Millimeter-wave radar RMA imaging acceleration method based on FPGA
Through the decomposition and reconstruction of the RMA algorithm and the implementation of FPGA, the problem of low efficiency of the RMA imaging method on the CPU is solved, efficient and low-cost imaging acceleration is achieved, and real-time and imaging resolution are improved.
Patent Information
- Application Number
- CN202211559250.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-12-06
- Publication Date
- 2025-08-19
- Estimated Expiration
- 2042-12-06
AI Technical Summary
The existing millimeter-wave radar RMA imaging methods have low execution efficiency on the CPU and slow computing speed, making it difficult to meet real-time requirements. Especially in three-dimensional imaging scenarios, there is no disclosed technology for FPGA acceleration processing.
The original RMA algorithm is reconstructed, and the FPGA-based method is used to decompose it into horizontal and vertical one-dimensional fast Fourier transforms, combining preprocessing calibration and matching filtering to simplify the calculation steps and reduce hardware resource usage.
Improve RMA imaging efficiency, reduce imaging time, enhance real-time, reduce cost and power consumption, and reduce hardware resource utilization.
Smart Images
Figure CN116224264B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of millimeter wave radar imaging, and in particular to a millimeter wave radar RMA imaging acceleration method based on FPGA. Background Art
[0002] The millimeter-wave radar RMA (range migration algorithm, abbreviated: RMA) imaging method is used to form images from synthetic aperture radar (SAR) data. SAR imaging algorithm research is already at a relatively mature stage. For example, RDA (range Doppler algorithm), RMA, and BPA (backprojection algorithm) have been widely used. However, in three-dimensional imaging scenarios, SAR systems are time-consuming and inefficient. Therefore, it is necessary to restructure the basic algorithm based on the characteristics of the hardware platform to achieve accelerated results. The following introduces the existing technologies for the three basic millimeter-wave imaging algorithms mentioned above. Their common disadvantage is low CPU execution efficiency and slow imaging speed:
[0003] 1. BPA is based on the concept of time-delayed addition, restoring imaging target information through the coherent superposition of echo data. Since BPA has no specific constraints on the array configuration or scanning method, it is applicable to any array type. MIMO arrays, on the other hand, have diverse array configurations. The echo data of MIMO radars is directly related to the positions of the transmitting and receiving antenna elements. However, other imaging algorithms, such as range Doppler algorithms, require certain constraints on the array's geometric model and scanning model. Therefore, BPA, which is not restricted by array form, has become the preferred MIMO imaging algorithm. However, for MIMO radar's three-dimensional spatial imaging, BPA is computationally intensive, making it difficult to ensure real-time performance using FPGAs.
[0004] 2. RDA achieves efficient modular processing through frequency-domain operations on range and azimuth, while retaining the simplicity of one-dimensional operations. Based on the large-scale temporal differences in range and azimuth, the algorithm employs range cell migration correction (RCMC) between the two one-dimensional operations, achieving an approximate separation of range and azimuth. However, RDA has inherent drawbacks: range cell migration correction requires interpolation, which increases the algorithm's computational burden and reduces imaging resolution. This makes it particularly incapable of achieving high-precision imaging in the squint SAR mode.
[0005] 3. RMA, also known as the wavenumber domain algorithm, originated in the field of seismic signal processing. After development, the range migration algorithm was applied to the field of SAR millimeter wave imaging. Because it uses fewer approximations and has better imaging results, it is considered to be the optimal algorithm. The advantage of RMA over other radar imaging algorithms is that it can not only fully compensate for wavefront curvature, but also achieve extremely high computational efficiency through fast Fourier transform. Compared with BPA, it has less computational complexity and better imaging results. Most of the processing in RMA is performed in the frequency domain, which is more suitable for computers. Its image inversion method is distortion-free, and its imaging effect is theoretically better than other algorithms. Although it is better than other imaging algorithms, the RMA imaging method requires a large number of fast Fourier transforms and matrix multiplication operations during the calculation process, resulting in a very slow calculation speed in the traditional CPU architecture hardware platform and cannot meet real-time requirements.
