Radar Moving Target Detection Method Based on Online Tensor Robust Principal Component Analysis

Abstract

The present invention discloses a radar moving target detection method based on online tensor robust principal component analysis. This method uses the robust principal component analysis method to regard the stationary background part as a low-rank component and the moving target part as a component with spatio-temporal continuity, respectively construct their corresponding regularization functions, and use the alternating direction multiplier method to iteratively solve these two components, thereby realizing the separation of the stationary background and the moving target. When multiple frames of radar data are input through an online mechanism, the results are updated in real time to achieve the online processing ability. The present invention solves the physical problem that when only single-frame data is used for Doppler FFT, the moving target obscures the stationary background part, resulting in the omission of the background part or the misidentification of the background part as a stationary target.

Description

Technical Field

[0001] The present invention relates to the field of millimeter-wave radar, and in particular to a radar moving target detection method based on online tensor robust principal component analysis. Background Art

[0002] With the continuous reduction of the cost and the continuous decrease of the volume of millimeter-wave radar, millimeter-wave radar has gradually played an important role in fields such as wireless communication and intelligent driving. In the field of communication, the frequency band of millimeter-wave is already very close to the current communication system, and millimeter-wave radar will also gradually play an important role in the integration of sensing and communication, which is one of the core contents of 6G. In the field of intelligent driving, compared with lidar, millimeter-wave radar has a lower cost and can work normally under harsh weather and lighting conditions. Linear frequency modulation continuous wave is the most commonly used waveform in current millimeter-wave radar due to its easy control and sensitivity to range and Doppler information. However, when the existing mainstream linear frequency modulation continuous wave millimeter-wave radar separates the stationary background and moving targets, it simply relies on the Doppler FFT in single-frame data to separate the velocity information. When a moving target occludes the stationary background, two physical problems will occur: First, the occluded stationary background will not be detected. Second, when the moving target just starts to occlude the stationary background and is about to stop occluding, the electromagnetic wave that irradiates the stationary background and returns to the radar will carry additional Doppler frequency shift due to the influence of the moving target during the propagation process, and the Doppler FFT will incorrectly identify the stationary background as a moving target. Using single-frame data will inevitably not be able to overcome the above physical problems. Summary of the Invention

[0003] The purpose of the present invention is to provide a radar moving target detection method based on online tensor robust principal component analysis for the deficiencies of the prior art. Regarding multi-frame radar data as a time series, using the tensor robust principal component analysis method, and utilizing the relationship between the front and back frame data to overcome the problems of missed detection of the stationary background and misjudgment as a moving target when a moving target occludes the stationary background in the task of separating moving targets and stationary backgrounds; through an online mechanism, when multi-frame radar data is input, the result is updated in real time to achieve the online processing ability.

[0004] The purpose of the present invention is achieved through the following technical solutions: A radar moving target detection method based on online tensor robust principal component analysis, which is divided into an offline initialization stage and an online processing stage;

[0005] The offline initialization stage includes the following steps:

[0006] Step 1: Use a time-division multiplexing multiple-input multiple-output millimeter-wave frequency-modulated continuous-wave radar for sampling and receive multiple frames of data; perform range FFT, Doppler FFT, and angle FFT preprocessing on the three-dimensional tensor radar input signal received in each frame to obtain a three-dimensional tensor data;

[0007] Step 2: Perform Doppler processing on the preprocessed three-dimensional tensor data in each frame in the Doppler dimension. Specifically, output a slice with Doppler equal to 0 as the static initialization range-angle matrix, and average the multiple slices with non-zero Doppler in the Doppler dimension as the dynamic initialization range-angle matrix;

[0008] Step 3: Stack the static initialization range-angle matrix and the dynamic initialization range-angle matrix frame by frame respectively and take the modulus of the complex numbers to obtain the static initialization tensor and the dynamic initialization tensor;

[0009] Step 4: Use the static initialization tensor and the dynamic initialization tensor as the initialization input, use the sum of the two tensors as the decomposition target, and use the traditional offline tensor robust principal component analysis method to decompose to obtain the static part tensor and the dynamic part tensor;

[0010] After the offline initialization stage is completed, the online processing stage is carried out. For each newly received frame of data, the following steps are executed:

[0011] Step 1: Use a time-division multiplexing multiple-input multiple-output millimeter-wave frequency-modulated continuous-wave radar to online receive each newly sampled frame of data, and perform range FFT, Doppler FFT, and angle FFT preprocessing on the three-dimensional tensor radar input signal received in the current frame to obtain a three-dimensional tensor data;

