embodiments.

[0031] The aforementioned and other technical contents, features, and effects of the present invention are clearly presented in the following detailed description of the specific embodiments in conjunction with the accompanying drawings. Through the description of the specific embodiments, a deeper and more specific understanding of the technical means and effects adopted by the present invention to achieve the intended purpose can be obtained. However, the accompanying drawings are provided for reference and illustration purposes only and are not intended to limit the technical solutions of the present invention.

[0032] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations are intended to cover non-exclusive inclusion, such that an article or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of additional identical elements in the article or device comprising the element.

[0033] See Figure 1 , Figure 1 This is a flowchart of a method for extracting micro-Doppler features of millimeter-wave radar signal gestures based on a GPU, provided by the present invention. The micro-Doppler feature extraction method includes:

[0034] Step 1: The multi-channel raw echoes of the millimeter-wave radar are collected and interpolated through GPU optimization to obtain a high-fidelity and phase-consistent multi-channel digital signal stream of gesture data.

[0035] Step 1 specifically includes:

[0036] 1a) Format conversion and data reorganization of millimeter-wave radar multi-channel raw sampling data

[0037] To clearly describe the structure of millimeter-wave radar multi-channel raw sampling data and how it is formed, the following key radar parameters are assumed and defined:

[0038] : Indicates the number of effective ADC sampling points during each FM pulse.

[0039] : Indicates the number of frequency modulated pulse sequences contained in each data frame.

[0040] : Indicates the number of physical receiving channels of the radar system.

[0041] : Indicates the total number of frames of millimeter-wave radar multi-channel raw sampling data.

[0042] Based on the above parameters, for the radar system receive channels (of which ), in the first observation of the system In the data frame (where represents continuous frames in time), the channel receives the FM pulses (where ) can be represented as a complex sampling sequence after being mixed, filtered, amplified, sampled and digitized by ADC at the radar front end. We denote this complex sampling sequence as ,in is the sampling point index along the fast time axis (i.e., the sampling time axis of a single pulse).

[0043] The complex sampling sequence It is composed of in-phase component and quadrature component, and its mathematical expression is:

[0044]

[0045] in, Is an imaginary unit. and Respectively represented in Sampling time, Channel, Frame, In-phase and quadrature sampling values ​​of the pulse. and It is extracted from the original binary data stream output by the radar hardware through a parsing program based on the preset digital interface protocol and data packaging format.

[0046] The present invention first reads the original int16_t type data from the binary file, then separates these original integer values ​​according to the rule that each channel contains two signals, I and Q, and combines them into two-dimensional arrays corresponding to the four receiving channels. In these arrays, each column represents the frequency modulation pulse corresponding to a specific data frame. The multiple columns of the array continuously store the sampling data of all FM pulses in all data frames. This processing provides a regularized and structured input basis for the subsequent signal processing steps sent to the GPU for parallelization.

[0047] 1b) Initial signal correction and enhancement

[0048] Actual radar hardware systems can introduce numerous non-ideal factors during signal acquisition, such as DC bias, I / Q signal imbalance, inter-pulse phase jitter caused by phase noise in the transmitter or local oscillator, and inconsistent amplitude and phase responses between multi-channel receivers. If not precisely corrected, these factors can significantly degrade signal quality, interfere with subsequent Doppler spectrum analysis, and ultimately hinder the accurate extraction of micro-Doppler signatures. This substep aims to correct these distortions through GPU parallel computing, thereby improving signal coherence and fidelity.

[0049] The formatted complex baseband signal As input, a series of correction algorithms are applied in parallel on the GPU to gradually obtain the corrected signal.

[0050] First, perform DC offset correction to remove the , each data frame Each pulse within The DC component that may exist in its I and Q signals is removed by calculating the mean of the I and Q signals of each pulse and subtracting these means from the original signal. Then, I / Q imbalance correction is performed, using the pre-calibrated DC component for each receiving channel. The specific calibration coefficient (relative gain error and phase mismatch ), corrects the signal distortion caused by the gain difference and phase non-orthogonality between the I and Q paths of the receiver. This correction process is achieved through GPU parallel computing.

[0051] Then the inter-pulse phase fluctuation correction is performed, and the phase error of each pulse relative to the stable reference is estimated by referring to the target echo and other methods. Then, the signal is subjected to pulse-by-pulse phase rotation correction on the GPU to compensate for the phase noise of the transmitting source or local oscillator in the same data frame. From a pulse To the next pulse This random phase jitter will broaden the Doppler spectrum and destroy the fine structure of the micro-Doppler signal.

