Compact high-frequency ground wave radar target detection method based on multi-frame information

Abstract

The invention belongs to the technical field of radio physics and image recognition, and provides a compact high-frequency ground wave radar target detection method based on multi-frame information. The method comprises the following steps: generating radar distance-Doppler spectrum images based on radar frequency domain data, performing target detection on each frame of distance-Doppler spectrum image by a first network, and taking a region suspected to contain a target as an interested image region of each frame of image; defining an image block of each frame of image by taking the interested image area as a center, and forming an image block sequence by the image blocks of multiple frames of images; inputting the image block sequences into a second network, flattening image blocks of each frame of image in the image block sequences into vectors, adding position codes to form image block sequence tokens, and inputting the token of each image block sequence and a learnable class token into the second network; and performing target detection by adopting the trained first network model and the second network model.

Description

Technical Field

[0001] The present invention belongs to the technical fields of radiophysics and image recognition, and provides a compact high-frequency ground wave radar target detection method based on multi-frame information. Background Art

[0002] High-frequency surface wave radar (HFSWR) is a long-distance ocean monitoring technology based on high-frequency electromagnetic wave bands. HFSWR utilizes the characteristics of high-frequency vertically polarized waves propagating along the coast on the earth's surface. After the electromagnetic wave is emitted towards the sea surface, it extends below the horizon along the curved earth's surface in the form of surface waves (ground waves), achieving over-the-horizon target detection. The maximum operating range of HFSWR can reach 600 km, and the operating frequency is 3 - 30 MHz. It has the advantages of over-the-horizon detection, anti-stealth, and all-weather operation, and can be used to monitor sea conditions and ship targets, safeguard the rights and interests of the exclusive economic zone (EEZ) of the ocean, etc., and has key application scenarios in both military and civilian fields. In practical applications, compact high-frequency ground wave radars can be deployed on the shore of islands or limited sites, and can also accompany ships into the deep sea or remote waters to expand the monitoring range and be flexibly deployed. Therefore, they have unique advantages in terms of mobility and flexibility.

[0003] As one of the important application fields of high-frequency ground wave radar, the technical development and application prospects of target detection have always been a research hotspot in the academic and industrial circles. In the development process of target detection technology, two major categories of methods have mainly emerged: traditional target detection methods and learning-based target detection methods. Among them, traditional target detection methods mainly improve traditional algorithms, with relatively complete theoretical derivations, strong theoreticality and interpretability. However, in order to adapt to different detection environments and situations, various complex theoretical tools are often required, and it is difficult to extract non-linear information in images; learning-based target detection methods benefit from the introduction of complex model structures, and the model can learn the non-linear features hidden in ship targets and background clutter, which is conducive to providing better target discrimination ability in actual non-stationary environments. However, affected by coherent accumulation and target maneuvering, the target signal may show dynamic changes in the range dimension and velocity dimension, and the existing learning-based target detection methods do not fully utilize the change information of target features in the time domain in radar images. Summary of the Invention

[0004] The object of the present invention is to propose a two-stage target detection framework based on multi-frame ROI extraction to fully utilize the features of the target in three dimensions and achieve refined target detection for the problems that traditional target detection methods are difficult to extract non-linear information of images, and the learning-based target detection methods currently mainly focus on detecting single-frame RD spectra and fail to utilize the characteristic information of the target in the time domain, which affects the accuracy of target detection results.

[0005] To achieve the above object, in some embodiments of the present invention, the following technical solutions are provided:

[0006] A compact high-frequency ground wave radar target detection method based on multi-frame information, characterized by comprising the following steps:

[0007] S1: Based on the radar frequency domain data, generate a radar range-Doppler spectrum image and construct a range-Doppler spectrum image data set;

[0008] S2: Input the range-Doppler spectrum image into the first network. The first network performs target detection on each frame of the range-Doppler spectrum image, and takes the region suspected to contain the target as the image region of interest for each frame of the image; centered on the image region of interest, delimit the image blocks of each frame of the image, and form an image block sequence from the image blocks of multiple frames of images; during training, the first network learns the input data to train the network model of the first network;

[0009] S3: Input the image block sequence into the second network. Flatten the image blocks of each frame of the image in the image block sequence into vectors, add position encoding, and form an image block sequence token. Add a learnable class token to each image block sequence, and input the token and class token of each image block sequence into the second network; during training, the second network learns the input data to train the network model of the second network;

[0010] S4: Use the trained first network model and second network model for target detection.