[0012] Step 2: Perform Doppler processing on the preprocessed three-dimensional tensor data in the current frame in the Doppler dimension. Specifically, average the range-angle dimension slices in the Doppler dimension and take the modulus of the complex numbers to obtain the range-angle matrix to be decomposed in the current frame;

[0013] Step 3: When processing the first frame of data in the online processing stage, use the static part tensor and the dynamic part tensor obtained in the offline initialization stage as the input, use the range-angle matrix to be decomposed as the decomposition target, and use the online tensor robust principal component analysis method to obtain the static background and moving targets in the current frame;

[0014] When processing subsequent frames of data in the online processing stage, use the Tucker tensor decomposition result of the previous frame as the input, use the range-angle matrix to be decomposed as the decomposition target, and use the online tensor robust principal component analysis method to obtain the static background and moving targets in the current frame;

[0015] After processing each frame of data in the online processing stage, the Tucker tensor decomposition results obtained during the decomposition process are saved.

[0016] Furthermore, in the offline initialization stage and the online processing stage, for the range-angle matrix obtained after Doppler processing, multiple frames of data are stacked in the time dimension to obtain a three-dimensional tensor, and the physical problems that cannot be overcome by a single frame of data are solved by mining the correlations in the time dimension after constructing the tensor.

[0017] Furthermore, the stationary background part is regarded as a low-rank component, and the moving target part is regarded as a component with spatio-temporal continuity. Corresponding regularization functions are constructed for them respectively, and the alternating direction method of multipliers (ADMM) is used to iteratively solve these two components, so as to realize the separation of the stationary background and the moving target.

[0018] Furthermore, the stationary background part is regarded as a low-rank component, and its corresponding regularization function is an approximation of the low-rank part of the Tucker tensor decomposition.

[0019] Furthermore, the moving target part is regarded as a component with spatio-temporal continuity, and its corresponding regularization function is the three-dimensional total variation (3D Total Variation) function of the tensor formed by the range-angle matrices of the moving target in the current frame and the previous frame.

[0020] The beneficial effects of the present invention are as follows: The present invention uses multiple frames of data to accurately detect the stationary background even when the moving target temporarily blocks the stationary background, and at the same time, the situation of misidentifying the stationary background as a moving target will not occur. When multiple frames of radar data are input through an online mechanism, the results are updated in real time, realizing the online processing ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 is a flowchart for obtaining the range-angle matrix to be decomposed, which needs to be carried out in both the offline initialization stage and the online processing stage;

[0022] Figure 2 is a real scene diagram and a schematic diagram of the test scenario;

[0023] Figure 3 is the range-angle diagram of the three-dimensional tensor to be decomposed in the direction of the alignment angle inverse 1 in the test scenario;

[0024] Figure 4 is a comparison diagram of the results of separating the stationary background and the moving target from the data to be decomposed by the Doppler FFT in the online processing stage and the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0025] In order to better understand the technical solution of the present application, the embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0026] It should be clear that the described embodiments are only a part of the embodiments of the present application, rather than all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present application.

[0027] The terms used in the embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the present application. The singular forms "a", "the" and "said" used in the embodiments of the present application and the appended claims are also intended to include the plural forms, unless the context clearly indicates otherwise.

[0028] Sampling is performed using a time-division multiplexing (TDM) multiple-input multiple-output (MIMO) millimeter-wave frequency-modulated continuous-wave radar. In the TDM mode, the MIMO antenna array can be equivalent to a virtual uniform linear array, and the number of virtual elements is the number of virtual channels. A radar moving target detection method based on online tensor robust principal component analysis provided by the present invention is divided into an offline initialization stage and an online processing stage; the implementation processes of each stage are elaborated in detail below.

[0029] The offline initialization stage includes the following steps:

[0030] Step 1: As Figure 1 shown, the.bin data received from the radar can be processed and converted into T-frame data. Each frame of data is a three-dimensional tensor, and the three dimensions are the number of virtual channels, the number of chirps within a single frame, and the number of sampling points within a single chirp. Perform range FFT, Doppler FFT, and angle FFT preprocessing on the three-dimensional tensor radar input signal received for each frame to obtain a three-dimensional tensor data, and the three dimensions of this tensor are the range FFT length, the Doppler FFT length, and the angle FFT length;

[0031] Step 2: Perform Doppler processing on the three-dimensional tensor data of each frame obtained in Step 1 in the Doppler dimension; the Doppler processing in the offline initialization stage and the online processing stage is different; the Doppler processing in this step is: output a slice with Doppler equal to 0 as the static initialization range-angle matrix, and average the multiple slices with non-zero Doppler in the Doppler dimension as the dynamic initialization range-angle matrix;

