Antenna phase center offset multi-stream asynchronous parallel clutter suppression method based on GPU

Abstract

The invention discloses an antenna phase center offset multi-stream asynchronous parallel clutter suppression method based on a GPU. A data stream is bound for the echo data of each channel, each data stream is responsible for data transmission and calculation processing of the bound channel, and asynchronous transmission and calculation between different channels are achieved. And the multi-pulse parallel compression operation is realized by using a fast Fourier transform batch processing technology, so that the time consumption of the algorithm is further reduced. In addition, real-time clutter suppression processing of echo data is realized through asynchronous parallel processing of registration and cancellation of data between adjacent channels. The method effectively improves the processing efficiency of an antenna phase center offset clutter suppression algorithm, and has certain robustness for different radar echo data.

Description

Technical Field

[0001] The invention belongs to the technical field of through-wall radar detection, and in particular relates to a GPU-based antenna phase center offset multi-stream asynchronous parallel clutter suppression method. Background Art

[0002] Through-the-wall radar can detect moving targets in a closed environment and obtain information about moving targets behind the wall, and its demand in urban perception is increasing. However, in the case that the posture and position of the moving targets behind the wall may change at any time, the system needs to have real-time processing capabilities to obtain the true status of the target.

[0003] In the field of through-wall radar detection, the ability to suppress clutter directly determines the performance of detecting targets behind walls. The antenna phase center offset method is a classic clutter suppression method, which is often used in the field of moving target detection on mobile platforms equipped with through-wall radars. However, in the actual detection process of through-wall radars, multiple sets of multi-channel radar echo data need to be processed in a short period of time, and the traditional antenna phase center offset method cannot meet the real-time requirements. Summary of the invention

[0004] The present invention proposes a GPU-based antenna phase center offset multi-stream asynchronous parallel clutter suppression method to solve the problem that the traditional antenna phase center offset method cannot process multiple groups of multi-channel radar echo data in a short time and cannot meet the real-time requirements. It can use the characteristics of GPU to realize parallel processing of echo data of different channels; at the same time, through batch processing design, it can realize parallel processing of echo data of the same channel, thereby improving the processing efficiency of the algorithm.

[0005] The technical solution of the present invention is as follows:

[0006] A GPU-based antenna phase center offset multi-stream asynchronous parallel clutter suppression method comprises the following steps:

[0007] Step 1. Create a number of Stream data streams on the GPU device that is the same as the number of channels;

[0008] Step 2: Map the data stream to the channel echo data, that is, each Stream data stream is bound to one of the channels and is responsible for processing the echo number of the channel;

[0009] Step 3: Apply for GPU memory resources by using the cudaMalloc instruction in the CUDA instruction set architecture to store multi-channel radar echo data, and use the cudaMemcpyAsync instruction to asynchronously transfer data from the CPU device to the GPU device according to different data streams to achieve data transmission;

[0010] Step 4, performing asynchronous parallel pulse compression on each channel data;

[0011] Step 5: Design a CUDA operator for registration to implement phase compensation for each frame of echo data in each channel in parallel, i.e., registration between channels;

[0012] Step 6: Design a cancellation CUDA operator to achieve cancellation between adjacent channels and obtain multi-channel data after clutter suppression.

[0013] Furthermore, the number of channels in step 1 specifically refers to: when using an antenna array to detect a moving target behind a wall, the antenna array has both a transmitting antenna and a receiving antenna, and the waveform transceiver channel between a transmitting antenna and a receiving antenna is called a channel. The number of channels is: the number of receiving antennas * the number of transmitting antennas. The Stream data stream can be applied through cudaStream_t in the CUDA architecture and created using cudaStreamCreate.

[0014] Furthermore, in step 4, the specific method of pulse compression is: the complex conjugate of the signal can be calculated according to the parameters of the radar, and then convolved with the echo signal, which is also called matched filtering. A compressed pulse can be obtained through matched filtering, which can not only improve the resolution, but also help to improve the signal-to-noise ratio of the echo signal. By combining the pulse compression algorithm logic with the parallel architecture characteristics of the GPU, a CUDA operator is designed to achieve parallel pulse compression.