[0006] There is currently no relatively public technology for FPGA acceleration processing of RMA. We can fully utilize the algorithm architecture characteristics of RMA, optimize the data processing link of the traditional RMA algorithm, form an RMA architecture suitable for parallel computing, and realize algorithm acceleration design on the FPGA platform.
[0007] Field Programmable Gate Array (FPGA) devices are widely used in algorithm hardware acceleration due to their parallel processing capabilities, reconfigurability, scalability and low cost. Summary of the Invention
[0008] Combining the characteristics of FPGA, the present invention proposes an FPGA-based millimeter-wave radar RMA imaging acceleration method, which reconstructs the original algorithm, simplifies its time complexity while facilitating hardware implementation, and achieves an acceleration effect.
[0009] In order to solve the problem that the RMA imaging method runs slowly on the CPU, the present invention reconstructs and optimizes the original RMA algorithm, and the technical solution adopted is:
[0010] The FPGA-based millimeter wave radar RMA imaging acceleration method includes the following steps:
[0011] Step 1: Preprocessing calibration: multiply the original radar signal by the preprocessing calibration factor;
[0012] Step 2: Horizontal one-dimensional fast Fourier transform;
[0013] Step 3: vertical one-dimensional fast Fourier transform;
[0014] Step 4: Matched filtering: perform point multiplication on the input frequency domain signal and the matched filter factor to achieve frequency domain filtering;
[0015] Step 5: Horizontal one-dimensional inverse fast Fourier transform;
[0016] Step 6: Perform one-dimensional inverse fast Fourier transform vertically.
[0017] The present invention has the following beneficial technical effects:
[0018] 1. The FPGA-based millimeter-wave radar RMA imaging acceleration method improves the efficiency of the RMA imaging method, reduces the time consumed by imaging, and enhances real-time performance.
[0019] 2. The present invention uses FPGA for algorithm acceleration, which is low in cost and power consumption compared to CPU / GPU.
[0020] 3. The present invention decomposes and reconstructs the original RMA imaging scheme, reducing the resource usage of the algorithm in FPGA hardware implementation. BRIEF DESCRIPTION OF THE DRAWINGS
[0021] Figure 1 It is a schematic diagram of the RMA imaging method in the prior art;
[0022] Figure 2 is a schematic diagram of the imaging method of the present invention;
[0023] Figure 3 FPGA hardware implementation flowchart framework for the present invention;
[0024] Figure 4 It is a detection effect diagram of the present invention. DETAILED DESCRIPTION
[0025] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings of the present invention. In addition, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this research direction without creative work are within the scope of protection of the present invention.
[0026] like Figure 1 As shown, the existing RMA imaging method includes the following steps:
[0027] The first step is time domain distance calibration, which involves multiplying the original radar data by a pre-known distance calibration factor to perform calibration in the distance direction.
[0028] The second step is time domain phase calibration, which involves multiplying the raw radar data by a pre-known phase calibration factor to calibrate the data phase.
[0029] The third step is a dimensionality reduction operation based on multiple one-dimensional fast Fourier transforms, which compresses the original radar data in the depth direction to achieve the purpose of dimensionality reduction, and ultimately compresses the data depth from 64 to 1.
[0030] The fourth step is two-dimensional fast Fourier transform, which converts the time domain signal into a frequency domain signal.
[0031] The fifth step is matched filtering, which multiplies the input frequency domain signal by the matched filter factor to achieve the filtering effect.
[0032] The sixth step is the two-dimensional inverse fast Fourier transform, which converts the filtered frequency domain signal into a time domain signal.
[0033] The above RMA imaging method has some problems in FPGA hardware acceleration:
[0034] 1) There are redundant operations in the calculation between the first and third steps.
[0035] 2) The 2D Fast Fourier Transform in the fourth and sixth steps cannot be directly implemented in an FPGA with low hardware resources.