[0052] Finally, channel equalization is performed, and the receiving channels obtained in advance are used Complex gain correction factor relative to a reference channel ( is the amplitude correction coefficient, The signal of each channel is corrected on the GPU to ensure that each channel in the multi-channel receiving system has consistent amplitude and phase response before subsequent processing.

[0053] After completing these GPU-based parallel correction operations, the output is the high-fidelity multi-channel radar signal that has been interferometrically processed, which is recorded as ,These signals will be used for subsequent gesture micro-Doppler feature extraction.

[0054] 1c) Optimize the frame organization and structure of the reorganized millimeter-wave radar multi-channel raw sampling data

[0055] After obtaining high-fidelity multi-channel radar signals, these data will be further organized and optimized for ease of management and subsequent efficient frame processing. Since the original parsed data is a continuous time series divided by channel, in this sub-step, these sequences will be divided according to the number of pulses per frame set by the radar system. and the total number of captured frames It is clearly structured into logical blocks called "frames".

[0056] Millimeter-wave radar multi-channel raw sampling data is organized into a four-dimensional structure, see Figure 2 , Figure 2 This is a schematic diagram of the optimized reorganization of the millimeter wave radar multi-channel raw sampling data structure provided by the present invention. Consisting of consecutive data frames, in each frame, the data contains For each receiving channel, such as RX1 to RX4 in the figure, the data in one frame is a two-dimensional matrix with dimensions of ,in Represents the number of FM pulses in the frame, Represents the number of fast time sampling points for each pulse. Such a four-dimensional data structure (channel pulse Sampling point Frame) is the basis for subsequent GPU parallel processing.

[0057] This optimized data organization of the raw, multi-channel millimeter-wave radar data is a key preprocessing step. It aligns the data structure with the logical hierarchy of the radar's operating mode, facilitating rapid and independent extraction of complete data blocks from any given data frame during subsequent processing. For large datasets, proper organization facilitates more efficient memory management and prepares for subsequent data transfer to the GPU.

[0058] Step 2: The multi-channel digital signal stream of the gesture data is processed by GPU parallel transformation and dynamic enhancement to obtain the gesture data range-Doppler spectrum of each channel.

[0059] 2a) GPU Parallel Advanced Clutter Suppression

[0060] This stage aims to weaken and eliminate non-target echoes, especially strong clutter from static or slow-moving large objects, which may mask the dynamic targets of interest and their micro-Doppler characteristics. The method adopted is adaptive averaging cancellation based on the slow time dimension.

[0061] Since static or slow-moving clutter has a high correlation phase in the slow time dimension, dynamic targets can be highlighted by estimating and subtracting this quasi-static background component. and each channel The clutter component is estimated by averaging the current pulse and its surrounding signals, and then the estimated clutter component is subtracted from the original signal to obtain the clutter-suppressed signal. This operation is performed on the GPU in parallel for all distance units and all channels.

[0062] 2b) GPU-optimized high dynamic range windowing

[0063] Before Fourier transform, a Hamming window is applied to the data to reduce spectral leakage, thereby improving the ability to detect weak signals. The window function reduces the sidelobe level by smoothing the edges of the signal.

[0064] Range windowing: Combine the fast time series after clutter suppression with the range Hamming window function Multiplying point by point, we get .

[0065] Doppler windowing: After completing the range FFT, the slow time series Doppler Hamming window function Multiplying point by point, we get .

[0066] Both windowing operations are executed in parallel on the GPU via CUDA kernel functions.

[0067] 2c) GPU-accelerated high-resolution 2D Fourier transform

[0068] The time domain signal is converted to the range-Doppler frequency domain by performing FFTs in both the fast and slow time dimensions. To improve spectral resolution and facilitate the observation of micro-Doppler details, the data is zero-padded before the FFT.

[0069] Range FFT: Fast time series with range windowing Pad zero to length , for , and then perform FFT to obtain the distance frequency domain signal .

[0070] Doppler FFT: Doppler windowed slow time series Pad zero to length , for , and then perform FFT to obtain the range-Doppler domain signal , is the Doppler unit index, .

[0071] Both FFT steps are completed using the GPU acceleration library.

