Radar real-time target detection method based on heterogeneous parallel computing

Abstract

The invention belongs to the technical field of signal processing, and discloses a radar real-time target detection method based on heterogeneous parallel computing. According to the method, through the steps of channel correction, pulse compression, moving target display, digital beam forming, coherent accumulation or non-coherent accumulation on multi-pulse signals including sum beams and difference beams, and constant false alarm processing, the problems of an existing radar real-time signal processing technology in the aspects of processing speed, accuracy, efficiency, flexibility and the like are solved. By adopting a domestic heterogeneous computing technology and applying a universal software technology, efficient and real-time processing of radar signals is realized, and the overall performance of a radar system is improved.

Description

Technical Field

[0001] The present invention mainly relates to the technical field of signal processing, and in particular to a radar real-time target detection method based on heterogeneous parallel computing. Background Art

[0002] Radar target detection technology mainly includes algorithms such as moving target indication (MTI), pulse compression, digital beamforming (DBF) and constant false alarm rate (CFAR). Due to the real-time requirements and usage requirements, when implementing products, most of them are implemented through FPGA, DSP or multiple CPUs, and accelerated computing is applied to achieve the real-time requirements of these algorithms.

[0003] The advantage of the real-time signal processing method for external radiation source radar based on multi-CPU processing is that it uses multiple CPUs to realize real-time signal processing of external radiation source radar through a general programming language, which increases the flexibility and real-time performance of signal processing; it effectively avoids signal loss and improves processing efficiency by splitting the signal processing process into multiple steps and running them on different CPU resources. The disadvantage is that it may require higher hardware configuration and increase costs; at the same time, the coordination and synchronization between multiple CPUs may be more complicated.

[0004] The advantage of the real-time processing system of intermediate frequency LFM-PD radar signals based on FPGA and DSP is that it combines FPGA and DSP and realizes real-time processing of radar signals through multiple processing modules (such as intermediate frequency sampling, digital down-conversion, pulse compression, etc.); this method has the characteristics of short development cycle and strong flexibility, and is suitable for large-scale radar detection systems. The disadvantages are that the system's specialization and complexity may increase the difficulty of maintenance and upgrading; at the same time, the performance of FPGA and DSP may become the bottleneck of the entire system; and the software of FPGA is relatively specialized and is not implemented with general software.

[0005] The advantage of the GPU-based time-division MIMO radar signal processing method is that it applies a general software method, realizes MTI, beamforming, and CFAR algorithm acceleration on the GPU, and realizes real-time signal processing. The disadvantage is that it uses both the CUDA core and the Tensor core on the GPU, and cannot be used on devices that only have Tensor cores.

[0006] As the demand for radar software and platform localization becomes increasingly strong, the present invention proposes a radar real-time target detection method based on heterogeneous parallel computing. The method applies a general software method to accelerate the radar target detection algorithm on the domestic heterogeneous computing acceleration platform Huawei NPU, Ascend 310p, which only has Tensor cores, and realizes real-time radar signal processing. Summary of the invention

[0007] The purpose of the present invention is to provide a radar real-time target detection method based on heterogeneous parallel computing. Aiming at the problems existing in the existing radar real-time signal processing technology in terms of processing speed, accuracy, efficiency and flexibility, the present invention adopts domestically produced heterogeneous computing technology and applies general software technology to realize efficient and real-time processing of radar signals and improve the overall performance of the radar system.

[0008] To achieve the above object, the present invention provides a radar real-time target detection method based on heterogeneous parallel computing, comprising the following steps: Step 1: Channel correction: Each frame of radar echo is converted into a digital signal through AD, and the radar echo amplitude and phase characteristics of each radar channel are made consistent after channel correction; Step 2: Pulse compression: restore the radar echo narrow pulse signal through pulse compression; Step 3: Moving target display: Calculate the moving target information through the moving target display; Step 4: Digital beamforming: Process the radar echoes of each radar channel to form sum beam and difference beam; Step 5: performing coherent accumulation or non-coherent accumulation on the multi-pulse signal including the sum beam and the difference beam; Step 6: CFAR processing: The target position is calculated by the CFAR algorithm.

[0009] Further, the pulse compression in step 2 includes the following steps: S1: Perform DFT processing on radar echo data with consistent amplitude and phase characteristics after channel correction; S2: Multiply the radar echo data processed in step S1 and the reference signal matrix; S3: Perform iDFT processing again.