[0032] Step 3: Stack the static initialization range-angle matrix and the dynamic initialization range-angle matrix frame by frame respectively and take the modulus of the complex numbers to obtain the static initialization tensor and the dynamic initialization tensor;

[0033] Step 4: Use the static initialization tensor and the dynamic initialization tensor as the initialization inputs, take the sum of the two tensors as the decomposition target, and decompose using the traditional offline tensor robust principal component analysis method to obtain the static part tensor S T,T-1 and the dynamic part tensor D T,T-1 , where the subscript (T, T - 1) represents the tensor formed by stacking the T-th frame matrix and the previous T - 1 frame matrices.

[0034] After the offline initialization phase is completed, the online processing phase is carried out. For each new frame of data received, the following steps are executed:

[0035] Step 1: As Figure 1 shown, use a time-division multiplexing multiple-input multiple-output millimeter-wave frequency-modulated continuous-wave radar to online receive each new sampled data frame, and perform range FFT, Doppler FFT, and angle FFT preprocessing on the three-dimensional tensor radar input signal received in the current frame t (t > T) to obtain a three-dimensional tensor data;

[0036] Step 2: Perform Doppler processing on the preprocessed three-dimensional tensor data in the Doppler dimension in the t-th frame. The specific Doppler processing in this step is: take the average of the range-angle dimension slices in the Doppler dimension and take the modulus of the complex numbers to obtain the distance-angle matrix X to be decomposed in the t-th frame t ;

[0037] Step 3: When processing the first frame of data in the online processing phase, that is, when t = T + 1, use the static part tensor S T,T-1 and the dynamic part tensor D T,T-1 obtained in the offline initialization phase as inputs, take the distance-angle matrix X t to be decomposed as the decomposition target, and use the online tensor robust principal component analysis method to obtain the stationary background S t and the moving target D t in the t-th frame;

[0038] When processing subsequent frames of data in the online processing phase, that is, when t > T + 1, use the Tucker tensor decomposition results of the previous frame: Ψ t-1,T-1 , U t-1,T-1 , V t-1,T-1 , W t-1,T-1 as inputs, and combine with the previously obtained stationary background tensor S t-1,T-1 and the moving target matrix D t-1 , take the distance-angle matrix X t to be decomposed as the decomposition target, and use the online tensor robust principal component analysis method based on Tucker tensor decomposition to obtain the stationary background S t and the moving target D t in the t-th frame, thereby realizing the separation of the stationary background and the moving target in this frame of data.

[0039] In one embodiment, the static background S t and the moving target D t are solved in the following manner:

[0040] 1) Using the Tucker tensor decomposition results Ψ t-1,T-1 , U t-1,T-1 , V t-1,T-1 and W t-1,T-1 as low-rank prior information, the static background matrix S t of the t-th frame is obtained, and stacked with the previously saved static background tensors S t-1,T-2 of the first T - 1 frames to form S t,T-1 ;

[0041] 2) Convert the tensor decomposition into the solution of the following optimization problem:

[0042]

[0043] where the constraint function for solving the part representing the static background is f s () based on Tucker tensor decomposition:

[0044] S t,T-1 ≈ Ψ t,T-1 ×1U t,T-1 ×2V t,T-1 ×3W t,T-1 ,

[0045]

[0046] The constraint function for solving the part representing the moving target is f d () based on 3D Total Variation:

[0047] f d (D t ) = ||D t,1 || TV ,

[0048] 3) Use the Alternating Direction Method of Multipliers (ADMM) to iteratively solve for S t ;

[0049] 4) Use ADMM to iteratively solve for D t .

[0050] Figure 2It is the actual scene diagram and schematic diagram of the test scenario. In the darkroom, 3 corner reflectors and 1 metal cuboid are distributed as the static background. There is 1 pedestrian walking along the pedestrian walkway as the moving target. During the forward movement of the pedestrian, corner reflector 1 will be blocked. After the data collected by the radar in this scenario is processed, the distance-angle matrix to be decomposed is stacked frame by frame into a three-dimensional tensor, and the distance and frame number dimension slices are as follows Figure 3 shown, and the angle of this tensor slice is the direction aligned with corner reflector 1. Figure 4 It can be seen that in the vicinity of the 30th frame of the Doppler FFT, that is, during the time period when the pedestrian blocks corner reflector 1, corner reflector 1 cannot be detected in the static background part. At the same time, when the pedestrian just starts to block corner reflector 1 and is about to stop blocking, the electromagnetic wave irradiating corner reflector 1 and returning to the radar will carry additional Doppler frequency shift during the propagation process due to the influence of the pedestrian, and the stationary corner reflector 1 will be wrongly identified as a moving target. However, the method of the present invention does not have the above problems, and corner reflector 1 can still be detected when the pedestrian blocks it, and corner reflector 1 will not be wrongly judged as a moving target either.