[0072] 2d) GPU parallel spectrum centralization

[0073] The standard FFT output places the zero-frequency component at the beginning of the spectrum. To facilitate observation and subsequent processing, the Doppler-dimensional spectrum is cyclically shifted so that the zero-Doppler frequency corresponds to the center of the spectrum.

[0074] For length Doppler spectrum The new Doppler index after centralization is achieved by moving the second half of the original spectrum to the first half and the first half to the second half. This rearrangement operation can be completed efficiently in parallel on the GPU. After completing all the above operations, for the currently processed Frame data, each receiving channel A complex valued, high-resolution range-Doppler spectrum will be obtained after clutter suppression, windowing, zero-filling, two-dimensional Fourier transform and spectrum centering. , is the Doppler unit index. This matrix is ​​stored in the GPU memory, and its amplitude and phase information are well preserved, laying a solid foundation for subsequent multi-channel fusion and gesture micro-Doppler feature analysis. Figure 3 , Figure 3 Schematic diagram of the range-Doppler spectrum of gesture data of a single receiving channel generated after processing in step 2 in an embodiment of the present invention.

[0075] Step 3: For the gesture data distance-Doppler spectrum of each channel, the enhanced gesture data distance-Doppler amplitude image is obtained by GPU-accelerated multi-channel information fusion.

[0076] This step aims to effectively fuse the range-Doppler information from multiple receiving channels in step 2 to generate a single range-Doppler image with a higher signal-to-noise ratio and more robust target indication in parallel on the GPU for subsequent gesture target detection and feature analysis. It should be noted that in the interferometric processing in step 1, the systematic and fixed phase differences between channels have already been calibrated through channel equalization. Therefore, this step is performed based on the basic phase alignment of the channel signals, aiming to further improve the signal-to-noise ratio while balancing signal fidelity and processing robustness.

[0077] First get all The range-Doppler complex matrix of the receiving channels obtained after step 2 ,in is the channel index, is the range gate index, is the Doppler unit index. Then GPU parallel amplitude accumulation is performed for each range-Doppler unit. , calculate the sum of the signal amplitudes on all channels and obtain the fused amplitude distance-Doppler map :

[0078]

[0079] in represents the modulus of a complex number. is the number of receiving channels. This accumulation process is performed on the GPU by designing a CUDA kernel function to calculate all the The units are executed in parallel, thus efficiently obtaining a single-frame enhanced range-Doppler amplitude map that integrates all channel information. This will be used as input for the next step 4. Figure 4 , Figure 4 This is a gesture data range-Doppler amplitude diagram with enhanced energy of a single-frame target echo signal obtained after processing in step 3 in an embodiment of the present invention.

[0080] Step 4: The enhanced gesture data range-Doppler amplitude image is identified by a two-dimensional adaptive operator in parallel on the GPU to obtain a gesture target detection result for micro-motion analysis.

[0081] 4a) Input data preparation and preprocessing

[0082] The cumulative amplitude value output from step 3 Perform scale conversion and convert to logarithmic value By calculation Come get ,in is a small positive constant to prevent taking logarithms of zero or negative values. This conversion operation is performed in parallel on the GPU using a dedicated CUDA kernel function.

[0083] 4b) Adaptive threshold calculation based on sliding window

[0084] This substep passes the input data A two-dimensional reference window is slid upward to calculate an adaptive detection threshold for each unit to be detected (CUT).

[0085] First, a reference window including a training window and a protection window is defined around the CUT, and the number of training units in the range and Doppler directions is preset. and , Number of protection units and Calculate the total number of training units Then the local background average level is estimated , by calculating all training units Sum of values And find the average value.

[0086] The detection threshold is obtained by subtracting the protection window (including CUT) from the total reference window. The local background average level estimated by Plus a probability that is equal to the expected false alarm probability ( ) related bias factors Obtain the bias factor Scaling factor by linear domain threshold By logarithmic transformation:

[0087]

[0088]

[0089] 4c) Target judgment and two-dimensional adaptive operator result graph generation

[0090] The logarithmic value of the unit CUT to be tested and the adaptive detection threshold calculated for it For comparison, if , it is judged as the target, otherwise it is judged as the background.

[0091] According to the judgment result, a two-dimensional adaptive operator output graph is generated If the unit If it is judged as a target, it will be given Its original value; otherwise, assign it zero.