[0015] Furthermore, the step 4 includes the following sub-steps:

[0016] Step 4.1, by using cufftHandle to create an FFT plan, and by calling the cufftPlanMany and cufftExecZ2Z API interfaces, parallel batch fast Fourier transform of each channel data is implemented to achieve the transformation of signal data from time domain to frequency domain;

[0017] Step 4.2: Generate a reference signal according to the radar antenna parameters and convolve it with the echo signal;

[0018] Step 4.3. Create an IFFT plan using cufftHandle and call the cufftPlanMany and cufftExecZ2Z API interfaces to implement parallel batch fast inverse Fourier transform of each channel data and transform the signal data from the frequency domain to the time domain.

[0019] Furthermore, in step 5, the specific implementation method of the registration CUDA operator is as follows: each frame of radar echo is a continuous sinusoidal waveform. Due to the limitation of hardware conditions, it is impossible to record all the data at each moment. Therefore, the time interval of the fast-time sampling points can be designed based on the azimuth interval, and the echo value corresponding to each fast-time sampling point is obtained to simulate the sinusoidal echo. Each frame of echo data contains several values ​​of fast-time sampling points. By mapping each frame of echo in a Block of the GPU, each sampling point data of the frame echo is mapped in each Thread in the Block, and phase compensation is implemented in parallel. The implementation method of phase compensation is as follows: after the platform moves in azimuth for a slow time interval, the displacement of each antenna array element must satisfy:

[0020]

[0021] Where d is the spacing between antenna elements, V is the speed of the moving platform, T is the pulse transmission period, n is the number of pulse transmission periods, and N is + represents any positive integer. After range pulse compression, the echo signals received by the m-1th and mth array elements are

[0022]

[0023]

[0024] Among them, s rc is the signal echo, the distance between the antenna element and the target is R, c is the speed of light, and the cutoff frequency is f c Since the distance between the target and the radar is much larger than the antenna array element spacing, assuming that the target only moves along the distance direction at a speed v, after a time interval η,

[0025] R1(t+η)≈R1(t)+v r η

[0026] R m (t+η)≈R m-1 (t)+v r η

[0027] Simplified

[0028]

[0029] It can be seen that the pulse compression signal received by the m-1th array element at time t and the pulse compression signal received by the mth array element at time t+η differ by a phase factor The purpose of image registration is to align echo signals obtained at different times in the same coordinate system. In the antenna phase center offset algorithm, image registration only needs to compensate the imaging result in azimuth.

[0030] Furthermore, in step 6, the specific implementation method of the cancellation CUDA operator is: after completing the alignment of the two adjacent channels, immediately apply for the corresponding GPU resources for the cancellation operation between the adjacent channel data; the cancellation method is: each Thread in the GPU is used to align the same distance sampling points of the echo data in the same azimuth direction in the two adjacent channels, and perform cancellation calculation to complete synchronous cancellation; the stationary clutter is effectively suppressed by pulse cancellation, and finally the moving target echo is obtained.

[0031] Beneficial effects:

[0032] 1. Taking advantage of the independence of channel data, a data stream is bound to the echo data of each channel. Each data stream is responsible for the data transmission and calculation processing of the bound channel, realizing asynchronous transmission and parallel calculation between different channels, greatly reducing the time consumption of clutter suppression.

[0033] 2. Taking advantage of the independence of pulse data within the channel, the fast Fourier transform batch processing technology is used to realize multi-pulse parallel compression operation, further reducing the time consumption of the algorithm.