[0011] In some embodiments of the present invention, the step S2 further includes: associating the targets in each frame of the range-Doppler spectrum image;

[0012] Determine the positions corresponding to each of the image regions of interest in the current frame, determine the distance of the center of each image region of interest from the target relative to the radar station, and the radial velocity of the target relative to the radar station; set the maximum motion state of the target, where the maximum motion state includes the maximum turning rate and the maximum acceleration, and based on the target stationary state and the maximum motion state, obtain the position range of the corresponding target in the next frame adjacent to the current frame; based on the matching between the position of the target in the current frame and the position range of the target in the next frame, associate the targets with adjacent positions, intercept the image regions of interest corresponding to the targets as image blocks, and obtain a sequence of image blocks;

[0013] If the corresponding target cannot be found in the next frame adjacent to the current frame, retain the position of the target in the current frame as the imaginary position of the target in the next frame for target association between the next frame and the frame after the next frame;

[0014] In step S3, input the sequence of image blocks obtained by processing the associated targets into the second network.

[0015] In some embodiments of the present invention, the steps of associating the targets in the range-Doppler spectral images of each frame further include:

[0016] Taking three consecutive frames of range-Doppler spectral images as a group, for each target in the first frame of range-Doppler spectral image, if:

[0017] The target is detected in the first frame of range-Doppler spectral image and all previous frames of range-Doppler spectral images have detected the target, or,

[0018] The target is detected in the first frame of range-Doppler spectral image and any previous frame of range-Doppler spectral image has detected the target, or,

[0019] The target is detected in the first frame of range-Doppler spectral image and all previous frames of range-Doppler spectral images have not detected the target, or,

[0020] The target is not detected in the first frame of range-Doppler spectral image and all previous frames of range-Doppler spectral images have detected the target;

[0021] Then in step S3, input the sequence of image blocks of this target in the three consecutive frames of range-Doppler spectral images into the second network, otherwise, discard the sequence of image blocks of this target in the range-Doppler spectral image.

[0022] In some embodiments of the present invention, in step S1, the steps of constructing the range-Doppler spectral image dataset further include:

[0023] Perform target extraction on each frame of the range-Doppler spectral image, and delimit the real target position area of each frame of the image, including:

[0024] Perform constant false alarm rate detection on each frame of range-Doppler spectrum image to extract all visible targets;

[0025] Use one of the following methods or a combination of the following methods to correct the constant false alarm rate detection process:

[0026] Reference cell bilateral clipping method: For each frame of range-Doppler spectrum image, perform constant false alarm rate detection for each minimum resolution cell. Sort the image cells according to the magnitude values of each image cell in the reference cell of the detector, remove a part of the image cells at the maximum magnitude end and the minimum magnitude end according to a set ratio, calculate the average value of the magnitude values of each image cell in the reference cell, and obtain the trimmed mean of the reference cell, which is used to calculate the detection threshold of the constant false alarm rate detector;

[0027] Upper half reference cell detection method: Set a range threshold, which is used to represent the range of the radar detection blind area. Consider the range-Doppler spectrum image area less than the range threshold as the near-range blind area, select the area above the near-range blind area as the target area for using the upper half reference cell detection method, and calculate the average value of the magnitude values of each image cell in the upper half area of the reference cell, which is used to calculate the detection threshold of the constant false alarm rate detector.

[0028] In some embodiments of the present invention, in step S1, the step of constructing the range-Doppler spectrum image data set further includes:

[0029] Perform target extraction on each frame of the range-Doppler spectrum image, and demarcate the real target position area of each frame of image, and further include:

[0030] Map the ship target information received by AIS to the range-Doppler spectrum image to obtain AIS targets;

[0031] Eliminate the targets in the non-radar detection area and the targets in the clutter interference area in the AIS targets to obtain real AIS targets;

[0032] Match all the visible targets extracted by constant false alarm rate detection with all the real AIS targets mapped by AIS information, and use the targets that can be matched as the final targets;

[0033] Perform correction processing on the final targets;

[0034] In step S2, input the data after correction processing into the first network.