[0092] The entire 2D adaptive operator detection process, from data preparation to final judgment, is executed in large-scale parallel on the GPU through efficient CUDA kernel functions, ensuring real-time processing. After processing is completed, a 2D adaptive operator detection result image is obtained in the GPU memory for the current data frame, clearly indicating the potential hand target location and its signal strength. Figure 5 , Figure 5 This is a diagram of the gesture target detection result obtained after processing in step 4 in an embodiment of the present invention.

[0093] Step 5: For the gesture target detection results, GPU information fusion and time-frequency analysis are performed to obtain the gesture target micro-Doppler characteristic spectrum that characterizes the fine motion characteristics of the gesture target.

[0094] 5a) Target range region selection and Doppler profile extraction per frame

[0095] In step 4, we have obtained the two-dimensional target detection result map after each data frame is processed by the two-dimensional adaptive operator. ,in is the frame index, is the range gate index, The core goal of step 5 is to use the GPU to integrate this target detection information from multiple consecutive data frames. By analyzing the time-varying Doppler characteristics of the hand target within a specific distance range, we can extract and generate a micro-Doppler feature spectrum that can accurately characterize the target's dynamic behavior.

[0096] This step first aggregates the data frames generated in step 4 for each The generated two-dimensional adaptive operator detection result image sequence ,These data are the basis for constructing the micro-Doppler spectrum.

[0097] In order to extract micro-Doppler features from a specific target area, it is necessary to first determine the distance range based on which the analysis is to be performed. The principle is as follows.

[0098] Multi-frame energy accumulation and target center distance positioning: First, the continuous multi-frame two-dimensional adaptive operator detection result map from step 4 is In Doplevi and frame dimensions Energy is accumulated on the surface to obtain a one-dimensional distance energy distribution profile. :

[0099]

[0100] The section It reflects the comprehensive distribution of the detected target energy at different range gates during the entire observation time.

[0101] Peak search and center distance determination: one-dimensional energy distribution profile Perform peak search to find the peak with the most concentrated energy. The range gate index corresponding to this peak is the center range gate of the potential target. .

[0102]

[0103] Adaptive range window selection based on energy threshold: to determine the center range gate As a benchmark, the boundary of the range window is adaptively determined. First, the center range gate At the beginning, search in the direction of increasing distance and the direction of decreasing distance respectively. During the search process, find the energy value that drops below the peak energy for the first time. We denote these two boundary range gate indices as and The final selected target distance area That is, all range gates between these two boundaries:

[0104]

[0105] This adaptive method can automatically adjust the range window width determined by the distance to the target according to its actual distance size or the range of range migration caused by movement, which is more accurate than using a fixed window width.

[0106] By using the above method, the distance area containing the main target energy can be adaptively determined , thus ensuring that the subsequent micro-Doppler analysis is performed on the target itself, eliminating the interference of irrelevant noise and clutter at other distances. , this step uses its two-dimensional adaptive operator to detect the result map , in the selected distance area For each Doppler unit The signal strength is accumulated to obtain the Doppler energy distribution profile of the frame in the specified distance area. . Reflected in the The accumulated energy of each Doppler element within a frame and a selected range region is calculated. This accumulation operation for each frame is efficiently performed in parallel on the GPU, accelerating the extraction of Doppler profiles.

[0107] 5b) Micro-Doppler characteristic spectrum construction

[0108] The Doppler energy profiles extracted from all data frames Arrange in chronological order to form a two-dimensional matrix In order to conform to the common display habits of micro-Doppler spectra, the matrix is ​​transposed to obtain the final micro-Doppler characteristic spectrum. .

[0109] The value of each element of the transposed matrix represents the time frame, the target is The energy intensity on a Doppler unit. This two-dimensional spectrum This clearly demonstrates the dynamic evolution of the hand's Doppler signature over time, providing a crucial basis for analyzing the target's fine motion. Matrix construction and transposition operations are also accelerated using the GPU's parallel computing capabilities.

[0110] Micro-Doppler characteristic spectrum It can be used for specific feature analysis and can also be used as feature input into machine learning models for tasks such as target recognition and classification. Figure 6 , Figure 6 Schematic diagram of micro-Doppler characteristics of a gesture target generated after processing in step 5 in an embodiment of the present invention.

[0111] Step 6: By performing feature analysis on the micro-Doppler characteristic spectrum of the gesture target, a quantitative micro-motion feature that can support fine recognition of the gesture target is obtained.

