A micro-doppler target discrimination method based on radar spatial difference fusion enhancement

Abstract

This application discloses a micro-Doppler target discrimination method based on radar spatial difference fusion enhancement, comprising: S1. Data acquisition and reconstruction; S2. Range and velocity feature extraction; S3. Cross-channel global normalization processing; S4. Spatial difference feature extraction; S5. Multi-dimensional feature tensor fusion stacking; S6. Target result discrimination. This application captures energy differences from different spatial perspectives, expanding one-dimensional intensity information into multi-dimensional spatial semantic features, ensuring rich feature dimensions. The spatial diversity information of this application significantly enhances the discriminability of targets with different physiological structures, ensuring high recognition accuracy.

Description

Technical Field

[0001] This application belongs to the field of radar signal processing technology, specifically relating to a micro-Doppler target discrimination method based on radar spatial difference fusion enhancement. Background Technology

[0002] Traditional human micro-Doppler recognition methods are mostly based on single-channel data processing, neglecting the phase and energy decorrelation characteristics captured by multiple-transmitter multiple-receiver (MIMO) radar from different spatial observation perspectives. This leads to insufficient utilization of complex gait information by the system, and limited recognition accuracy in non-line-of-sight or interference environments. Summary of the Invention

[0003] This application provides a micro-Doppler target discrimination method based on radar spatial difference fusion enhancement to solve the above-mentioned technical problems.

[0004] To address the aforementioned technical problems, this application adopts the following technical solution: a micro-Doppler target discrimination method based on radar spatial difference fusion enhancement, comprising:

[0005] S1. Based on the multi-channel raw echo signal of the multi-antenna radar system, the raw data is analyzed and reconstructed into a multi-dimensional raw echo matrix containing the dimensions of frames, pulses and sampling points;

[0006] S2. Perform range transformation on the multidimensional original echo matrix of each channel to obtain target spatial distance information; then simultaneously perform velocity-to-Doppler transformation and energy accumulation processing to generate the original micro-Doppler energy spectrum of each channel reflecting different spatial observation perspectives;

[0007] S3. Traverse the original micro-Doppler energy spectrum of all channels, obtain the global maximum energy value in the entire spatiotemporal range, and perform linear normalization scaling on the energy intensity of each channel to generate normalized intensity features;

[0008] S4. Based on the automatic selection mechanism, the channel with the most stable target motion center of mass energy is locked as the reference observation view channel, and the difference components between the reference observation view channel and other different spatial view channels are calculated.

[0009] S5. Stack the intensity features of the baseline observation channel with the extracted multiple spatial difference components along the dimensional direction to construct a multimodal feature tensor;

[0010] S6. Input the multidimensional feature tensor into the preset discrimination model, and use the feature complementarity brought about by spatial subsets to complete the target classification and discrimination.

[0011] Furthermore, the method in step S1 includes:

[0012] S11. Based on a multi-transmitter, multi-receiver radar system, acquire the raw echoes from multiple observation channels;

[0013] S12. Analyze the original echo and reconstruct it into a radar data cube containing the dimensions of frames, pulses and sampling points, providing a foundation for subsequent extraction of multi-view features.

[0014] Furthermore, the method in step S2 includes:

[0015] S21. Spatial distance distribution is extracted by distance fast Fourier transform, and static clutter in the indoor environment is filtered out by pulse dimension mean coherent cancellation technique;

[0016] S22. Perform Doppler frequency transformation and energy accumulation to synchronously generate the original micro-Doppler energy spectrum that reflects the evolution characteristics of the target motion over time.

[0017] Furthermore, the method in step S4 includes:

[0018] S41. Select the channel with the most stable energy feedback from multiple channels as the benchmark observation perspective;

[0019] S42. Calculate the difference components between the original microDoppler energy spectrum and other channels with different observation perspectives.

[0020] Furthermore, the method in step S5 includes:

[0021] S51. Stack the normalized baseline channel intensity features and the extracted multiple spatial difference components along the dimensional direction to construct a multimodal feature tensor containing intensity information and spatial semantic information.