[0034] 3. The present invention realizes real-time clutter suppression processing of echo data by asynchronously and parallelly processing the registration cancellation of data between adjacent channels, effectively improves the processing efficiency of the antenna phase center offset clutter suppression algorithm, and has a certain robustness for different radar echo data. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0036] Figure 2 Schematic diagram of data flow mapping channel;

[0037] Figure 3 It is a schematic diagram of asynchronous parallel pulse compression between channels;

[0038] Figure 4 This is a schematic diagram of parallel registration within a channel;

[0039] Figure 5 It is the GPU thread structure;

[0040] Figure 6 This is the schematic diagram of antenna phase center offset;

[0041] Figure 7 Schematic diagram of parallel cancellation between channels;

[0042] Figure 8 Time trend chart for different antenna phase center offset clutter suppression algorithm processing solutions. DETAILED DESCRIPTION

[0043] The drawings are for illustrative purposes only and should not be construed as limiting the present patent;

[0044] In order to better illustrate the present embodiment, some parts in the drawings may be omitted, enlarged or reduced, and do not represent actual sizes;

[0045] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0046] The detection area is tested using a 2-transmit 4-receive radar antenna array. The relevant parameters of the antenna radar are as follows: the radar system is a linear frequency modulated continuous wave; the starting frequency is 2.7 GHz, denoted as f0; the signal bandwidth is 500 MHz, denoted as B; the modulation rate is 1.136×10 12 ; The pulse repetition frequency is 1923, recorded as PRF; The sampling rate is 10MHz, recorded as fs. The GPU device uses a graphics card model NVIDIA RTX 4070, which has 5888CUDA cores and 12G video memory. A GPU-based antenna phase center offset multi-stream asynchronous parallel clutter suppression method is used for experimental verification. The specific implementation steps are shown in the flowchart. Figure 1 As shown, the following steps are included:

[0047] Step 1. Create a number of Stream data streams on the GPU device that is the same as the number of channels:

[0048] Stream data flow is a feature of GPU devices. Data flow is a computing technology. When the pre-operand of a node in the data flow is ready, the device can start the node, and the execution result of the node will be directly transmitted to its child nodes. At the same time, a GPU device can have multiple Stream data flows at the same time. The operation execution mode under the same data flow is serial execution, while different Stream data flows can execute computing instructions independently and in parallel. The number of channels refers to the use of antenna arrays to detect moving targets behind walls. The antenna array has both transmitting antennas and receiving antennas. The waveform transmission and reception channel between a transmitting antenna and a receiving antenna is called a channel. The number of channels is: the number of receiving antennas * the number of transmitting antennas. Combining the above characteristics, asynchronous parallel computing between different channel data can be achieved. Define the number of channels as n, and record n ​​as 2, 3, ..., N; the way to create a Stream data flow on a GPU device is to use CUDA to call the cudaStreamCreate (cudaStream_t*pStream) function to create n Stream data flows. Among them, the data type of the Stream data flow is cudaStream_t, and the pStream pointer is used to index the data flow group with a length of n.

[0049] Using a 2-transmit 4-receive radar antenna array, define the number of channels n as 8 and create 8 data streams: cudaStream_tpStream[8]. Use pStream[i] to index the i-th data stream, i = 0, 1, ..., 7. These 8 data streams are independent of each other and do not affect each other.

[0050] Step 2: Map the data stream to the channel echo data, that is, each Stream data stream is bound to one of the channels and is responsible for processing the echo number of the channel:

[0051] Channel echo data refers to the signal data reflected by the target and received by the receiving antenna in the channel formed by each group of transmitting and receiving antennas. The echo data in each channel are independent of each other and do not affect each other. Combining the above characteristics with the asynchronous parallel processing characteristics of the Stream data flow in the GPU, each Stream data flow is bound to the data processing process of one of the channels to achieve independent asynchronous parallel processing of data between channels. That is, the i-th data stream pStream[i] is responsible for the radar echo data of all frames of the i-th channel, i = 0, 1, ..., 7. For example Figure 2 shown.