[0112] 6a) Spectral preprocessing and normalization

[0113] To ensure the robustness and accuracy of gesture feature extraction, the generated micro-Doppler feature spectrum is first Standardized preprocessing operations were performed, and a two-dimensional Gaussian filter was applied for smoothing to suppress random noise interference. At the same time, histogram equalization technology was used to enhance the contrast of the spectrum, so that the weak but discriminative micro-motion details could be highlighted.

[0114] 6b) Accurate quantification and extraction of key micro-motion parameters

[0115] Based on the preprocessed spectrogram, specific algorithms and analytical methods are used to quantitatively extract key micro-motion parameters, including: main frequency and harmonic analysis of the modulation features generated by periodic motion in the spectrogram; quantification of the period and frequency of periodic scintillation features by analyzing the time series of the energy envelope within a specific Doppler frequency band; determination of the maximum micro-Doppler frequency shift throughout the entire observation time; calculation of the instantaneous bandwidth occupied by the micro-Doppler signal on the frequency axis for each time frame; identification of the duration of significant micro-motion feature segments on the time axis by setting an energy threshold and combining connected domain analysis; extraction of the envelope shape parameters of the main micro-Doppler spectral line energy over time; and estimation of its model parameters such as amplitude, modulation frequency, initial phase, center frequency, slope, and intercept for the specific morphology presented in the spectrogram using curve fitting techniques.

[0116] 6c) Feature vector construction and optimization

[0117] All the extracted Quantified micro-motion parameters, including frequency, bandwidth, duration, amplitude and morphological parameters, are organized and constructed into a high-dimensional feature vector of fixed dimension. :

[0118]

[0119] For each component in the eigenvector Z-score normalization was performed:

[0120]

[0121] in and They are The mean and standard deviation of each feature on the training set are calculated. This operation eliminates the impact of differences in dimension and numerical range between different features. Based on the prior knowledge base and a deep understanding of gesture target characteristics, combined with feature importance assessment, this step selects a core subset of features from the high-dimensional feature vector that is most discriminative for the specific gesture target recognition task. Through this series of precise analysis and extraction steps, the resulting quantized micro-motion feature vector provides high-quality, information-rich input for the subsequent gesture target fine-grained recognition model.

[0122] In addition, these detailed micro-Doppler feature spectra and their quantitative characteristics have also opened up new avenues for the emerging field of contactless human-computer interaction, especially showing great potential in gesture recognition. By analyzing the unique micro-Doppler signal patterns generated by small, rapid hand movements, it is possible to accurately classify and recognize complex gestures, providing an intuitive and efficient means of interaction for scenarios such as smart homes, medical assistance, and virtual reality. The method of the present invention is used to extract micro-Doppler features for various gestures, see Figure 7 , which is the micro-Doppler characteristic spectrum of the waving gesture. Figure 8 , which is the micro-Doppler characteristic spectrum of the downward-swinging hand gesture. Figure 9 , which is the micro-Doppler characteristic spectrum of the check gesture. Figure 10 , which is the micro-Doppler characteristic spectrum of the open fist gesture. Figure 11 , the image is the characteristic spectrum of the left and right waving gesture.

[0123] To evaluate the performance of this method for extracting micro-Doppler features from gestures in millimeter-wave radar signals, the following tests further demonstrate its effectiveness. We compared it with a traditional CPU-based method using an NVIDIA GeForce GTX 40 series graphics card. Table 1 shows the runtimes of the two methods for the same data at different processing steps and overall. The experimental results are the average of 100 tests.

[0124] Table 1 Each step and overall test time

[0125] Processing Name CPU time (ms) GPU time (ms) Speedup Data preprocessing 11.28 1.19 9.479 Dynamic enhancement processing 411.76 103.36 3.984 Fusion Enhancement Processing 11.03 1.25 8.824 Target identification processing 1266.44 23.67 53.504 Feature extraction processing 23.17 4.52 5.126 Total duration 1723.68 133.99 12.864

[0126] In summary, the present invention, through GPU parallel acceleration, not only achieves an order of magnitude increase in millimeter-wave radar signal processing speed, but more importantly, it can ensure the accuracy and detail richness of gesture micro-Doppler feature extraction while maintaining high-speed processing. This successfully solves the problem of traditional methods that struggle to balance speed and precision. This significantly speeds up the previously time-consuming micro-Doppler feature extraction process, meeting the urgent need for real-time perception of target dynamic characteristics in fields such as autonomous driving, human-computer interaction, and security monitoring.