[0022] S52. The channel dimension of the feature tensor is dynamically expanded according to the actual scale of the radar array antenna, and the radar hardware system of different scales is adapted by adding additional observation view difference terms.

[0023] S53. By organically integrating the differential terms of multiple observation perspectives after dynamic expansion with the baseline intensity component, a high-dimensional spatiotemporal-frequency joint feature characterizing the gait details of the target is constructed.

[0024] Furthermore, the method in step S6 includes:

[0025] S61. Input the synthetic feature map containing the strength base and multiple spatial difference information into the discrimination model as the core input tensor for target classification and discrimination;

[0026] S62. The discriminant model uses the spatial phase change information contained in the multidimensional tensor to extract motion semantic features that reflect the evolution of the target's gait;

[0027] S63. By utilizing the complementary nature of intensity features and spatial components in the synthetic spectrum, the discriminability of targets with different physical forms can be enhanced, enabling accurate classification and identification of targets with different physiological structures, such as humans and animals.

[0028] The beneficial effects of this application are: by capturing energy differences from different spatial perspectives, this application expands one-dimensional intensity information into multi-dimensional spatial semantic features, ensuring rich feature dimensions; the spatial diversity information of this application significantly enhances the recognizability of targets with different physiological structures, ensuring a high recognition accuracy. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating an embodiment of the micro-Doppler target discrimination method based on radar spatial difference fusion enhancement of this application;

[0030] Figure 2 This is a normalized single-channel original micro-Doppler image of an embodiment of the micro-Doppler target discrimination method based on radar spatial difference fusion enhancement of this application;

[0031] Figure 3 This is a synthesized multidimensional feature fusion map of an embodiment of the micro-Doppler target discrimination method based on radar spatial difference fusion enhancement of this application. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0033] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0034] The specific embodiments described in this application all use a 60GHz three-transmit four-receive (MIMO) millimeter-wave radar as the hardware platform.

[0035] See Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the micro-Doppler target discrimination method based on radar spatial difference fusion enhancement according to this application. The method includes:

[0036] S1. Data Acquisition and Reconstruction: Acquire the multi-channel raw echo signals of the multi-antenna radar system, and reconstruct them into a multi-dimensional raw echo matrix containing the dimensions of frames, pulses, and sampling points by parsing the raw data.

[0037] Specifically, the method in step S1 includes:

[0038] S11. Multi-channel signal synchronous acquisition: The raw echoes of multiple observation channels are acquired using a multiple transmit multiple receive (MIMO) radar system.

[0039] S12. Multidimensional Tensor Reconstruction and Transpose: The original data is reconstructed into a radar data cube containing the dimensions of frames, pulses, and sampling points using a data parsing program, providing a foundation for subsequent extraction of multi-view features.

[0040] S2. Range and velocity feature extraction: Perform range transformation on the original echo matrix of each channel to obtain target spatial distance information and filter out environmental static clutter. Then, perform velocity-to-Doppler transformation and energy accumulation processing simultaneously to generate the original micro-Doppler energy spectrum of each channel reflecting different spatial observation perspectives.

[0041] Specifically, the method in step S2 includes:

[0042] S21. Range Transform and Clutter Suppression: Spatial range distribution is extracted using the Fast Fourier Transform of Range (FFT), and static clutter in the indoor environment is filtered out using pulse-dimensional mean coherent cancellation technology. The formula is as follows:

[0043] (1);

[0044] in, The original signal, This represents the total number of pulses within a single frame.

[0045] S22. Velocity Transformation and Spectrum Generation: Perform Doppler frequency transformation and energy accumulation to simultaneously generate a raw micro-Doppler energy spectrum that reflects the target's motion evolution over time. .

[0046] S3. Cross-channel global normalization processing: Traverse all channel spectra and obtain the global maximum energy value in the entire spatiotemporal range; use this maximum value as a uniform scale to linearly normalize and scale the energy intensity of each channel to generate normalized intensity features.