[0035] In some embodiments of the present invention, the method for eliminating the targets in the non-radar detection area in the AIS targets includes:

[0036] Calculate the angle between the line connecting the AIS target and the radar relative to the normal of the radar antenna array, and eliminate the targets outside the radar detection area;

[0037] According to the ship length of the AIS target, eliminate the targets with incorrect ship length information;

[0038] Estimate the ratio of the radar cross section corresponding to the AIS target to the theoretically maximum radar cross section, and eliminate the targets whose radar cross section ratio is not within the set range.

[0039] In some embodiments of the present invention, the method for eliminating the targets in the clutter interference area among the AIS targets includes:

[0040] Calculate the theoretical positions of sea clutter and ground clutter in the range-Doppler spectral image, use the region growing algorithm to cluster the constant false alarm rate detection results, search for the occurrence positions of large clutter such as ionospheric clutter and radio frequency interference from the clustering results, and eliminate the targets located in various clutter and interference areas.

[0041] In some embodiments of the present invention, in step S2, input the range-Doppler spectral image data set into the trained first network, find real targets and false targets from the output results of the first network, and generate image patches to input into the second network, including:

[0042] Divide the sequence of image patches output by the first network into positive samples and negative samples;

[0043] The positive sample data includes the targets correctly detected in the current frame, the targets that were not detected in the current frame but were correctly detected in any previous frame, and the targets that were incorrectly detected in the current frame but were correctly detected in all previous frames;

[0044] The negative sample data includes the targets that were incorrectly detected in the current frame and no correct detection results exist in any previous frame;

[0045] In step S3, combine the positive samples and negative samples and input them into the second network.

[0046] In some embodiments of the present invention, the first network is a YOLOv8 network.

[0047] In some embodiments of the present invention, the second network is a Vision Transformer network.

[0048] Compared with the prior art, the beneficial effects of the technical solution of the present invention are as follows:

[0049] (1) The target detection method based on multi-frame information disclosed by the present invention mainly includes two parts: a two-stage target detection framework for multi-frame ROI extraction and a network training method based on AIS data matching, which is used for detecting marine ship targets by high-frequency ground wave radar. The improved YOLOv8 network (the first network) improves the single-frame small target detection performance of a single network. At the same time, a multi-frame ViT network (the second network) is proposed to realize the discrimination of multi-frame target image block sequences.

[0050] (2) The present invention only needs to accumulate the RD spectrum detection results of several frames of the first network in a short time, and then the second network can be used to realize multi-frame information extraction and discrimination, effectively improving the final target detection performance.

[0051] (3) Based on three types of data, namely high-frequency ground wave radar frequency domain data, INS data, and AIS data, all real and visible targets are extracted, automatically mapped and matched to realize the generation of an automated dataset.

[0052] (4) The present invention entirely uses measured data for the training and verification of the two-stage network, and the dataset is completely automatically generated by the program. On the one hand, this can avoid problems such as low efficiency and inconsistent standards caused by manually annotating the dataset; on the other hand, this will facilitate the subsequent use of online learning methods to generate datasets in real time to achieve network fine-tuning and domain adaptation in different environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for use in the embodiments or the description of the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0054] Figure 1 It is a schematic diagram of the range-doppler (RD) spectrum image of the radar;

[0055] Figure 2 It is the overall structure diagram of the two-stage target detection framework of the present invention;

[0056] Figure 3 It is the multi-frame image target association logic diagram;

[0057] Figure 4 It is the architecture diagram of the second network;

[0058] Figure 5 It is the dataset generation logic diagram;

[0059] Figure 6a It is the effect diagram of the processed data with bilateral clipping turned off;

[0060] Figure 6b Effect diagram of bilateral clipping (weak clipping) data processing enabled

