Method for parallel and rapid realization of fractal target detection

Abstract

The invention discloses a method for parallelly and rapidly realizing fractal target detection, which comprises the following steps of: acquiring a sea radar signal matrix, carrying out stipulation summation and normalization processing, then carrying out peak value search, and determining fractal detection points; constructing a corresponding fractal matrix for each fractal detection point; constructing different detection sequences and detection values for each fractal matrix; setting different detection signal-to-noise ratios, and judging each fractal detection point by adopting a decision tree for multi-level condition judgment so as to determine the signal-to-noise ratio corresponding to the fractal detection point; and performing target detection based on the size of the signal-to-noise ratio. In order to solve the problem of long time consumption when a fractal target detection algorithm is carried out by using a CPU platform, the fractal target detection algorithm is parallelized, calculation is carried out by using a GPU, the realization of the fractal detection algorithm is accelerated, the calculation speed is improved, and radar signal processing is rapidly realized to meet the real-time requirement of radar data.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of radar, and relates to a fast implementation of a sea radar signal data processing algorithm, in particular to a parallel fast implementation fractal target detection method, which uses the parallel computing capability of a GPU to accelerate the processing of sea radar signal data. BACKGROUND

[0002] With the development of electronic information technology, radar detection technology is also developing; along with the digitization and intelligentization of modern radars and more and more high-performance radar signal processing algorithms, the amount of data required for processing is also increasing, which puts certain performance requirements on the radar signal data processing equipment.

[0003] In traditional radar signal data processing equipment, FPGA and DSP are mostly used for radar signal data processing, but the hardware platform is complex to build and lacks flexibility, and it is difficult to deploy in the first time in the face of modern radar algorithm iteration, and the development cycle is long. Due to the demand for intelligentization and visualization, modern radar processing platforms mostly use PC platforms for radar signal processing, and pure CPU processing platforms consume a lot of time in processing large amounts of data and complex radar algorithms, and it is difficult to achieve real-time performance.

[0004] The CUDA (Compute Unified Device Architecture) parallel computing platform and programming model launched by NVIDIA make GPU devices responsible for graphics processing available as general-purpose computing devices, and the parallel processing capability of GPU can better process radar data. The cooperative processing platform of GPU and CPU takes CPU as the scheduling center and GPU as the computing center, which makes up for the lack of parallel computing capability of CPU and realizes the processing of radar signal data more quickly. In the face of increasing radar data and more and more complex radar algorithms, the cooperative processing platform of GPU and CPU can take advantage of the parallel processing capability of GPU to realize the processing of radar data more quickly and meet the real-time requirements of data.

[0005] In the data processing of sea radar signals, the signals received by the radar contain a large amount of sea clutter, and the use of traditional detection methods often results in a large number of false alarms. The fractal target detection algorithm can achieve better sea surface target detection effect, but the fractal target detection algorithm has a large amount of calculation, and pure CPU platform calculation cannot meet the real-time requirements of radar data. SUMMARY

[0006] The present application aims to provide a parallel fast implementation fractal target detection method, which solves the problem that the data processing capability of CPU is limited when using a PC as a radar data processing platform in modern sea detection radars, and the fractal detection algorithm with a large amount of calculation takes a long time.

[0007] To achieve the above-mentioned tasks, the present application adopts the following technical solutions: A parallel fast fractal target detection method, comprising: Obtaining a sea radar signal matrix, performing parallel reduction summation on all Doppler units corresponding to each range unit in the radar signal matrix, performing data normalization processing using the reduction summation result, and obtaining a normalized signal matrix; Based on the normalized signal matrix, performing parallel peak value search on the Doppler channel of each range unit, constructing a structure array for storing peak values and their position information; sorting each range unit using the structure array, and selecting a preset number of peak values as fractal detection points; For each fractal detection point, a corresponding fractal matrix is constructed; for each fractal matrix: Taking the matrix on the cross angle centered on the fractal detection point, and obtaining the maximum value of each row of data in the matrix, from which the global maximum value is determined to construct a feature peak value sequence; Constructing a detection region centered on the fractal detection point, extracting the maximum value of each row in the detection region using threads, and determining the global maximum value of these maximum values to construct a maximum value array; Taking a plurality of data around the fractal detection point in a preset rule, and calculating the maximum value, minimum value, and ratio of the fractal point to the average value of the data; Setting the required detection signal-to-noise ratio for temporary track association data, the required detection signal-to-noise ratio for stable breakpoint track association, and the required detection signal-to-noise ratio for stable continuous track association; based on these detection signal-to-noise ratios, a multi-level condition judgment decision tree is used to determine the signal-to-noise ratio corresponding to each fractal detection point; based on the size of the signal-to-noise ratio, target detection is performed.