[0036] To solve this problem, the present invention decomposes and reconstructs the algorithm for optimization. The optimized algorithm principle diagram is shown in the following figure: Figure 2 As shown, the specific steps include:
[0037] The first step is preprocessing calibration. This step combines the time-domain range calibration, time-domain phase calibration, and dimensionality reduction operations of multiple one-dimensional fast Fourier transforms in the original algorithm, performing a dot product between the original radar signal and the preprocessing calibration factors. The time complexity of this step is reduced from 2xDxNxM+2xDxlog(D)xNxM in the original algorithm to DxNxM, where D is the radar data depth, N is the radar data length, and M is the radar data width.
[0038] The second step is a horizontal one-dimensional fast Fourier transform, and the third step is a vertical one-dimensional fast Fourier transform. These two steps decompose and reconstruct the original two-dimensional fast Fourier transform, using several one-dimensional fast Fourier transforms to achieve the equivalent. This equivalence significantly reduces the hardware resources required to implement this imaging method in an FPGA.
[0039] The fourth step is matched filtering, which performs point multiplication on the input frequency domain signal and the matched filter factor to achieve the effect of frequency domain filtering.
[0040] The fifth step is the horizontal one-dimensional inverse fast Fourier transform, and the sixth step is the vertical one-dimensional inverse fast Fourier transform. These two steps decompose and reconstruct the original two-dimensional inverse fast Fourier transform and use several one-dimensional inverse fast Fourier transforms to achieve the equivalent.
[0041] like Figure 3 As shown in the FPGA hardware implementation flow chart of the present invention, a millimeter wave radar RMA imaging acceleration method based on FPGA is proposed based on this, and its specific operation is described as follows:
[0042] 1. Collect radar data: Collect three-dimensional radar data.
[0043] 2. Read radar data:
[0044] 1) Save a set of 3D radar data to the SD card.
[0045] 2) When the FPGA is powered on, the radar data is read from the SD card and sent to the pre-processing calibration for calculation.
[0046] 3. Preprocessing calibration:
[0047] 1) Calculate the time domain calibration matrix and the time domain distance offset calibration matrix.
[0048] Time domain calibration matrix: First calculate the distance difference between the transmitting and receiving antennas to the target center and the equivalent phase center to the target center
[0049]
[0050] Since the MIMO array is arranged in the y direction, represents the distance difference between the transmitting and receiving antennas corresponding to the equivalent phase center of the first group. Then, the time domain calibration matrix can be calculated, and z0 represents the vertical distance to the target.
[0051] Range offset correction matrix: The range offset correction matrix can be obtained by performing range FFT on the uncorrected echo signal and obtaining the phase value at the range peak index.
[0052] 2) Calculate the FFT frequency domain matrix corresponding to the fast Fourier transform with a step size of 64 using the single-point discrete Fourier transform calculation formula.
[0053] 3) Multiply the time domain calibration matrix, the time domain distance offset calibration matrix and the FFT frequency domain matrix to obtain the preprocessing calibration matrix.
[0054] 4) Perform dot multiplication of the original 3D radar data with the preprocessed calibration matrix.
[0055] 5) Write the calculated results into RAM / DDR
[0056] 6) The specific calculation formula is as follows:
[0057] O (i,j,k) =X (i,j,k) P (i,j,k)
[0058] where X(i,j,k) Represents the input 3D radar data, P (i,j,k) represents the preprocessing calibration matrix, O (i,j,k) Represents the output result data, with a size of 256x256x64. i represents the horizontal coordinate, j represents the vertical coordinate, and k represents the depth coordinate.
[0059] 4. Horizontal one-dimensional fast Fourier transform (FFT):
[0060] 1) Use the axi_full bus protocol to read the calculation results of the previous step from DDR.
[0061] 2) Convert data from memory mapped transfer to streaming data transfer.
[0062] 3) Feed the data stream into the one-dimensional fast Fourier transform calculation kernel.
[0063] 4) The one-dimensional fast Fourier transform calculation kernel is reused based on each row of data.
[0064] 5) Transpose the calculated results and write them into RAM / DDR.
[0065] 6) The specific calculation formula is as follows:
[0066]
[0067] in x(n) represents the input row data, N represents the number of rows of data, which is 64, W represents the frequency components, and X(k) represents the FFT output. n represents the sequence number of the time-domain discrete sequence, and k represents the sequence number of the frequency-domain discrete sequence.