[0052] Step 3: Apply for GPU memory resources by using the cudaMalloc instruction in the CUDA instruction set architecture to store multi-channel radar echo data, and use the cudaMemcpyAsync instruction to asynchronously transfer data from the CPU device to the GPU device according to different data streams to achieve data transmission;

[0053] To process data on a GPU device, you first need to copy the data from the main memory to the GPU memory. The cudaMalloc function is used to apply for a memory size on the GPU device that can store all radar echo data. The cudaMalloc instruction in the CUDA instruction set architecture is used to apply for GPU memory resources to store multi-channel radar echo data. The data size is: number of channels * distance sampling points * number of radar echo cumulative frames. The cudaMemcpyAsync instruction is used to asynchronously transfer data from the CPU device to the GPU device through the corresponding data stream according to different channels to achieve data transmission. The pStream[i] parameter is passed to the cudaMemcpyAsync function, where i = 0, 1, ..., 7; it is used to transfer radar echo data of different channels to the GPU memory according to the specified data stream.

[0054] Step 4: Perform asynchronous parallel pulse compression on each channel data, such as Figure 3 As shown:

[0055] The specific method of pulse compression is: the complex conjugate of the signal can be calculated according to the parameters of the radar, and then convolved with the echo signal. This process is also called matched filtering. A compressed pulse can be obtained through matched filtering, which can not only improve the resolution, but also help improve the signal-to-noise ratio of the echo signal. By combining the logic of the pulse compression algorithm with the parallel architecture characteristics of the GPU, a CUDA operator is designed to achieve parallel pulse compression.

[0056] Step 4.1. Create an FFT plan using cufftHandle and call the cufftPlanMany and cufftExecZ2Z APIs to implement parallel batch fast Fourier transform of each channel data and transform the signal data from the time domain to the frequency domain:

[0057] cufftHandle, cufftPlanMany and cufftExecZ2Z are the cufft library functions of CUDA. cufftHandle is used to create FFT task plans; cufftPlanMany is used to design task plans. By passing the CUFFT_FORWARD parameter to cufftPlanMany, you can specify the execution of the Fast Fourier Transform task; cufftExecZ2Z is ​​used to execute the task plan.

[0058] Step 4.2: Generate a reference signal based on the radar antenna parameters and convolve it with the echo signal:

[0059] The reference signal is generated from a known form of the transmitted signal. Here a linear frequency modulated pulse is used and the transmitted signal can be expressed as: Where A is the amplitude of the signal, f0 is the initial frequency, Δf is the bandwidth, T is the pulse duration, and t is the time. The reference signal is a linear frequency modulation signal in the same form as the transmitted signal and can be expressed as: Multiply the reference signal with the echo signal to obtain a convolution result and improve the resolution.

[0060] Step 4.3, by using cufftHandle to create an IFFT plan, and by calling the cufftPlanMany and cufftExecZ2Z API interfaces, parallel batch fast inverse Fourier transform of each channel data is implemented to achieve the transformation of signal data from frequency domain to time domain:

[0061] cufftHandle is used to create an IFFT task plan; cufftPlanMany is used to design a task plan. By transmitting the CUFFT_INVERSE parameter to cufftPlanMany, the fast inverse Fourier transform task can be specified; cufftExecZ2Z is ​​used to execute the task plan.

[0062] Step 5: Design the “registration CUDA operator” to implement phase compensation for each frame of echo data in each channel in parallel, that is, the registration between channels, such as Figure 4 As shown:

[0063] The GPU thread structure is as follows Figure 5 As shown. The specific implementation method of the registered CUDA operator is as follows: Each frame of radar echo is a continuous sinusoidal waveform. Due to hardware limitations, it is impossible to record all the data at each moment. Therefore, the time interval of the fast-time sampling points can be designed based on the azimuth interval, and the echo value corresponding to each fast-time sampling point is obtained to simulate the sinusoidal echo. Each frame of echo data contains several values ​​of fast-time sampling points. By mapping each frame of echo to a Block of the GPU, each sampling point data of the frame of echo is mapped to each Thread in the Block, and phase compensation is achieved in parallel. The principle of antenna phase center offset is as follows Figure 6 As shown. The implementation method of phase compensation is: after the platform moves along the azimuth direction for a slow time interval, the displacement of each antenna array element must satisfy: Where d is the spacing between antenna elements, V is the speed of the moving platform, T is the pulse transmission period, n is the number of pulse transmission periods, and N is + Represents any positive integer. After range pulse compression, the echo signals received by the m-1th and mth array elements are: Among them, s rc is the signal echo, the distance between the antenna element and the target is R, c is the speed of light, and the cutoff frequency is f c Since the distance between the target and the radar is much larger than the antenna array element spacing, assuming that the target only moves along the distance direction at a speed v, after a time interval η, R1(t+η)≈R1(t)+v r η; R m (t+η)≈R m-1 (t)+v r η, simplified to