[0008] Further, performing parallel reduction summation on all Doppler units corresponding to each range unit in the radar signal matrix, performing data normalization processing using the reduction summation result, and obtaining a normalized signal matrix, comprising: Putting the data in each range unit in the radar signal matrix into an array, and using threads to add every two adjacent data in the array; adding every two adjacent results again to obtain the final reduction summation result, and storing all summation results According to the order of the range unit; The reduction summation result of the first range unit; wherein represents the total number of Doppler units in the radar signal matrix, represents the total number of range units, represents the first range unit, the first Data at each Doppler unit; open up Each of the n threads independently computes the nth operation. The distance unit, the first Normalized values ​​of data at each Doppler cell .

[0009] Furthermore, a parallel peak search is performed on the Doppler channels of each distance unit to construct a structure array for storing the peaks and their location information, including: Parallel peak search is performed on the Doppler channels of each range cell; before-and-after comparisons are made on the Doppler channel data of each range cell: if think This data represents the peak value. Create and initialize an array of structures, the size of which is equal to the size of the array. Consistent; the members in the structure array are numerical values ​​and position numbers, and the original position of each distance unit is determined; peak search is performed simultaneously on the Doppler channels of all distance units, and after the peak value is determined, it is assigned to the corresponding numerical segment of the structure array; otherwise, the amplitude of the corresponding numerical segment is set to 0.

[0010] Furthermore, in the structure array after peak search, the data corresponding to each distance unit is sorted in ascending order using a bitonic sorting algorithm; the positions of the peaks in the structure array are swapped according to the sorting results and their positions are retained; the 20 peaks with the highest energy in each distance unit are selected as fractal detection points.

[0011] Furthermore, a corresponding fractal matrix is ​​constructed for each fractal detection point, including: The fractal matrix in the th OK Column elements are ;in, and These represent the ranges of the distance unit and the Doppler unit, respectively, set during fractal processing. Based on the position of the fractal detection point in the structure array, the area surrounding the fractal detection point is extracted. )× The data is used as the fractal detection matrix for that fractal detection point, and a thread of the corresponding matrix size is opened for each fractal matrix to move the data.

[0012] Furthermore, for each fractal matrix, four points are taken at the cross corners centered on the fractal detection point. Matrix, for each The matrix is ​​compared using 3 threads, each thread being used to... The maximum value in each row of the matrix is extracted, and each fractal matrix can generate 12 maximum values; all of the 12 maximum values extracted from the matrix are subjected to maximum value judgment by one thread to obtain a global maximum value; the global maximum values extracted from all fractal matrices are stored in sequence according to the order of the fractal matrices to form a characteristic peak value sequence . .

[0013] Further, for each fractal matrix, a 5 range centered on the fractal detection point is taken as a detection region, and the data in the detection region is grouped and processed by row, 5 threads are allocated to the detection region of each fractal matrix, each thread is responsible for comparison operation of all data elements in the corresponding row, and the maximum value of the row is extracted; 1 thread is enabled to perform a secondary comparison operation on the maximum values output by each row in the detection region to determine the global maximum value in the entire detection region; the global maximum values extracted for all fractal matrices are sequentially stored in a maximum value array according to the storage order of the fractal matrices .

[0014] Further, for each fractal matrix, a 5 The four data of the four top corners of the 5 The four data of the four top corners of the 5 The four data of the four top corners of the 5 , minimum value , fractal detection and 12 data average value ratio are extracted by 12 threads at a time, and 3 threads are used to perform parallel operations and store them in sequence according to the storage order of the fractal matrices.