[0061] Figure 6c Effect diagram of bilateral clipping data processing enabled

[0062] Figure 7a Effect diagram of data processing using all reference unit data

[0063] Figure 7b Effect diagram of data processing using upper half reference unit data

[0064] Figure 8 Three growth rule diagrams of region growing algorithm

[0065] Figure 9 Positive and negative sample classification logic diagram

[0066] Figure 10 Training logic diagram of the first network and the second network in the target detection method provided by the present invention

[0067] Figure 11 Logic diagram of target detection by the first network and the second network in the target detection method provided by the present invention

[0068] Figure 12a Detection result diagram of the 103rd frame RD spectrum image

[0069] Figure 12b Detection result diagram of the 104th frame RD spectrum image

[0070] Figure 12c Detection result diagram of the 105th frame RD spectrum image

[0071] Figure 13 Comparison diagram of detection results of four methods for complete RD spectrum images

[0072] Figure 14a Detection result diagrams of each frame of the target by four methods under the state of the target passing through ground clutter (circle: normally detected target data, cross: missed alarm data, triangle: target data supplemented and detected by the second network)

[0073] Figure 14b Detection result diagrams of each frame of the target by four methods under the state of the target being near ground clutter (circle: normally detected target data, cross: missed alarm data, triangle: target data supplemented and detected by the second network)

[0074] Figure 15 Detection result diagrams of each frame of the target by four methods under the state of unstable signal-to-noise ratio (circle: normally detected target data, cross: missed alarm data, triangle: target data supplemented and detected by the second network). Detailed implementation manners

[0075] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention clearer and more understandable, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0076] The compact high-frequency ground wave radar involved in the present invention is carried on a ship, and all the radar data used comes from a compact HFSWR system named CORMS (compact over-horizon radar for marine surveillance). The general process of radar signal processing is as follows: The high-frequency signal generated by the radar transmitter is transmitted through the antenna and diffracts along the ocean surface; the radar receiver receives the echo signal through its array antenna system and performs mixing processing to obtain the original data; the original data is subjected to two FFTs to respectively achieve range demodulation and velocity demodulation, and finally the frequency-domain data that can be used to draw the RD spectrum is obtained.

[0077] Based on the radar frequency-domain data, a range-doppler (RD) spectrum image of the radar can be generated, as Figure 1 shown. In addition to Figure 1 the five types of data characteristics of sea clutter, ground clutter, ionospheric clutter, radio frequency interference and ship targets introduced, there is also a near-range blind area in the radar RD spectrum image: due to the characteristics of the FMICW modulation wave emitted by the high-frequency ground wave radar, the receiver stops receiving signals during the closing of the receiving gate, so the radar will have a detection blind area in the short-distance area within about 10 km. The reflection signals of any targets in this area will not be received by the radar system, which is manifested as a dark blue strip area with the bottom side parallel to the horizontal axis in the RD spectrum image. Figure 1 The part circled by the black box in

[0078] is the part where a suspected target exists. The radar range-doppler (RD) spectrum images can form multiple frames of images. Due to the changes in the ship's position and the position of the target over time, each frame of the image reflects the detection of target data at each moment. Based on the radar range-doppler (RD) spectrum image and the information of multiple frames of images, the present invention proposes a two-stage target detection framework.

[0079] Refer to Figure 2, which is the two-stage target detection framework proposed by the present invention. It is a schematic diagram of the target detection process based on multi-frame radar images. The "two-stage" method uses two networks to process multi-frame images. In the present invention, the two networks are respectively defined as the Stage-1 network or the first network, and the Stage-2 network or the second network. Among them, the first network is used to detect each frame of the image frame by frame and select the image regions of interest, and the second network receives the data processed by the first network, screens the multi-frame data, and performs target recognition based on the screening results.

[0080] Next, the step process of the compact high-frequency ground wave radar target detection method based on multi-frame information provided by the present invention will be elaborated in detail, including the following steps.

[0081] S1: Based on the radar frequency domain data, generate the radar range-Doppler spectrum image. Construct a range-Doppler spectrum image data set.

[0082] The generation method of the range-Doppler spectrum image is a prior art and will not be elaborated.