[0010] Further, in step S1, DFT processing is performed on the radar echo data with consistent amplitude and phase characteristics after channel correction, including the following steps: S4: For a finite length sequence x[n] of N points, its DFT is mathematically expressed as a multiplication and addition operation, and the formula is ; S5: The process of step S4 is represented by a matrix, and X represents the N-point DFT frequency domain sequence vector, and x represents the time domain sequence vector. Then , the DFT matrix of N points is defined as ; S6: The N-point DFT of the sequence x[n] can be expressed as ; S7: DFT processing is completed through matrix calculation in steps S4-S6. The matrices and vectors are all complex numbers. When implemented on the NPU, the complex matrix needs to be transformed into a real matrix calculation to obtain the final result of DFT processing of the radar echo data with consistent amplitude and phase characteristics after channel correction. The method is: Assume Y=A*B, A, B, Y Are all complex matrices, first complex matrices A and complex matrices B Decompose into real and imaginary parts respectively, and calculate the matrix A and matrix B The product of Y， The process is , where real(Y) means taking the real part of Y, and imag(Y) means taking the imaginary part of Y, then Y can be expressed as .

[0011] Furthermore, in step S2, the radar echo data processed in step S1 is multiplied by the reference signal matrix, which includes the following steps: S8: O=P × Q ,in P is the signal after DFT, Q is the reference signal, O is the result of multiplying the reference signal; because P, Q, O Are all complex matrices, first complex matrices P and complex matrices Q Decompose into real and imaginary parts respectively, and calculate the matrix P and matrix Q The product of O， The process is , where real( O ) means to take O The real part, imag( O ) means to take O The imaginary part of O It can be expressed as O = real(O) + imag(O)j .

[0012] Furthermore, iDFT processing is performed in step S3, including the following steps: S9: Repeat steps S4-S7.

[0013] Furthermore, in step 3, the moving target information is calculated by displaying the moving target, including the following steps: S10: Assume that the order of the moving target display filter is n, then the coefficient vector of the moving target display is ,in ; S11: Assume the number of radar pulses is , the number of distance gates is , then the received signal matrix The dimension is ; S12: Construct auxiliary matrix To achieve MTI, The dimension is , ,in Elements in , then the result of MTI can be expressed as ; S13: The dimension is ,in .

[0014] Furthermore, the digital beamforming in step 4 includes the following steps: S14: Assume that the radar has channels, the multi-channel data received by the radar can be expressed as ; S15: Assuming that n beams need to be formed, the digital beamforming matrix is ​​constructed as ,in, Indicates that the beam is pointed to The weight vector of ; S16: Combining step S14 and step S15, the digital beamforming process can be expressed as .

[0015] Further, in step 5, coherent accumulation or non-coherent accumulation is performed on the multi-pulse signal including the sum beam and the difference beam, including the following steps: S17: Non-coherent accumulation is performed in the slow time dimension. Assume that the radar data after pulse compression is ,in Represents the data of each pulse, with dimension , then the non-coherent accumulation process can be expressed as ,in is the result of non-coherent accumulation, with dimension .

[0016] Furthermore, the constant false alarm processing in step 6 includes the following steps: S18: Assume that the radar data to be detected is , the number of reference units of CFAR is , the number of protection units is , construct the constant false alarm processing convolution kernel as ,in It means to construct a row vector with N elements all set to 1. It means constructing a row vector with N elements all being 0; S19: Yes To fill the edge with zero, the expression is , then the CFAR processing result can be expressed as , where conv(x,y) represents the convolution of x and y vectors.

[0017] Beneficial effects: The present invention provides a radar real-time signal processing method based on heterogeneous computing. Compared with the prior art, the advantages are as follows: (1) Improved processing speed: The parallel processing of matrix operations significantly improves the execution speed of key algorithms such as Fourier transform, pulse compression, MTI, non-coherent accumulation and CFAR, thereby shortening the signal processing cycle and improving the real-time performance of the radar system. (2) Reduced energy consumption: The present invention reduces unnecessary calculation steps and redundant operations, reducing energy consumption in the radar signal processing process; the heterogeneous computing architecture can dynamically allocate computing resources according to processing requirements, avoiding waste of resources and further reducing overall energy consumption. (3) Reduced cost: The present invention uses software optimization to improve processing efficiency, reduce dependence on external hardware devices, and thus reduce the cost of the radar system; by using NPU and Ascend technology, a more cost-effective signal processing solution can be achieved, reducing investment in expensive hardware. (4) Flexibility and scalability: The heterogeneous computing-based approach provides higher flexibility and can adapt to different types of radar systems and diverse application scenarios, whether on land, at sea or in the air, to maintain stable performance; the programming flexibility of the NPU allows the algorithm to be adjusted and optimized according to different mission requirements, enhancing the scalability of the system. (5) Accuracy and reliability: The present invention improves the accuracy and reliability of radar signal processing through precise matrix operations; in complex electromagnetic environments, it can more effectively suppress clutter and improve the accuracy of target detection and tracking. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a radar target detection processing flow chart of a radar real-time signal processing method based on heterogeneous computing involved in an embodiment of the present invention; Figure 2 This is the simulation result of Huawei NPU Ascend 310; Figure 3 This is the simulation result of Nvidia GPU GV100. DETAILED DESCRIPTION

[0020] The preferred mechanism and implementation method of the present invention are further described below in conjunction with the accompanying drawings and specific implementation methods.