[0047] Specifically, the method in step S3 includes:

[0048] S31. Global Peak Retrieval: Simultaneously traverse the energy maps of all observation channels to obtain the global maximum energy amplitude across the entire spatiotemporal range. ;

[0049] S32. Gain Difference Preservation: Using the global maximum amplitude. Based on this, each channel is scaled linearly and proportionally, as shown in the following formula:

[0050] (2);

[0051] in, For the first The original logarithmic energy matrix of the channel, This serves as the noise floor. The process ensures that the original spatial gain and intensity characteristics, which vary due to antenna mounting location and target scattering properties, are preserved during normalization.

[0052] S4. Spatial difference feature extraction: The channel with the most stable energy of the target's center of mass is locked as the benchmark observation channel through an automated selection mechanism, and the difference components between the benchmark channel and other different spatial perspective channels are calculated.

[0053] Specifically, the method in step S4 includes:

[0054] S41. Reference Viewpoint Locking: Lock the channel with the most stable energy feedback (such as TX1) from multiple channels as the reference viewpoint. ;

[0055] S42. Spatial Correlation Feature Extraction: Calculating the baseline channel and other spatial correlation features The difference components between channels from different spatial perspectives ;

[0056] (3);

[0057] This component is used to characterize the spatial correlation characteristics of target micro-motion at different detection angles.

[0058] S5. Multi-dimensional feature tensor fusion and stacking: The intensity features of the baseline channel and the extracted multiple spatial difference components are stacked along the dimensional direction to construct a multi-modal feature tensor.

[0059] Specifically, the method in step S5 includes:

[0060] S51. Construction of Multimodal Feature Tensor: Constructing a multimodal feature tensor The calculation method is as follows:

[0061] (4);

[0062] S52. Dynamic expansion of channel dimension: the feature tensor The channel dimension is dynamically expanded according to the actual size of the radar array antenna by adding additional observation view difference terms to the tensor to adapt to radar hardware systems of different sizes.

[0063] S53. Construction of high-dimensional joint features: By organically fusing the difference terms of multiple observation perspectives after dynamic expansion with the reference intensity component, a high-dimensional spatiotemporal-frequency joint feature characterizing the gait details of the target is constructed for subsequent semantic extraction and discrimination.

[0064] S6. Target Result Discrimination. Input the multidimensional feature tensor into the preset discrimination model, and use the feature complementarity brought about by spatial subsets to complete the target classification and discrimination.

[0065] Specifically, the method in step S6 includes:

[0066] S61. Multidimensional synthetic feature input: The synthetic feature map containing intensity base and multiple spatial difference information is input into the discrimination model as the core input tensor for target classification and discrimination.

[0067] S62. Motion semantic feature extraction: The discrimination model uses the spatial phase change information contained in the multidimensional tensor to extract motion semantic features that reflect the evolution law of the target gait.

[0068] S63. Multi-view complementary classification and recognition: By utilizing the feature complementarity between intensity features and spatial components in the synthetic spectrum, the recognition of targets with different physical forms is enhanced, and accurate classification and recognition of targets with different physiological structures such as humans and animals is achieved.

[0069] Example 1

[0070] This embodiment focuses on illustrating the general processing flow of the algorithm and its performance in extracting complex gait features.

[0071] Step 1: Data Acquisition and Reconstruction

[0072] Radar data acquisition was conducted in a typical laboratory setting. This embodiment uses a 60GHz millimeter-wave radar as the hardware platform, with a frame rate of 20Hz and a single acquisition duration of 500 frames. The acquired raw echo signals were analyzed and reconstructed into a three-dimensional raw matrix containing the dimensions of frames, pulses, and sampling points.

[0073] Step 2: Range Dimension Processing and Clutter Suppression

[0074] Windowing is applied to the sampling point dimension to suppress sidelobes, followed by range FFT to obtain the spatial distance information of the target. To effectively filter out static clutter in the indoor environment (such as reflections from walls and furniture), DC component cancellation is performed using pulse dimension mean subtraction.