[0064]

[0065] It can be seen that the pulse compression signal received by the m-1th array element at time t and the pulse compression signal received by the mth array element at time t+η differ by a phase factor The purpose of image registration is to align echo signals obtained at different times in the same coordinate system. In the antenna phase center offset algorithm, image registration only needs to compensate the imaging result in azimuth.

[0066] In the GPU, all frame data of each channel are mapped to a Grid. A Grid contains x Blocks, and each Block corresponds to a frame of data. Because the data of all sampling points of each frame of data need to be aligned and calibrated through calculation, each Thread in each Block is used to execute the "alignment CUDA operator" to calculate a certain sampling point of the frame data it is responsible for. All Grids process data of different channels asynchronously and in parallel, and all Blocks and Threads in the same Grid process all sampling points of all frame echo data in the same channel synchronously and in parallel.

[0067] Step 6: Design a “cancellation CUDA operator” to achieve cancellation between adjacent channels and obtain multi-channel data after clutter suppression, such as Figure 7 As shown:

[0068] The specific implementation method of the cancellation CUDA operator is: after the two adjacent channels have completed the "alignment CUDA operator" task, immediately apply for the corresponding GPU resources to perform the cancellation operation between the data of the two adjacent channels. The cancellation method is: each Thread in the GPU is used to align the same distance sampling points of the echo data in the same azimuth direction in the two adjacent channels, and perform cancellation calculations to complete synchronous cancellation. That is, apply for a Grid grid of the same size as the Grid applied for in the "alignment CUDA operator", and the size of each Block in the Grid is also the same, that is, each Block contains the same number of Threads as the Block applied for in the "alignment CUDA operator". Each Thread is responsible for canceling the data with the same relative position in the two adjacent channels. The stationary clutter is effectively suppressed by pulse cancellation, and the moving target echo is finally obtained.

[0069] The following is an experimental verification.

[0070] Under the conditions of 8 channels and 6361 range sampling numbers, the GPU-based antenna phase center offset clutter suppression method is used. The average time consumption in processing radar echo data with different azimuth sampling numbers is counted under two parallel processing schemes: single-stream parallel processing (Scheme 1) and multi-stream asynchronous parallel processing (Scheme 2). The statistical results are shown in Table 1. It can be seen that when the number of sampling points in the range and azimuth directions is the same, the processing efficiency of the GPU-based antenna phase center offset multi-stream asynchronous parallel clutter suppression method is higher than that of the single-stream parallel technology.

[0071] Table 1

[0072]

[0073] The time consumption trend of the GPU-based antenna phase center offset clutter suppression algorithm single-stream parallel processing technology and multi-stream asynchronous parallel processing technology when processing radar echo data with different azimuth sampling numbers is shown in the figure below. Figure 8 As shown, it can be seen that with the increase in the number of azimuth sampling points, the GPU-based antenna phase center offset multi-stream asynchronous parallel clutter suppression method proposed in this paper has more and more obvious advantages in processing efficiency.

[0074] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A GPU-based antenna phase center offset multi-stream asynchronous parallel clutter suppression method, characterized in that: The following steps are involved: Step 1. Create a number of Stream data streams on the GPU device that is the same as the number of channels; Step 2: Map the data stream to the channel echo data, that is, each Stream data stream is bound to one of the channels and is responsible for processing the echo number of the channel; Step 3: Apply for GPU memory resources by using the cudaMal loc instruction in the CUDA instruction set architecture to store multi-channel radar echo data, and use the cudaMemcpyAsync instruction to asynchronously transfer data from the CPU device to the GPU device according to different data streams to achieve data transmission; Step 4, performing asynchronous parallel pulse compression on each channel data; Step 5: Design a CUDA operator for registration to implement phase compensation for each frame of echo data in each channel in parallel, i.e., registration between channels; Step 6: Design a cancellation CUDA operator to achieve cancellation between adjacent channels and obtain multi-channel data after clutter suppression.