[0021] like Figures 1 to 3As shown, an embodiment of the present invention discloses a technical solution for a radar real-time signal processing method based on heterogeneous computing. Figure 1 A radar target detection processing flow chart of a radar real-time signal processing method based on heterogeneous computing according to an embodiment of the present invention; Figure 2 This is the simulation result of Huawei NPU Ascend 310. In the figure, the horizontal axis is the range gate, and the vertical axis is the signal gain after CFAR. It can be seen that there is a signal exceeding 15dB at 2000 range gates, which is the detected target. Figure 3 This is the simulation result of Nvidia GPU GV100. In the figure, the horizontal axis is the range gate and the vertical axis is the signal gain after CFAR. It can be seen that there is a signal exceeding 15dB at 2000 range gates, which is the detected target.

[0022] Due to different architectures, the calculation results of GPU and NPU are slightly different. The calculation results of GPU are consistent with those of CPU. If the calculated value varies greatly, the NPU result is often incorrect. The present invention uses actual results to prove that when calculating the final result of cfar, the results of NPU and GPU are basically consistent. Example

[0023] A radar real-time signal processing method based on heterogeneous computing includes the following steps: Step 1: Channel correction: Each frame of radar echo is converted into a digital signal through AD (analog-to-digital conversion), and the radar echo radiation phase characteristics of each radar channel are made consistent after channel correction; Step 2: Pulse compression: restore the radar echo narrow pulse signal through pulse compression; The pulse compression algorithm consists of three parts. First, DFT is performed on the radar echo data with consistent amplitude and phase characteristics after channel correction, then it is multiplied with the reference signal matrix, and iDFT is performed on the calculated result to restore the original signal.

[0024] (1) Perform DFT processing on the radar echo data with consistent amplitude and phase characteristics after channel correction: The Chinese name of FFT is Fast Fourier Transform, which uses the butterfly algorithm to reduce the computational complexity from O(n2) to O(n*logn). However, the butterfly algorithm uses a loop statement, which is extremely inefficient on the NPU (neural network processor) because the components on the NPU are mainly matrix multipliers. As long as matrix multiplication operations are performed, the highest NPU can reduce the computational complexity from O(n3) to O(1). Therefore, the corresponding functions of FFT must be implemented using matrix methods. Returning to the essence of the problem, FFT and iFFT both perform Fourier transforms, so can Fourier transforms be implemented using matrix methods? This is the most important technology of the first core technology in the present invention. The present invention designs a Fourier transform matrix, and obtains the Fourier transform result by multiplying the original signal with the Fourier transform matrix. The specific method is as follows: Matrix representation of Discrete Fourier Transform (DFT): For a finite-length sequence x[n] of N points, its discrete Fourier transform (DFT) is mathematically expressed as a multiplication-addition operation, as shown in the following formula:

[0025] The above process is represented by a matrix, and X represents the N-point DFT frequency domain sequence vector, and x represents the time domain sequence vector, then:

[0026] At this point, the DFT matrix of N points is defined as:

[0027] The N-point DFT of the sequence x[n] can be expressed as:

[0028] It can be seen that the number of DFT multiplications and additions of N points is the same as Directly proportional.

[0029] As mentioned earlier, NPU does not support complex number operations. Therefore, matrix multiplication needs to be split into real and imaginary parts for calculation. The calculation method is as follows: dimensional complex matrix A and The calculation of the product Y of the dimensional complex matrix B needs to be decomposed into real and imaginary parts. The process is as follows:

[0030] Where real(x) means taking the real part of variable x, and imag(x) means taking the imaginary part of variable x. Therefore, Y can be expressed as: .

[0031] (2) Multiply it by the reference signal matrix: Multiply it by the reference signal as follows: O=P × Q ,in P is the signal after DFT, Q is the reference signal, O is the result of multiplying the reference signal; because P, Q, O Are all complex matrices, first complex matrices P and complex matrices Q Decompose into real and imaginary parts respectively, and calculate the matrix P and matrix Q The product of O， The process is , where real( O ) means to take O The real part, imag( O ) means to take O The imaginary part of O It can be expressed as O = real(O) + imag(O)j .