[0051] The above are only the preferred embodiments of one or more embodiments of this specification, and are not intended to limit one or more embodiments of this specification. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of one or more embodiments of this specification shall be included within the scope protected by one or more embodiments of this specification.

Claims

1. A radar moving target detection method based on online tensor robust principal component analysis, characterized in that, It is divided into an offline initialization stage and an online processing stage; The offline initialization stage includes the following steps: Step 1: Use a time-division multiplexing multiple-input multiple-output millimeter-wave frequency-modulated continuous-wave radar for sampling and receive multiple frames of data; perform range FFT, Doppler FFT, and angle FFT preprocessing on the three-dimensional tensor radar input signal received in each frame to obtain a three-dimensional tensor data; Step 2: Perform Doppler processing on the preprocessed three-dimensional tensor data in each frame in the Doppler dimension. Specifically: Output a slice with Doppler equal to 0 as the static initialization range-angle matrix, and average the multiple slices with non-zero Doppler in the Doppler dimension as the dynamic initialization range-angle matrix; Step 3: Stack the static initialization range-angle matrix and the dynamic initialization range-angle matrix frame by frame respectively and take the modulus of the complex numbers to obtain the static initialization tensor and the dynamic initialization tensor; Step 4: Use the static initialization tensor and the dynamic initialization tensor as the initialization input, use the sum of the two tensors as the decomposition target, and use the offline tensor robust principal component analysis method to decompose to obtain the static part tensor and the dynamic part tensor; After the offline initialization stage is completed, the online processing stage is carried out. For each new frame of data received, execute the following steps: Step 1: Use the time-division multiplexing multiple-input multiple-output millimeter-wave frequency-modulated continuous-wave radar to online receive each new sampled data, and perform range FFT, Doppler FFT, and angle FFT preprocessing on the three-dimensional tensor radar input signal received in the current frame to obtain a three-dimensional tensor data; Step 2: Perform Doppler processing on the preprocessed three-dimensional tensor data in the current frame in the Doppler dimension. Specifically: Average the range-angle dimension slices in the Doppler dimension and take the modulus of the complex numbers to obtain the range-angle matrix to be decomposed in the current frame; Step 3: When processing the first frame of data in the online processing stage, use the static part tensor and the dynamic part tensor obtained in the offline initialization stage as the input, use the range-angle matrix to be decomposed as the decomposition target, and use the online tensor robust principal component analysis method to obtain the static background and moving targets in the current frame; When processing subsequent frames of data in the online processing stage, use the Tucker tensor decomposition result of the previous frame as the input, use the range-angle matrix to be decomposed as the decomposition target, and use the online tensor robust principal component analysis method to obtain the static background and moving targets in the current frame; After processing each frame of data in the online processing stage, save the Tucker tensor decomposition result obtained during the decomposition process.

2. The radar moving target detection method based on online tensor robust principal component analysis according to claim 1, characterized in that In the offline initialization stage and the online processing stage, for the range-angle matrix obtained after Doppler processing, stack multiple frames of data in the time dimension to obtain a three-dimensional tensor. The physical problems that cannot be overcome by single-frame data are solved by mining the correlation in the time dimension after constructing the tensor.

3. A radar moving target detection method based on online tensor robust principal component analysis according to claim 1, characterized in that Regard the static background part as the low-rank component and the moving target part as the component with spatio-temporal continuity, construct their corresponding regularization functions respectively, and use the alternating direction multiplier method to iteratively solve these two components, so as to realize the separation of the static background and the moving target.

4. A radar moving target detection method based on online tensor robust principal component analysis according to claim 3, characterized in that, The static background part is regarded as a low-rank component, and its corresponding regularization function is an approximation of the low-rank part of the Tucker tensor decomposition.

5. The radar moving target detection method based on online tensor robust principal component analysis according to claim 3, characterized in that, The moving target part is regarded as a component with spatio-temporal continuity, and its corresponding regularization function is the three-dimensional total variation function of the tensor formed by the distance-angle matrices of the moving targets in the current frame and the previous frame.

Images