[0083] The data sets used in the embodiments of the present invention include the data set for the first network and the data set for the second network. The data set of the first network is obtained based on the range-Doppler spectrum image data set. In some embodiments, it is jointly generated according to external AIS data and the CFAR result based on the radar RD spectrum image; the data set of the second network is selected from the output of the first network (to obtain this output, the data set of the first network just constructed needs to be input to the first network). Details will be described later.

[0084] S2: Input the range-Doppler spectrum image into the first network. The first network performs target detection on each frame of the range-Doppler spectrum image, and takes the region suspected of containing a target as the image region of interest of each frame of the image; centered on the image region of interest, delimit the image blocks of each frame of the image, and form an image block sequence of multi-frame images. During training, the first network learns the input data and trains the network model of the first network.

[0085] The first network is responsible for the candidate extraction of the region of interest (ROI). The first network adopts a target detection network, with the input being consecutive batches of radar RD spectrum images, and the output being the ROI extraction results of each image. The main purpose of the first network is to identify and extract as many ROI candidate regions as possible, provide a data basis for subsequent multi-frame target association and target fine screening verification, and at the same time provide samples for the training of the Stage-2 network. The first network needs to first perform feature extraction and integration on the RD spectrum image, and finally output the confidence level and position coordinate values of each ROI through the classification head and the regression head respectively.

[0086] In the present invention, the first network adopts an improved YOLOv8 + SPD network. The number of parameters of the network is approximately 3M. Based on the original model, in order to solve the problem that the performance of the traditional CNN architecture deteriorates rapidly due to the loss of detailed information when detecting small targets in low-resolution RD spectrum images, it is necessary to improve the original network structure: use space-to-depth convolution (SPD-Conv) to replace the original strided convolution and pooling layers of YOLO, transfer the spatial dimension data of the image height and width to the depth dimension with multiple feature map channels, and then perform a non-strided convolution operation on the new feature map. This method can achieve the downsampling process without losing information and improve the detection performance of the network for small targets in low-resolution images. Through experimental tests, the effect of this improvement is more obvious for small-scale models. Therefore, the YOLOv8n model improved by SPD is selected as the target detection network for Stage-1. Considering the significant differences between radar RD spectrum images and traditional object detection training sets (such as the COCO dataset, etc.), a dataset is generated based on actual radar data, and the network is trained in a zero-training manner.

[0087] It should be understood that in the embodiments of the present invention, the region of interest is the region where the detection target may exist. Similar to other single-frame detection methods, the first network performs target detection on the complete RD spectrum image, outputs all ROI position regions containing suspected targets, and then intercepts square image blocks of a fixed size centered on each ROI position region for subsequent multi-frame target association and the detection process of the second network. The size of the image block can be preset. For convenience of processing, the usually defined image block is a square image block.

[0088] Since radar RD spectrum images are vulnerable to noise interference, in order to improve the accuracy of the data, the present invention also includes the step of correcting the RD spectrum image and the setting of generating a dataset based on AIS data. Among them, the data obtained by matching AIS data and CFAR results is used as the dataset for the first network. Specifically, the dataset generation generally adopts the CFAR + AIS method, that is, the constant false alarm rate (CFAR) detection results based on the RD spectrum image are matched with the real ship data based on AIS data, so as to screen out all real and visible targets. The dataset generation process is mainly divided into three steps: CFAR detection, target mapping, and data matching. The process is referred to Figure 5 .

[0089] In some embodiments of the present invention, the step of constructing a range-Doppler spectrum image dataset, that is, the dataset for the first network, further includes:

[0090] Target extraction is performed on each frame of the range-Doppler spectrum image, and the real target position area of ​​each frame of the image is delineated, including:

[0091] Perform constant false alarm rate detection on the range-Doppler spectrum image to extract all visible targets;

[0092] The constant false alarm rate detection process is corrected by using one of the following methods or a combination of the following methods:

[0093] Reference unit bilateral trimming method: For each frame of range-Doppler spectrum image, constant false alarm rate detection is performed for each minimum resolution unit, and the image units are sorted according to the amplitude values ​​of each image unit in the reference unit of the detector, and the image units with the maximum amplitude value and the minimum amplitude value are removed according to the set ratio. The average amplitude value of each image unit in the reference unit is calculated to obtain the trimmed average value of the reference unit, which is used to calculate the detection threshold of the constant false alarm rate detector;

[0094] The reference unit bilateral clipping method is used to reduce the constant false alarm rate detection results in the background noise area of ​​the range-Doppler spectrum image to avoid interference with the subsequent dataset generation process.