[0032] (3) Perform IDFT processing again: The methods of iDFT and DFT are exactly the same.

[0033] Step 3: Moving target display: Calculate the moving target information through the moving target display; The implementation of Moving Target Indication (MTI) uses matrix multiplication to process the Doppler filter bank. By multiplying the MTI auxiliary matrix (MTIauxiMat) with the input signal matrix, the static and slow-moving clutter can be effectively filtered out, improving the detection capability of moving targets. The Doppler filter bank in MTI is implemented through matrix operations, and the position and speed of the moving target are quickly calculated by using the parallel processing capability of the NPU. The specific implementation method of Moving Target Indication (MTI) is as follows: Assuming that the order of the MTI filter is n, the coefficient vector of the MTI is:

[0034] in

[0035] Assume the number of radar pulses is , the number of distance gates is , then the received signal matrix The dimension is .

[0036] Constructing the auxiliary matrix To achieve MTI, The dimension is . , in, Elements in At this time, the result of MTI can be expressed as: .

[0037] The dimension is ,in ; This part will also encounter imaginary matrix calculations, and the matrix multiplication needs to be split into the real part and the imaginary part for calculation separately.

[0038] Step 4: Digital beamforming: Process the radar echoes of each radar channel to form sum beam and difference beam; The beamforming algorithm is implemented using matrix operations, and the required beam pointing direction is synthesized by adjusting the weight matrix. The method for implementing digital beamforming (DBF) is as follows: If radar is channels, the multi-channel data received by the radar can be expressed as: .

[0039] Assuming that n beams need to be formed, the DBF matrix is ​​constructed: , in, Indicates that the beam is pointed to The weight vector of .

[0040] Then the DBF process can be expressed as: ; This part will also encounter imaginary matrix calculations, and the matrix multiplication needs to be split into the real part and the imaginary part for calculation separately.

[0041] Step 5: performing coherent accumulation or non-coherent accumulation on the multi-pulse signal including the sum beam and the difference beam; Non-coherent accumulation is accomplished through matrix operations, which sums the signal matrix after pulse compression to enhance the signal strength. This method uses the parallel processing capability of the NPU to quickly calculate the total energy of the signal, thereby increasing the probability of target detection. The implementation method of non-coherent accumulation is as follows: Non-coherent accumulation is performed in the slow time dimension. Assuming that the radar data after pulse compression is , in Represents the data of each pulse, with dimension .

[0042] Then the non-coherent accumulation process can be expressed as:

[0043] in, is the result of non-coherent accumulation, with dimension .

[0044] Step 6: CFAR processing: The target position is calculated by the CFAR algorithm.

[0045] The implementation of constant false alarm rate (CFAR) uses convolution operation, which convolves the convolution kernel (ConvCore) with the auxiliary matrix (AuxiliaryMatrix) to adaptively determine the detection threshold. This method can maintain a stable false alarm rate in complex environments and improve the reliability of target detection. The sliding window and threshold calculation in the CFAR algorithm are converted into matrix operations, and the target detection threshold is calculated in real time using the parallel processing capability of the NPU. The implementation method of constant false alarm rate (CFAR) is as follows: Assume that the radar data to be detected is , the number of reference units of CFAR is , the number of protection units is .

[0046] Construct the CFAR convolution kernel: ; in It means to construct a row vector with N elements all set to 1. It means constructing a row vector whose N elements are all 0.

[0047] At this time, it is necessary to To perform zero padding on the edges:

[0048] Then the CFAR result can be expressed as:

[0049] Among them, conv(x,y) represents the convolution of x and y vectors.

[0050] The present invention provides a radar real-time target detection method based on heterogeneous parallel computing. Aiming at the problems existing in the existing radar real-time signal processing technology in terms of processing speed, accuracy, efficiency and flexibility, the present invention adopts domestic heterogeneous computing technology and applies general software technology to realize efficient and real-time processing of radar signals and improve the overall performance of the radar system. Finally, it should be noted that the above are only preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention is described in detail with reference to the embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments or to replace some of the technical features therein by equivalents. However, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A radar real-time target detection method based on heterogeneous parallel computing, characterized in that: The following steps are involved: Step 1: Channel correction: Each frame of radar echo is converted into a digital signal through AD, and the radar echo amplitude and phase characteristics of each radar channel are made consistent after channel correction; Step 2: Pulse compression: restore the radar echo narrow pulse signal through pulse compression; Step 3: Moving target display: Calculate the moving target information through the moving target display; Step 4: Digital beamforming: Process the radar echoes of each radar channel to form sum beam and difference beam; Step 5: performing coherent accumulation or non-coherent accumulation on the multi-pulse signal including the sum beam and the difference beam; Step 6: CFAR processing: The target position is calculated by the CFAR algorithm.