[0015] Further, based on the signal-to-noise ratio of these detections, a decision tree with multiple levels of conditional judgment is used to determine the signal-to-noise ratio corresponding to each fractal detection point, including: The decision tree with multiple levels of conditional judgment is specifically constructed as follows: The first priority condition: when is less than 0.1, is less than or equal to 0.99, and is equal to 1, it is considered that the radar signal quality is high; at this time, the relationship between and (SNR_temp_track + 0.1) is compared, and the larger one is selected as the signal-to-noise ratio corresponding to the fractal detection point ; wherein SNR_temp_track represents the required detection signal-to-noise ratio of the temporary track correlation data; The second priority condition: when SNR_temp_track and is equal to 1, directly set to (SNR_temp_track+0.1); The third to fifth priority conditions correspond to different signal quality levels respectively: less than 0.2 and a high-quality signal less than or equal to 0.99; less than 0.3 and a medium-quality signal less than 0.6; less than 0.4 and a low-quality signal less than 0.5; For these cases, the same processing strategy is adopted: under the premise of meeting SNR_temp_track and is equal to 1, compare with SNR_temp_track relationship, select the larger one as ; The sixth priority condition: when less than 0.7, less than 0.8 and SNR_lose_track, while meeting is equal to 1, it is judged as a critical tracking state, and is set to (SNR_temp_track+0.1); wherein SNR_lose_track represents the required detection signal-to-noise ratio of the stable breakpoint track correlation; The seventh priority condition: when less than 0.7, less than 0.8 and is equal to 1, according to the difference comparison result of and SNR_steady_track, select the target value closest to the difference between SNR_lose_track and SNR_steady_track, and assign to the target value+0.1.

[0016] A terminal device comprising a processor, a memory, and a computer program stored in the memory; when the processor executes the computer program, the parallel fast fractal target detection method is realized.

[0017] A computer readable storage medium, wherein the medium stores a computer program; the computer program is executed by a processor to implement the parallel fast fractal target detection method.

[0018] Compared with the prior art, the application has the following technical characteristics: 1. The fractal detection algorithm is parallelized in the application, the serial calculation existing in the fractal detection algorithm is disassembled, the algorithm is used to decouple some strongly coupled calculation processes to reduce the number of serial calculations, the parallel calculation capability of the GPU is utilized to perform parallel calculation on a large amount of data, the intermediate data required for calculation is uniformly calculated, and the result is stored in the GPU end, and finally the final signal-to-noise ratio decision is made after all the intermediate data calculation is completed, so that the fractal detection algorithm is accelerated.

[0019] 2. The fractal target detection algorithm is parallelized in the application, the GPU is used for calculation, the fractal detection algorithm is accelerated, the calculation speed is improved, the radar signal processing is rapidly realized to meet the real-time requirement of radar data. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 is a GPU processing flowchart of the method; Figure 2 is a GPU reduction summation schematic diagram of the method; Figure 3 is a GPU parallel reduction summation operation total energy schematic diagram of each distance unit of the method; Figure 4 is a GPU multi-thread peak value search schematic diagram of the method; Figure 5 is a GPU end effective peak value data storage mode schematic diagram of the method; Figure 6 is a GPU double sorting schematic diagram of the method; Figure 7 is a GPU end fractal matrix storage mode schematic diagram of the method; Figure 8 is a GPU multi-thread search cross matrix maximum value in the fractal matrix schematic diagram of the method; Figure 9 is a GPU multi-thread search maximum value around the detection point schematic diagram of the method; Figure 10 is a GPU multi-thread calculation different separation distance maximum value and average value schematic diagram of the method; Figure 11 is a GPU operation and CPU operation error diagram of the method. DETAILED DESCRIPTION

[0021] The application discloses a kind of parallel fast implementation fractal target detection method, mainly solve modern sea detection radar using PC as radar data processing platform, the data processing capability of CPU is limited, in the long time of large amount of calculation fractal detection algorithm problem.This application will fractal detection algorithm parallelization, the serial computation in fractal detection algorithm is disassembled, and some strong coupling calculation process is decoupled to reduce the number of serial computation;Using the parallel computing capacity of GPU, large quantities of data parallel computing is carried out, the intermediate data required for calculation is uniformly operated, and the results are stored in the GPU end, and finally the final signal-to-noise ratio decision is made after all intermediate data operation is completed, so as to speed up the implementation of fractal detection algorithm.

[0022] The application provides a kind of parallel fast implementation fractal target detection method, fractal target detection algorithm parallelization and using GPU for calculation, to speed up the implementation of fractal target detection algorithm, improve the calculation speed, quickly realize sea radar signal processing to meet the real-time requirement of radar data;The GPU parallelization method of each step of fractal target detection algorithm is as follows: Step 1, data normalization calculation.