Claims

1. A method for extracting micro-Doppler features of gestures from millimeter-wave radar signals based on a GPU, the method comprising: Step 1: The multi-channel raw echoes of the millimeter-wave radar are collected and interpolated through GPU optimization to obtain a high-fidelity and phase-consistent multi-channel digital signal stream of gesture data. Step 2: Performing GPU parallel transformation and dynamic enhancement processing on the multi-channel digital signal stream of the gesture data to obtain the range-Doppler spectrum of the gesture data of each channel; Step 3: For the gesture data range-Doppler spectrum of each channel, GPU-accelerated multi-channel information fusion is performed to obtain an enhanced gesture data range-Doppler amplitude image; Step 4: The enhanced gesture data range-Doppler amplitude image is subjected to GPU parallel two-dimensional adaptive operator discrimination to obtain gesture target detection results that can be used for micro-motion analysis; Step 5: For the gesture target detection results, GPU information fusion and time-frequency analysis are performed to obtain a gesture target micro-Doppler characteristic spectrum representing the fine motion characteristics of the gesture target; Step 6: By performing feature analysis on the micro-Doppler characteristic spectrum of the gesture target, a quantitative micro-motion feature that can support fine recognition of the gesture target is obtained.

2. The method according to claim 1, characterized in that The phase interference processing in step 1 specifically includes: performing at least one of DC offset correction, I / Q (in-phase / quadrature) imbalance correction based on predetermined calibration coefficients (including relative gain error and phase mismatch), inter-pulse phase fluctuation correction based on phase error estimation, and channel equalization based on a complex gain correction factor in parallel on the GPU.

3. The method according to claim 1, characterized in that The GPU parallel transformation and dynamic enhancement processing in step 2 specifically includes: first, using an adaptive average cancellation method based on a slow time dimension to suppress clutter; then, applying a Hamming window function to the clutter-suppressed data in the range and Doppler directions respectively; finally, performing a zero-padding operation on the windowed data in the range and Doppler directions and performing a fast Fourier transform to obtain the range-Doppler spectrum.

4. The method according to claim 1, wherein The multi-channel information fusion in step 3 adopts the amplitude accumulation method, specifically: the modulus value of the complex signal of the corresponding unit in the range-Doppler spectrum of each channel is calculated in parallel on the GPU, and then the modulus values ​​of all channels are accumulated to obtain the enhanced gesture data range-Doppler amplitude image.

5. The method according to claim 1, characterized in that The two-dimensional adaptive operator discrimination in step 4 is a two-dimensional constant false alarm rate detection algorithm, which is executed in parallel on a GPU and specifically includes: performing logarithmic scaling conversion on the input range-Doppler amplitude image; defining a two-dimensional reference window including a training window and a protection window around the unit to be detected; estimating the local background level by averaging the unit values ​​within the training window; calculating an adaptive detection threshold based on the local background level and a preset false alarm probability; and comparing the value of the unit to be detected with the threshold to determine whether it is a target.

6. The method according to claim 1, characterized in that The specific method for obtaining the gesture target micro-Doppler characteristic spectrum in step 5 includes: first, accumulating the energy of the target detection results of multiple consecutive frames in the Doppler dimension and the frame dimension to obtain a one-dimensional range energy distribution profile; then, performing a peak search on the profile to determine the center range gate of the target; then, using the center range gate as a reference, adaptively determining a range window containing the main energy of the target by searching for a boundary where the energy is lower than a preset percentage of the peak energy; finally, for each frame, accumulating the detection values ​​of each Doppler unit within the range window only to form the Doppler profile of the frame, and then arranging the Doppler profiles of all frames in chronological order and transposing them to construct the micro-Doppler characteristic spectrum.

7. The method according to claim 1, characterized in that The quantitative micro-motion features obtained in step 6 include at least one of the following: the main frequency and its harmonic components obtained by Fourier transform analysis of the micro-Doppler characteristic spectrum; the period and frequency of the periodic scintillation feature obtained by autocorrelation function analysis; the maximum micro-Doppler frequency shift during the entire observation time; the instantaneous micro-Doppler bandwidth of each time frame; the duration of the micro-motion event identified by connected domain analysis; or the spectral line morphology parameters obtained by curve fitting.

Images