[0068] 5. Vertical one-dimensional fast Fourier transform (FFT):
[0069] 1) Use the axi_full bus protocol to read the calculation results of the previous step from DDR.
[0070] 2) Convert data from memory mapped transfer to streaming data transfer.
[0071] 3) Feed the data stream into the one-dimensional fast Fourier transform calculation kernel.
[0072] 4) Taking each row of data as a unit (the transposition operation has been performed in step 4), the one-dimensional fast Fourier transform calculation kernel is reused.
[0073] 5) Write the calculated results into RAM / DDR.
[0074] 6) The specific calculation formula is the same as step 4.
[0075] 6. Matched filtering:
[0076] 1) Calculate the matched filter matrix and write it into RAM / DDR. Matched filter matrix: in k x ,k y They represent the Fourier space angular frequencies corresponding to the spatial sampling positions. k represents the wave number.
[0077] 2) Use the axi_full bus protocol to read the calculation results and matched filter matrix of the previous step from DDR.
[0078] 3) Convert data from memory mapped transfer to streaming data transfer.
[0079] 4) Perform dot multiplication of the data and the matched filter matrix in a streaming data manner.
[0080] 5) Write the calculated results into RAM / DDR.
[0081] 6) The specific calculation formula is as follows:
[0082] O (i,j,k) =X fft(i,j,k) F (i,j,k)
[0083] where X fft The output after two FFTs, F represents the matched filter matrix, O represents the output data, and the size is 256x256x64. i represents the horizontal coordinate, j represents the vertical coordinate, and k represents the depth coordinate.
[0084] 7. Horizontal one-dimensional inverse fast Fourier transform (inverse FFT):
[0085] 1) Use the axi_full bus protocol to read the calculation results of the previous step from DDR.
[0086] 2) Convert data from memory mapped transfer to streaming data transfer.
[0087] 3) Feed the data stream into the one-dimensional inverse fast Fourier transform calculation kernel.
[0088] 4) The one-dimensional inverse fast Fourier transform calculation kernel is reused based on each row of data.
[0089] 5) Transpose the calculated results and write them into RAM / DDR.
[0090] 6) The specific calculation formula is as follows:
[0091]
[0092] in X filter (k) represents the input row data after matched filtering, N represents the number of rows of data, which is 64, W represents the frequency component, x filter (n) represents the inverse FFT output result.
[0093] 8. Vertical one-dimensional inverse fast Fourier transform (inverse FFT):
[0094] 1) Use the axi_full bus protocol to read the calculation results of the previous step from DDR.
[0095] 2) Convert data from memory mapped transfer to streaming data transfer.
[0096] 3) Feed the data stream into the one-dimensional fast inverse Fourier transform calculation kernel.
[0097] 4) Taking each row of data as a unit (the transposition operation has been performed in step 7), reuse the one-dimensional inverse fast Fourier transform calculation kernel.
[0098] 5) Write the calculated results into RAM / DDR.
[0099] like Figure 4 As shown, the results of the detection using the method of the present invention are as follows Figure 4 As shown in the middle right picture.
[0100] After FPGA accelerated processing, the imaging resolution can reach a theoretical 3mm, and the imaging speed is only 20ms.