[0023] Step 1.1, obtain the sea radar signal matrix, all doppler units corresponding to each distance unit in the radar signal matrix are parallelly reduced sum, as shown in Figure 2 ; The reduction sum can accelerate the time of long array addition in GPU operation, in this step, the data in each distance unit is put into an array, and the data in the array is added every adjacent two using threads;All addition results are added every adjacent two again until the final reduction sum result is obtained, and all sum results are stored according to the order of distance unit.

[0024] The reduction sum can accelerate the time of long array addition in GPU operation;In the process of using GPU reduction sum for sea radar signal matrix, as shown in Figure 3 , each distance unit is parallelly operated by reduction sum, the total number of calculations is , the total number of threads required is , and the reduction sum result of the th distance unit is ; Wherein represents the total number of doppler units in the radar signal matrix, represents the total number of distance units, represents the data at the th distance unit and the th doppler unit in the radar signal matrix.

[0025] Step 1.2, open threads, each thread independently operates the normalization of data at the first distance unit, the second Doppler unit , the total number of calculations is one; each distance unit is indexed by the index of the thread as follows:

[0026] wherein, indicates a thread block, indicates a thread grid, indicates a thread; and indicates the and directional coordinate position of the thread grid in which the current thread block is located; indicates the dimension of the direction of the thread grid, and indicates the dimension of the and direction of the current thread block, and indicates the and directional coordinate position of the thread block in which the current thread is located; is the total number of threads.

[0027] Each distance unit corresponds to the total number of threads, is the index number of , which is also the index number of the distance unit; the ratio of to the modulus is the normalization result , and each thread only needs to perform one normalization operation.

[0028] Step 2, perform parallel peak search on the distance unit.

[0029] Step 2.1, perform parallel peak search on the Doppler channel of each distance unit; the parallel peak search is as shown in Figure 4 , for each distance unit of the Doppler channel data, compare the front and back: if consider the data (data at the first distance unit, the first Doppler unit) as the peak value.

[0030] As shown in Figure 5 , create and initialize an array of structure bodies, the size of the array is equal to Consistent; the members of the structure array are numerical values ​​and position indices, and the original position of each distance unit is determined; in GPU computation, the Doppler channel data of each distance unit uses... Each thread performs the computation, using a total of [number] threads. One thread; perform peak search on the Doppler channels of all distance cells simultaneously. Each thread is responsible for comparing the current data with the data before and after it. If the current data is greater than the data before and after it, it is the peak value. Then, the peak value is assigned to the corresponding numerical segment of the structure array. Otherwise, the amplitude of the corresponding numerical segment is set to 0. After the peak detection is completed, a structure array is obtained in which all values ​​except the peak position are 0.

[0031] Step 2.2: Sort the data corresponding to each distance cell in the structure array after peak search in ascending order using the bitonic sorting algorithm; swap the positions of the peaks in the structure array according to the sorting results and retain their indexes; each Data opening ( ) threads perform indexing, if Zeros must be padded if the expression does not satisfy a power of 2; the time complexity of parallel bitonic sort is O(n). For each distance unit, the 20 peak values ​​with the highest energy are selected as fractal detection points.

[0032] A diagram of a bitone sort is shown below. Figure 6 As shown, the 8 data points in the distance unit are sorted by first breaking the sequence into sequences with opposite polarities, and then performing a bitonic sort. In the bitonic sort, the direction of each arrow represents ascending or descending order, and each arrow represents a thread. Each thread operates independently, and there is no conflict between threads. The Doppler channels of each distance unit are sorted simultaneously.

[0033] Step 3: Construct fractal processing data.

[0034] Step 3.1: Generate a corresponding fractal matrix for each fractal detection point. The fractal matrix contains the fractal matrix with the first fractal point as the first fractal point. OK Column elements ;in, and These represent the ranges of the distance unit and the Doppler unit, respectively, set during fractal processing. The size of the fractal matrix is ​​then... Based on the position of the fractal detection point in the structure array, the area surrounding the fractal detection point is extracted. )× The data is used as the fractal detection matrix for that fractal detection point. A thread corresponding to the size of each fractal matrix is ​​created for data transfer, thus forming a matrix group from all fractal matrices in sequence, such as... Figure 7 As shown.

[0035] In one embodiment of the present invention, it is set That is, taking the original data as the center of the fractal detection point. middle A fractal matrix of size. A fractal matrix is ​​generated for all fractal detection points. For each fractal detection point, 11 threads are used for data migration, and each thread migrates the data 11 times. All generated fractal matrices are stored as follows: Figure 7 As shown, it is stored in column-major order.