[0095] Upper half reference unit detection method: set a distance threshold, which is used to characterize the distance of the radar detection blind area, regard the distance-Doppler spectrum image area less than the distance threshold as the short-range blind area, and select the area above the short-range blind area as the target area for the upper half reference unit detection method. Calculate the average value of the amplitude value of each image unit in the upper half of the reference unit to calculate the detection threshold of the constant false alarm rate detector.

[0096] The upper half reference unit detection method is used to reduce the influence of the extreme amplitude value of the unit in the close-range blind zone on the average value of the reference unit, so as to avoid the interference of the close-range blind zone on the normal operation of the detector.

[0097] Specifically, CFAR detection is first performed on the radar RD spectrum data to extract all visible targets in the image. At the same time, in order to minimize the detection of targets in the background noise area and correct the erroneous estimation of the background power in the close-range blind area, two improvement measures can be adopted respectively:

[0098] a. Bilateral Trimming CA-CFAR

[0099] In the measured data, the background noise will not perfectly conform to the Gaussian distribution, but may contain some outliers. In order to reduce the detection of CFAR in the background noise, CA-CFAR based on bilateral clipping is proposed, which uses the trimmed mean of the reference unit as the detection threshold. That is, before averaging the reference unit, the data of the reference unit is sorted and a certain proportion of the maximum and minimum values ​​are removed to avoid the wrong estimation of the background power caused by abnormal data.

[0100] Reference Figures 6a to 6c Figure, for the RD spectrum image, is the processing effect diagram of adopting different bilateral clipping processing strategies. In the bilateral clipping CFAR, compared with the original CA-CFAR, the truncated mean is used instead of the ordinary mean as the source of the detection threshold. The truncated mean, also known as the trimmed mean, is a common data processing method, that is: before taking the average of the reference cells, all data are sorted, and a certain proportion of the maximum and minimum values are excluded. The sorting is based on the amplitude of each reference cell, and the sorting direction is not important; the above-mentioned maximum and minimum values, amplitude, and data all refer to the amplitude value of a minimum resolution cell in the RD spectrum, which can be understood as the amplitude value of the minimum unit composed of multiple pixels of the same color in the RD spectrum image.

[0101] b. Upper half reference cell CFAR

[0102] Since there is a certain range of near-range blind area at the bottom side of the RD spectrum image, when performing CFAR detection on the targets near the blind area, if all or part of the lower half of the reference cells fall into the blind area range, it will lead to an underestimation of the background power, resulting in a large number of false alarms being detected. Therefore, a CFAR method based on the upper half reference cells is proposed. This method is only used for the area near the blind area, and the blind area range is determined by a manually set threshold.

[0103] Reference Figure 7a and Figure 7b Figure, are the processing effect diagrams of processing the RD spectrum image with and without using the upper half reference cells. Upper half reference cell CFAR: Take the upper half reference cells → Obtain the detection threshold → Determine whether each cell is detected according to the threshold; Method for determining the blind area: By setting a distance threshold, the area smaller than this threshold is regarded as the near-range blind area.

[0104] The AIS system is a technology for automatic identification, communication, and information exchange between ships and aircraft. It is based on the Global Positioning System (GPS) and communication network technology, and realizes real-time update and sharing of the dynamic information of ships or aircraft through on-board or on-board equipment.

[0105] In some embodiments of the present invention, in step S1, the step of constructing the range-Doppler spectrum image dataset further includes:

[0106] Performing target extraction on each frame of the range-Doppler spectrum image, and demarcating the real target position area of each frame of the image, further including:

[0107] Mapping the ship target information received by AIS to the range-Doppler spectrum image to obtain the AIS target;

[0108] Eliminate the targets in the non-radar detection area and the targets in the clutter interference area from the AIS targets to obtain the real AIS targets;

[0109] Match all visible targets extracted by constant false alarm rate detection with all real AIS targets mapped by AIS information, and regard the targets that can be matched as the final targets;

[0110] Perform correction processing on the final targets. In step S2, input the data after correction processing into the first network.