[0075] Step 3: Velocity Dimension Processing and Atlas Generation

[0076] See Figure 2 The processed signal is subjected to Doppler FFT in the pulse dimension, and energy is accumulated along the distance dimension. Noise is suppressed by median background removal and logarithmic scaling is performed to generate the original micro-Doppler energy maps for each observation channel.

[0077] Step 4: Baseline Locking and Global Normalization

[0078] An automated selection mechanism is used to identify the channel with the most stable target center of mass energy as the reference viewpoint. To preserve the original energy solution correlation characteristics caused by the spatial arrangement of antennas from different observation viewpoints, this method extracts the global maximum energy value of all channels and performs linear normalization.

[0079] Step 5: Spatial Difference Fusion and Stacking

[0080] See Figure 3 The absolute energy difference component between the reference intensity characteristic and the other channels is calculated to characterize the spatial phase or energy difference of the target micro-motion at different detection angles.

[0081] Step Six: Verification of the Judgment Results

[0082] The constructed multimodal feature matrix was input into the deep learning model for testing. Experimental results show that when using only single-channel data, the system's accuracy in recognizing the human body is 91.50%; however, after applying the spatial difference fusion enhancement method described in this invention, the final recognition accuracy jumps to 98.67%.

[0083] Example 2

[0084] This embodiment aims to verify the universality of the present invention in multi-target recognition and cross-category feature extraction.

[0085] Step 1: Animal Target Data Collection

[0086] The radar equipment was deployed indoors to collect data on the test target (a cat). Considering the intermittent nature of animal movement, this embodiment introduced an energy detection mechanism to eliminate stationary periods, retaining only valid data when the target was in motion, accumulating approximately 20,000 frames of raw echoes.

[0087] Step 2: Cross-channel feature fusion application

[0088] The collected cat micro-Doppler data was processed using the same workflow as in Example 1. First, logarithmic energy maps of each channel were generated. Then, through global normalization and spatial difference calculation, a multimodal feature matrix containing the correlation characteristics of cat gait features and spatial features was constructed.

[0089] Because of the fundamental differences between cats and humans in terms of body size and gait frequency, this difference in energy characteristics from a spatial perspective presents a unique feature mapping in the fusion map.

[0090] Step 3: Multi-target joint training and recognition

[0091] The processed cat target feature data and the human target data from Example 1 were sliced ​​into 80-frame segments and then fed into a deep learning model for joint training. Experimental results show that the model achieved 100% accuracy in recognizing both human and cat targets.

[0092] Step 4: Experimental Conclusion

[0093] This embodiment demonstrates that, because the method of the present invention can significantly capture the essential differences in spatial observation perspective and motion semantics of targets with different physiological structures, it exhibits extremely high discrimination reliability and classification robustness in target recognition tasks with large morphological differences.

[0094] In summary, the micro-Doppler target discrimination method proposed in this application, based on multi-antenna radar spatial difference fusion enhancement, significantly enhances the target's feature representation capability in three-dimensional space by utilizing multi-view spatial complementarity.

[0095] (1) Significance of spatial fusion: Data from Example 1 shows that by constructing a multimodal feature matrix through stacking spatial difference features, the target recognition accuracy further increased from 91.5% of single-channel enhanced data to 98.67%, which fully demonstrates the key role of spatial diversity information in improving recognition stability in complex environments.

[0096] (2) Multi-target category discrimination ability: In Example 2, the recognition accuracy reached 100% through joint training on targets with different physiological structures such as humans and cats, which verified the unique advantages of the present invention in capturing the phase and energy decorrelation characteristics of humans and animals under different spatial observation perspectives.

[0097] (3) High real-time performance and scalability: The technical solution supports multi-channel parallel processing, and the dimension of the feature matrix can be flexibly expanded according to the size of the radar array antenna, which can be adapted to various MIMO radar systems and meet the real-time monitoring requirements.