[0036] Step 3.2: For each fractal matrix, take 4 points at the cross corners centered on the fractal detection point. Matrix, for each The matrix is ​​compared using 3 threads, such as Figure 8 As shown; each thread is used to... Extracting the maximum value from each row of the matrix, each fractal matrix can generate 12 maximum values, requiring 12 threads for comparison and data movement; [and then...] The 12 maximum values ​​extracted from the matrix are then used by a single thread to determine the global maximum value. All global maximum values ​​extracted from the fractal matrices are stored in the order of the fractal matrices, forming a characteristic peak sequence. .

[0037] like Figure 8 As shown in the diagram, the arrows in the fractal matrix represent a thread that extracts the 12 maximum values ​​from all fractal matrices. Finally, another thread is used to determine and extract the maximum value, arranging them according to the order of the fractal matrices. .

[0038] Step 3.3: For each fractal matrix, take the fractal detection point as the center. The range is used as the detection area. The data within the detection area is grouped and processed by row. Five threads are allocated to the detection area of ​​each fractal matrix. Each thread is responsible for the comparison operation of all data elements in the corresponding row and extracting the maximum value of the row. Figure 9 The horizontal arrows indicate the row scanning direction of each thread; one thread is used to perform a second comparison operation on the maximum value output of each row within the detection area to determine the global maximum value within the entire detection area; the global maximum values ​​extracted for all fractal matrices are stored sequentially into the maximum value array according to the storage order of the fractal matrices. .

[0039] Step 3.4, for each fractal matrix, 5 points centered on the fractal detection point. The four data points at the four vertices of the 5 matrix are numbered 1-4 in a clockwise direction; the 7 data points centered on the fractal detection point... 5 The four data of the four top corners of the matrix are numbered 5-8 in clockwise direction, 5 7 The four data of the four top corners of the matrix are numbered 9-12 in clockwise direction, the data position is shown as Figure 10 ; each fractal matrix uses 12 threads to take out the 12 data above at one time, then uses 3 threads to complete the operation of maximum , minimum , fractal detection and the ratio of 12 data average value in parallel, and stores according to the storage order of the fractal matrix.

[0040] Step 4, SNR calculation and target detection.

[0041] Step 4.1, the required detection SNR of temporary track association data SNR_temp_track, the required detection SNR of stable breakpoint track association SNR_lose_track, and the required detection SNR of stable continuous track association SNR_steady_track are fixed values that can be adjusted freely according to different environments; for each fractal detection point, a decision tree with multiple levels of conditional judgment is used to determine the corresponding SNR according to different signal feature combinations : First priority condition: when is less than 0.1, is less than or equal to 0.99, and is equal to 1, it is considered that the radar signal quality is high; at this time, the size relationship between and (SNR_temp_track + 0.1) is compared, and the larger one is selected as , to ensure the optimal tracking performance under high-quality signal conditions.

[0042] Second priority condition: when exceeds SNR_temp_track and is equal to 1, directly set to (SNR_temp_track + 0.1) to ensure the stable tracking ability of the system.

[0043] The third to fifth priority conditions correspond to different signal quality levels: medium-high quality signal when is less than 0.2 and is less than or equal to 0.99; medium quality signal when is less than 0.3 and is less than 0.6; Medium to low quality signals less than 0.5; For all these situations, the same processing strategy is adopted: ... (The sentence is incomplete and requires more context to be translated accurately.) Exceeding SNR_temp_track and Under the premise that it equals 1, compare In relation to SNR_temp_track, choose the larger one. To ensure tracking performance.

[0044] Sixth priority condition: when Less than 0.7 Less than 0.8 and When exceeding SNR_lose_track, simultaneously satisfying If the value equals 1, it is determined to be a critical tracking state, and... Set to (SNR_temp_track+0.1).

[0045] Seventh priority condition (lowest priority): When Less than 0.7 Less than 0.8 and When equal to 1, according to Compare the difference between SNR_steady_track and SNR_lose_track, select the target value that is closest to the difference, and assign the value. (Target value + 0.1).

[0046] Each fractal detection point uses a separate thread to make decisions using the decision tree, and the total signal-to-noise ratio is obtained after parallel computation. The target can be determined based on the signal-to-noise ratio; for example, if the signal-to-noise ratio is greater than a set threshold, then the target is considered to exist.