[0111] In some embodiments of the present invention, the method for eliminating the targets in the non-radar detection area from the AIS targets includes:

[0112] Calculate the angle between the line connecting the AIS target and the radar relative to the normal of the radar antenna array, and eliminate the targets not within the radar detection area;

[0113] According to the ship length of the AIS target, eliminate the targets with incorrect ship length information;

[0114] Estimate the ratio of the radar cross section corresponding to the AIS target to the theoretically maximum radar cross section, and eliminate the targets whose cross section ratio is not within the set range.

[0115] All AIS target information received by the AIS system comes from real targets. Matching the AIS targets with the CFAR detection results can eliminate the invisible AIS targets.

[0116] Before the matching process, steps for correcting the AIS targets are required. For example, calculating the angle between the line connecting the AIS target and the radar relative to the normal of the radar antenna array can eliminate the targets not within the radar detection area; estimating the ratio of the radar cross section corresponding to the AIS target to the theoretically maximum radar cross section can eliminate the targets with very weak echo signals due to various factors (such as the size, attitude, height, material, etc. of the target ship); other information such as the ship length of the AIS target can eliminate the targets with incorrect AIS information.

[0117] In the preferred embodiment, before matching the AIS targets with the CFAR detection results, perform clutter removal processing on the CFAR detection results processed by the reference cell bilateral clipping method and the upper half reference cell detection method, and eliminate the CFAR detection results in the clutter and interference areas, where the clutter and interference include: sea clutter, ground clutter, ionospheric clutter, radio frequency interference, etc.

[0118] The specific method is as follows.

[0119] First, the region growing algorithm is used to cluster the detection results after correction processing to facilitate subsequent clutter recognition and elimination. For the implementation of the region growing algorithm, in addition to the relatively common 4-connectivity (basic method) and 8-connectivity (full method) growth rules, a feasible 12-connectivity (star method) growth rule is proposed, as Figure 8 shown. Through experimental tests, the 12-connectivity (star method) growth rule with better clustering effect is selected to process the CFAR detection results.

[0120] Then, for each target cluster obtained by clustering, the samples whose target cluster center is near the theoretical sea clutter or ground clutter velocity and the number of targets contained is greater than the specified threshold are regarded as sea clutter or ground clutter and eliminated.

[0121] Finally, for other target clusters with a large number of targets, they are processed in two possible cases:

[0122] Traverse all target clusters, calculate their width and height in the RD spectrum, and directly regard the samples with a high aspect ratio and the number of targets greater than the specified threshold as ionospheric clutter and eliminate them. The target clusters have irregular shapes, and the width and height of the target clusters generally refer to the width and height of the minimum bounding rectangle of the target clusters.

[0123] Since there may still be cases where large clutter such as ionospheric clutter is connected to sea clutter / ground clutter, it is necessary to perform morphological closing operations on the target clusters to handle the hollow and rough edges caused by clutter fragmentation. Subsequently, the part within the theoretical sea clutter / ground clutter velocity range in the target clusters is eliminated, and then the aspect ratios of the remaining connected regions are judged in turn. The samples with too large or too small aspect ratios are regarded as ionospheric clutter or radio frequency interference and eliminated respectively.

[0124] After processing the corrected results, for each AIS target, find its adjacent target recognition results, delimit the target box range with each AIS target as the center, and eliminate the highly overlapping target boxes based on IoU (Intersection over Union) to obtain the final Stage-1 dataset.

[0125] According to the research results of relevant scholars, in all seas around the world including the Yellow Sea, a considerable proportion of ships lack AIS signals. At the same time, considering that AIS devices may not always receive the signals of all ships, it is necessary to bypass and introduce CFAR detection results with strong signal-to-noise ratio after the CFAR+AIS matching process to avoid missing potential AIS-free targets and prevent the network training from having a tendency of false alarms.