2. The radar real-time target detection method based on heterogeneous parallel computing according to claim 1 is characterized in that: The pulse compression in step 2 includes the following steps: S1: Perform DFT processing on radar echo data with consistent amplitude and phase characteristics after channel correction; S2: Multiply the radar echo data processed in step S1 and the reference signal matrix; S3: Perform iDFT processing again.

3. The radar real-time target detection method based on heterogeneous parallel computing according to claim 2 is characterized in that: In step S1, DFT processing is performed on the radar echo data with consistent amplitude and phase characteristics after channel correction, which includes the following steps: S4: For a finite length sequence x[n] of N points, its DFT is mathematically expressed as a multiplication and addition operation, and the formula is ; S5: The process of step S4 is represented by a matrix, and X represents the N-point DFT frequency domain sequence vector, and x represents the time domain sequence vector. Then , the DFT matrix of N points is defined as ; S6: The N-point DFT of the sequence x[n] can be expressed as ; S7: DFT processing is completed through matrix calculation in steps S4-S6. The matrices and vectors are all complex numbers. When implemented on the NPU, the complex matrix needs to be transformed into a real matrix calculation to obtain the final result of DFT processing of the radar echo data with consistent amplitude and phase characteristics after channel correction. The method is: Assume Y=A*B, A, B, Y Are all complex matrices, first complex matrices A and complex matrices B Decompose into real and imaginary parts respectively, and calculate the matrix A and matrix B The product of Y， The process is , where real(Y) means taking the real part of Y, and imag(Y) means taking the imaginary part of Y, then Y can be expressed as .

4. The radar real-time target detection method based on heterogeneous parallel computing according to claim 2 is characterized in that: In step S2, the radar echo data processed in step S1 is multiplied by the reference signal matrix, which includes the following steps: S8: O=P × Q ,in P is the signal after DFT, Q is the reference signal, O is the result of multiplying the reference signal; because P, Q, O Are all complex matrices, first complex matrices P and complex matrices Q Decompose into real and imaginary parts respectively, and calculate the matrix P and matrix Q The product of O， The process is , where real( O ) means to take O The real part, imag( O ) means to take O The imaginary part of O It can be expressed as O = real(O) + imag(O)j .

5. The radar real-time target detection method based on heterogeneous parallel computing according to claim 2 is characterized in that: In step S3, iDFT processing is performed again, including the following steps: S9: Repeat steps S4-S7.

6. The radar real-time target detection method based on heterogeneous parallel computing according to claim 1 is characterized in that: In step 3, the moving target information is calculated by displaying the moving target, including the following steps: S10: Assume that the order of the moving target display filter is n, then the coefficient vector of the moving target display is ,in ; S11: Assume the number of radar pulses is , the number of distance gates is , then the received signal matrix The dimension is ; S12: Construct auxiliary matrix To achieve MTI, The dimension is , ,in Elements in , then the result of MTI can be expressed as ; S13: The dimension is ,in .

7. The radar real-time target detection method based on heterogeneous parallel computing according to claim 1 is characterized in that: The digital beam forming in step 4 includes the following steps: S14: Assume that the radar has channels, the multi-channel data received by the radar can be expressed as ; S15: Assuming that n beams need to be formed, the digital beamforming matrix is ​​constructed as ,in, Indicates that the beam is pointed to The weight vector of ; S16: Combining step S14 and step S15, the digital beamforming process can be expressed as .

8. The radar real-time target detection method based on heterogeneous parallel computing according to claim 1 is characterized in that: In step 5, coherent accumulation or non-coherent accumulation is performed on the multi-pulse signal including the sum beam and the difference beam, including the following steps: S17: Non-coherent accumulation is performed in the slow time dimension. Assume that the radar data after pulse compression is ,in Represents the data of each pulse, with dimension , then the non-coherent accumulation process can be expressed as ,in is the result of non-coherent accumulation, with dimension .

9. The radar real-time target detection method based on heterogeneous parallel computing according to claim 1 is characterized in that: The constant false alarm processing in step 6 includes the following steps: S18: Assume that the radar data to be detected is , the number of reference units of CFAR is , the number of protection units is , construct the constant false alarm processing convolution kernel as ,in It means to construct a row vector with N elements all set to 1. It means constructing a row vector with N elements all being 0; S19: Yes To fill the edge with zero, the expression is , then the CFAR processing result can be expressed as , where conv(x,y) represents the convolution of x and y vectors.

Images