[0098] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for micro-Doppler target discrimination based on radar spatial diversity fusion enhancement, characterized in that, include: S1. Based on the multi-channel raw echo signal of the multi-antenna radar system, the raw data is analyzed and reconstructed into a multi-dimensional raw echo matrix containing the dimensions of frames, pulses and sampling points; S2. Perform range transformation on the multidimensional original echo matrix of each channel to obtain target spatial distance information; Subsequently, a velocity-to-Doppler transformation is performed synchronously, and energy accumulation processing is carried out to generate the original micro-Doppler energy spectrum of each channel reflecting different space observation perspectives; S3. Traverse the original micro-Doppler energy spectrum of all channels, obtain the global maximum energy value in the entire spatiotemporal range, and perform linear normalization scaling on the energy intensity of each channel to generate normalized intensity features; S4. Based on the automatic selection mechanism, the channel with the most stable target motion center of mass energy is locked as the reference observation view channel, and the difference components between the reference observation view channel and other different spatial view channels are calculated. S5. Stack the intensity features of the reference observation view channel and the extracted multiple spatial difference components along the dimensional direction to construct a multimodal feature tensor; S6. Input the multidimensional feature tensor into a preset discrimination model, and use the feature complementarity brought about by spatial diversity to complete the target classification and discrimination.

2. The method of claim 1, wherein, The method of step S1 includes: S11. Based on a multi-transmitter, multi-receiver radar system, acquire the raw echoes from multiple observation channels; S12. Analyze the original echo and reconstruct it into a radar data cube containing the dimensions of frames, pulses and sampling points, providing a basis for subsequent extraction of multi-view features.

3. The method of claim 1, wherein, The method of step S2 includes: S21. Spatial distance distribution is extracted by distance fast Fourier transform, and static clutter in the indoor environment is filtered out by pulse dimension mean coherent cancellation technique; S22. Perform Doppler frequency transformation and energy accumulation to synchronously generate the original micro-Doppler energy spectrum that reflects the evolution characteristics of the target motion over time.

4. The method of claim 1, wherein, The method of step S3 includes: S31. Global Peak Retrieval: Simultaneously traverse the original micro-Doppler energy spectrum of all observation channels to obtain the global maximum energy amplitude across the entire spatiotemporal range; S32. Perform equal-proportional linear scaling on each channel based on the global maximum amplitude, generate normalized intensity features while mapping the data to a unified range, and ensure that the original spatial gain and intensity feature differences caused by the antenna physical location and target scattering characteristics are preserved.

5. The method of claim 1, wherein, The method of step S4 includes: S41. Select the channel with the most stable energy feedback from multiple channels as the benchmark observation perspective; S42. Calculate the difference components between the original microDoppler energy spectrum and other channels with different observation perspectives.

6. The method of claim 1, wherein, The method of step S5 includes: S51. Stack the normalized baseline channel intensity features and the extracted multiple spatial difference components along the dimensional direction to construct a multimodal feature tensor containing intensity information and spatial semantic information. S52. The channel dimension of the feature tensor is dynamically expanded according to the actual scale of the radar array antenna, and the radar hardware system of different scales is adapted by adding additional observation view difference terms. S53. By organically integrating the differential terms of multiple observation perspectives after dynamic expansion with the baseline intensity component, a high-dimensional spatiotemporal-frequency joint feature characterizing the gait details of the target is constructed.

7. The method of claim 1, wherein, The method of step S6 includes: S61. Input the synthetic feature map containing the strength base and multiple spatial difference information into the discrimination model as the core input tensor for target classification and discrimination; S62. The discrimination model uses the spatial phase change information contained in the multidimensional tensor to extract motion semantic features that reflect the evolution law of the target's gait; S63. By utilizing the complementary nature of the intensity features and spatial components in the synthesized spectrum, the discriminability of targets with different physical forms is enhanced, thereby achieving accurate classification and identification of targets with different physiological structures, such as humans and animals.

Images