Claims
1. The FPGA-based millimeter wave radar RMA imaging acceleration method is characterized by: The following steps are involved: Step 1: Preprocessing calibration: perform dot multiplication of the original radar signal and the preprocessing calibration matrix; 1) Calculate the time domain calibration matrix and the time domain distance offset calibration matrix; Time domain calibration matrix: First calculate the distance difference between the transmitting and receiving antennas to the target center and the equivalent phase center to the target center Since the MIMO array is arranged in the y direction, represents the distance difference between the transmitting and receiving antennas corresponding to the equivalent phase center of the lth group, and then the time domain calibration matrix can be calculated. z0 represents the vertical distance to the target; Range offset calibration matrix: The range offset calibration matrix is obtained by performing range FFT on the uncorrected echo signal and obtaining the phase value at the range peak index; 2) Calculate the FFT frequency domain matrix corresponding to the fast Fourier transform with a step size of 64 using the single-point discrete Fourier transform calculation formula; 3) Multiplying the time domain calibration matrix, the time domain distance offset calibration matrix and the FFT frequency domain matrix to obtain a preprocessing calibration matrix; 4) Perform dot multiplication of the original 3D radar data and the preprocessed calibration matrix. The specific calculation formula is as follows: O(i,j,k)=X (i,j,k) P (i,j,k) where X (i,j,k) Represents the input 3D radar data, P (i,j,k) represents the preprocessing calibration matrix, O (i,j,k) Represents the output result data, i represents the horizontal coordinate, j represents the vertical coordinate, and k represents the depth coordinate; 5) Write the calculated results into RAM / DDR; Step 2: Horizontal one-dimensional fast Fourier transform; Step 3: vertical one-dimensional fast Fourier transform; Step 4: Matched filtering: perform point multiplication on the input frequency domain signal and the matched filter factor to achieve frequency domain filtering; Step 5: Horizontal one-dimensional inverse fast Fourier transform; Step 6: Perform one-dimensional inverse fast Fourier transform vertically.
2. The FPGA-based millimeter wave radar RMA imaging acceleration method according to claim 1, characterized in that: The second step specifically includes: 1) Read the calculation result of the previous step from DDR; 2) Convert data from memory-mapped transfer to streaming data transfer; 3) Feed the data stream into the one-dimensional fast Fourier transform calculation kernel; 4) Reuse the one-dimensional fast Fourier transform calculation kernel based on each row of data; The specific calculation formula is as follows: in x(n) represents the input row data, N represents the number of data in each row, W represents the frequency component, X(k) represents the FFT output result, n represents the sequence number of the time domain discrete sequence, and k represents the sequence number of the frequency domain discrete sequence.
3. The FPGA-based millimeter wave radar RMA imaging acceleration method according to claim 1, characterized in that: The step three specifically includes: 1) Read the calculation result of the previous step from DDR; 2) Convert data from memory-mapped transfer to streaming data transfer; 3) Feed the data stream into the one-dimensional fast Fourier transform calculation kernel; 4) The one-dimensional fast Fourier transform calculation kernel is reused based on each row of data.
4. The FPGA-based millimeter wave radar RMA imaging acceleration method according to claim 1, characterized in that: The step 4 specifically includes: 1) Calculate the matched filter matrix and write it into RAM / DDR; 2) Read the calculation results and matched filter matrix of the previous step from DDR; 3) Convert data from memory-mapped transfer to streaming data transfer; 4) Perform dot multiplication of the data and the matched filter matrix in a streaming manner. The specific calculation formula is as follows: O (i,j,k) =X (i,j,k) F (i,j,k) Among them F (i,j,k) represents the matched filter matrix, i represents the horizontal coordinate, j represents the vertical coordinate, and k represents the depth coordinate.
5. The FPGA-based millimeter wave radar RMA imaging acceleration method according to claim 1, characterized in that: The step five specifically includes: 1) Read the calculation result of the previous step from DDR; 2) Convert data from memory-mapped transfer to streaming data transfer; 3) Feed the data stream into the one-dimensional inverse fast Fourier transform calculation kernel; 4) Taking each row of data as a unit, the one-dimensional inverse fast Fourier transform calculation kernel is reused. The specific calculation formula is as follows: in X(k) represents the inverse fast Fourier transform input row data, N represents the number of data in each row, W represents the frequency component, x(n) represents the inverse fast Fourier transform output result, n represents the sequence number of the time domain discrete sequence, and k represents the sequence number of the frequency domain discrete sequence.
6. The FPGA-based millimeter wave radar RMA imaging acceleration method according to claim 1, characterized in that: The step six specifically includes: 1) Read the calculation result of the previous step from DDR; 2) Convert data from memory-mapped transfer to streaming data transfer; 3) Feed the data stream into the one-dimensional fast inverse Fourier transform calculation kernel; 4) The one-dimensional inverse fast Fourier transform calculation kernel is reused based on each row of data.