[0047] Example: (1) Experimental verification conditions: The performance of the present invention is tested below. The method provided by the present invention uses Visual Studio 2019 and CUDA version 12.4 programming under the Windows Server 2019 system as the software framework, and the hardware is a maritime intelligent radar data processing platform. The CPU uses Intel(R) Xeon(R) Silver4310 and the GPU uses NVIDIA Telsa A100. The experimental data of the maritime intelligent radar is used as the test data.

[0048] (2) Experimental Results Figure 11For the error map of using GPU for fractal target detection algorithm operation and CPU for fractal target detection algorithm operation, the result of the accelerated operation according to the above method is consistent with the result of the pure CPU operation without error. Considering the problem of hardware cold start, in the process of time comparison between CPU and GPU, the first 10 frames of time comparison are removed, and the time of the last 10 frames of operation is taken for comparison. Table 1 records the time consumption required for data operation of the last 10 frames after the device is started. By using the method, the operation time of the fractal target detection algorithm can be greatly accelerated, and the real-time performance of the radar data is guaranteed.

[0049] Table 1 GPU and CPU operation time record table

[0050] The above examples are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A parallel fast implementation of a fractal target detection method, characterized in that, include: The radar signal matrix for maritime surveillance is obtained. All Doppler cells corresponding to each range cell in the radar signal matrix are reduced and summed in parallel. The reduction and summation results are used to perform data normalization processing to obtain the normalized signal matrix. Based on the normalized signal matrix, parallel peak search is performed on the Doppler channel of each range cell to construct a structure array for storing peak values ​​and their location information; the structure array is used to sort each range cell and a preset number of peak values ​​are selected as fractal detection points. Construct a corresponding fractal matrix for each fractal detection point; for each fractal matrix: Take a matrix at the cross angle centered on the fractal detection point, and obtain the maximum value of each row of data in the matrix. Determine the global maximum value from this matrix to construct a feature peak sequence. A detection region is constructed with the fractal detection point as the center. In each row of the detection region, the maximum value is extracted using a thread, and the global maximum value of these maximum values ​​is determined to construct a maximum value array. Centered on the fractal detection point, multiple data points are extracted around the fractal detection point according to preset rules, and the maximum value, minimum value, and ratio of the fractal point to the average value of these data points are calculated. The required detection signal-to-noise ratio (SNR) for temporary track association data, the required detection SNR for stable breakpoint track association, and the required detection SNR for stable continuous track association are set. Based on these detection SNRs, a multi-level conditional decision tree is used to determine the SNR corresponding to each fractal detection point. Target detection is then performed based on the magnitude of the SNR.

2. The parallel fast implementation of fractal target detection method according to claim 1, characterized in that, For each range cell in the radar signal matrix, all Doppler cells are reduced and summed in parallel. The results of the reduction and summation are then used for data normalization to obtain the normalized signal matrix, including: The data in each range cell in the radar signal matrix is put into an array, and the data in the array are added by two adjacent data using threads; all the addition results are added by two adjacent data again until the final reduction summation result is obtained, and all the summation results are stored in a result array in the order of range cells The reduction summation result of the i-th range cell is stored in the i-th position of the result array , where represents the total number of Doppler cells in the radar signal matrix, represents the total number of range cells, represents the data at the i-th range cell and the j-th Doppler cell in the radar signal matrix .​​ opening up one thread, each thread independently operating on normalized values of data at the first distance unit, the first Doppler unit .

3. The parallel fast implementation of fractal target detection method according to claim 1, wherein, Parallel peak search is performed on the Doppler channels of each range cell, and a structure array is constructed to store the peaks and their location information, including: performing a parallel peak search on the Doppler channel for each range cell; performing a forward-backward comparison on the Doppler channel data for each range cell; if it is determined that the data is a peak; Create and initialize an array of structures, the size of which is equal to the size of the array. Consistent; the members in the structure array are numerical values ​​and position numbers, and the original position of each distance unit is determined; peak search is performed simultaneously on the Doppler channels of all distance units, and after the peak value is determined, it is assigned to the corresponding numerical segment of the structure array; otherwise, the amplitude of the corresponding numerical segment is set to 0.

4. The parallel and rapid fractal target detection method according to claim 1, characterized in that, After peak search, the data corresponding to each distance unit in the structure array is sorted in ascending order using a bitonic sorting algorithm. The positions of the peaks in the structure array are swapped according to the sorting results, and their positions are retained. The 20 peaks with the highest energy in each distance unit are selected as fractal detection points.