2. The method according to claim 1, characterized in that: In step 1, the number of channels specifically refers to: when using an antenna array to detect a moving target behind a wall, the antenna array has both a transmitting antenna and a receiving antenna, and a waveform transceiver channel between one transmitting antenna and one receiving antenna is called one channel; The number of channels is: number of receiving antennas * number of transmitting antennas.

3. The method according to claim 1, characterized in that: In step 3, the multi-channel radar echo data is stored, and the data size is: number of channels * distance sampling points * number of radar echo accumulated frames.

4. The method according to claim 1, characterized in that: In step 4, the pulse compression method is: the complex conjugate of the signal can be calculated according to the parameters of the radar, and then convolved with the echo signal, which is also called matched filtering; A compressed pulse can be obtained through matched filtering.

5. The method according to claim 1, characterized in that: The step 4 includes the following sub-steps: Step 4.1, by using cufftHandle to create an FFT plan, and by calling the cufftPlanMany and cufftExecZ2Z API interfaces, parallel batch fast Fourier transform of each channel data is implemented to achieve the transformation of signal data from time domain to frequency domain; Step 4.2: Generate a reference signal according to the radar antenna parameters and convolve it with the echo signal; Step 4.

3. Create an IFFT plan using cufftHandle and call the cufftPlanMany and cufftExecZ2Z API interfaces to implement parallel batch fast inverse Fourier transform of each channel data and transform the signal data from the frequency domain to the time domain.

6. The method according to claim 1, characterized in that: In step 5, the specific implementation method of the alignment CUDA operator is as follows: each frame of radar echo is a continuous sinusoidal waveform. Due to the limitation of hardware conditions, it is impossible to record all the data at each moment. Therefore, the time interval of the fast time sampling point can be designed based on the azimuth interval, and the echo value corresponding to each fast time sampling point is obtained to simulate the sinusoidal echo; each frame of echo data contains several values ​​of the fast time sampling points. By mapping each frame of echo in a Block of the GPU, each sampling point data of the frame echo is mapped in each Thread in the Block, and phase compensation is realized in parallel.

7. The method according to claim 6, characterized in that: In step 5, the phase compensation is implemented as follows: after the platform moves in azimuth for a slow time interval, the displacement of each antenna array element must satisfy: Where d is the spacing between antenna elements, V is the speed of the moving platform, T is the pulse transmission period, n is the number of pulse transmission periods, and N is + represents any positive integer; after range pulse compression, the echo signals received by the m-1th and mth array elements are Among them, s rc is the signal echo, the distance between the antenna array element and the target is R, c is the speed of light, and the cutoff frequency is f c Since the distance between the target and the radar is much larger than the antenna array element spacing, assuming that the target only moves along the distance direction at a speed v, after a time interval η, R1(t+η)≈R1(t)+v r η R m (t+η)≈R m-1 (t)+v r or Simplified It can be seen that the pulse compression signal received by the m-1th array element at time t and the pulse compression signal received by the mth array element at time t+η differ by a phase factor 8. The method according to claim 1, characterized in that: In step 6, the specific implementation method of the cancellation CUDA operator is: after completing the alignment of the two adjacent channels, immediately apply for the corresponding GPU resources for the cancellation operation between the adjacent channel data; the cancellation method is: each Thread in the GPU is used to align the same distance sampling points of the echo data in the same azimuth direction in the two adjacent channels, and perform cancellation calculation to complete synchronous cancellation; the stationary clutter is effectively suppressed by pulse cancellation, and finally the moving target echo is obtained.

Images