5. The parallel and rapid fractal target detection method according to claim 1, characterized in that, Construct a corresponding fractal matrix for each fractal detection point, including: The fractal matrix in the th OK Column elements are ;in, and These represent the ranges of the distance unit and the Doppler unit, respectively, set during fractal processing. Based on the position of the fractal detection point in the structure array, the area surrounding the fractal detection point is extracted. )× The data is used as the fractal detection matrix for that fractal detection point, and a thread of the corresponding matrix size is opened for each fractal matrix to move the data.

6. The parallel and rapid fractal target detection method according to claim 1, characterized in that, For each fractal matrix, four points are selected at the cross corners centered on the fractal detection point. Matrix, for each The matrix is ​​compared using 3 threads, each thread being used to... If the maximum value is extracted from each row of the matrix, then each fractal matrix can produce 12 maximum values; [and all...] The 12 maximum values ​​extracted from the matrix are then used by a single thread to determine the global maximum value. All global maximum values ​​extracted from the fractal matrices are stored in the order of the fractal matrices, forming a characteristic peak sequence. .

7. The parallel and rapid fractal target detection method according to claim 1, characterized in that, For each fractal matrix, with the fractal detection point as the center... The detection area is defined as the range. Data within the detection area is grouped and processed by row. Five threads are allocated to the detection area of ​​each fractal matrix. Each thread is responsible for comparing all data elements in its corresponding row and extracting the maximum value of that row. One thread is then used to perform a second comparison operation on the maximum values ​​output in each row within the detection area to determine the global maximum value for the entire detection area. The global maximum values ​​extracted for all fractal matrices are then stored sequentially in the maximum value array according to the storage order of the fractal matrices. .

8. The parallel and rapid fractal target detection method according to claim 1, characterized in that, For each fractal matrix, 5 points are centered on the fractal detection point. The four data points at the four vertices of the 5 matrix are numbered 1-4 in a clockwise direction; the 7 data points centered on the fractal detection point... The four data points at the four vertices of the 5 matrix are numbered 5-8 in a clockwise direction, with the fractal detection point as the center. The four data points at the four vertices of the 7-matrix are numbered 9-12 in a clockwise direction; for each fractal matrix, 12 threads retrieve the 12 data points at once, and then 3 threads work in parallel to maximize the result. Minimum value The ratio of fractal detection to the average of 12 data points The operations are performed and stored in the order of fractal matrix storage.

9. The parallel and rapid fractal target detection method according to claim 1, characterized in that, Based on these detected signal-to-noise ratios (SNRs), a multi-level decision tree with conditional judgments is used to determine the SNR corresponding to each fractal detection point, including: The decision tree for multi-level conditional judgments is constructed as follows: First priority condition: when Less than 0.1 Less than or equal to 0.99 and When the value equals 1, the radar signal quality is considered to be relatively high; at this point, the comparison... The larger of the values ​​(SNR_temp_track + 0.1) and (SNR_temp_track + 0.1) is chosen as the signal-to-noise ratio corresponding to the fractal detection point. SNR_temp_track represents the required signal-to-noise ratio for detecting temporary track association data; Second priority condition: when Exceeding SNR_temp_track and When it equals 1, directly... Set to (SNR_temp_track+0.1); The third to fifth priority conditions each correspond to different signal quality levels: Less than 0.2 and Medium to high quality signals less than or equal to 0.99; Less than 0.3 and Medium quality signal less than 0.6; Less than 0.4 and Medium to low quality signals less than 0.5; For all these situations, the same processing strategy is adopted: ... (The sentence is incomplete and requires more context to be translated accurately.) Exceeding SNR_temp_track and Under the premise that it equals 1, compare In relation to SNR_temp_track, choose the larger one. ; Sixth priority condition: when Less than 0.7 Less than 0.8 and When exceeding SNR_lose_track, simultaneously satisfying If the value equals 1, it is determined to be a critical tracking state, and... Set to (SNR_temp_track+0.1); where SNR_lose_track represents the detection signal-to-noise ratio required for stable breakpoint track association; Seventh priority condition: when Less than 0.7 Less than 0.8 and When equal to 1, according to Compare the difference between SNR_steady_track and SNR_lose_track, select the target value that is closest to the difference, and assign the value. The target value is +0.

1.

10. A computer-readable storage medium storing a computer program; characterized in that, When the computer program is executed by the processor, it implements the parallel and rapid fractal target detection method according to any one of claims 1-9.

Images