[0126] For the test dataset, to fully verify the multi-frame target detection performance of the proposed method, it is necessary to manually remove the targets that cannot form tracks and correct the track-breaking targets that are lost in some frames due to low signal-to-noise ratio or partial occlusion by clutter. That is to say, if a target is visible in its previous or subsequent frames but not visible or difficult to distinguish in the current frame, the target will still be added as a sample to be detected in the current frame to ensure the multi-frame continuity of each target. It should be noted that the manually repaired samples will not participate in the training and verification processes of the two-stage network, but are only used for the multi-frame detection performance evaluation of the final detection framework.

[0127] In the present invention, the measured data obtained from experiments carried out in the Yellow Sea in July and December 2021 are selected as the source data, including 3888 batches of frequency-domain data with 8 channels. Each batch of frequency-domain data has AIS data and inertial navigation system (INS) data in the same time period. To improve the efficiency of the dataset generation program and facilitate the management of experimental data, before generating the dataset, it is necessary to synchronize and match the radar frequency-domain data, INS data, and AIS data, and package the three types of matched data into a unified PIA data.

[0128] Each batch of PIA data can generate an RD spectrum image and its target information. The time interval between batches of data is 1 minute. The statistical situation of the radar measured data is shown in Table 1.

[0129] Table 1 Statistical table of radar measured dataset

[0130]

[0131] Among the total 3888 batches of data, 2995 batches are used as the training dataset for the Stage-1 network, 749 batches are used as the validation dataset, and 144 batches are used as the test dataset for the final detection result evaluation. The ratio of data used for training, validation, and testing in the dataset is approximately 20:5:1.

[0132] Through the foregoing processing steps, real ship targets within the radar detection range can be comprehensively screened out.

[0133] Since each frame image contains information on target images at different time points, therefore, on the images of different frames, the positions reflected by the same target point may be different. It should be understood that considering factors such as the radar detection range and the noise influence of individual frame images, at least some target points are correlated on the images of different frames.

[0134] In order to achieve target detection based on multi-frame information, in some embodiments of the present invention, the targets in each frame of range-Doppler spectrum images are associated. The specific method includes: determining the position corresponding to each image region of interest in the current frame, determining the distance of the target corresponding to the center of this region relative to the radar station, and the radial velocity of the target relative to the radar station; setting the maximum motion state of the target, where the maximum motion state includes the maximum turning rate and the maximum acceleration, and based on the target stationary state and the maximum motion state, obtaining the position range of the corresponding target in the next frame adjacent to the current frame; based on the matching between the position of the target in the current frame and the position range of the target in the next frame, associating the targets with adjacent positions, intercepting the image region of interest corresponding to the target as an image block, and obtaining an image block sequence.

[0135] It should be understood that the method for target point association defines two extreme motion states of the ship: namely, the stationary state and the maximum motion state. Based on the two extreme motion states of the ship, the position range of the target in the image of the next time frame can be predicted. According to the principle of proximity of positions, the association of targets on adjacent frame images can be completed, and the association of targets on each frame image can be gradually completed.

[0136] Based on the above method, the present invention utilizes a multi-frame target association method implemented according to the range prediction and nearest neighbor principle. Based on the ROI detection results of each frame output by the first network, within a fixed number of frames, the current frame and several past frames of RD spectrum images are selected, the positions of each suspected target are associated in proximity, and the corresponding multi-frame image block sequence is obtained. Then, according to the detection status of each frame, the image block sequence that meets the specified multi-frame mode is sent to the second network for further fine screening of each target. It should be understood that the length of the image block sequence can be set according to requirements. In the present invention, the sequence length is set to 3.

[0137] Considering that there may be false alarm targets (i.e., in the case of interference, etc., noise information is misidentified as a target) and missed alarm targets (i.e., in the case of interference, etc., the target is not detected) in the RD spectrum image data. Therefore, it is necessary to comprehensively refer to the single-frame target detection results of consecutive multi-frame images to more accurately complete target detection.

[0138] Based on this, in some embodiments of the present invention, the steps for associating the targets in each frame of range-Doppler spectrum images further include:

[0139] Taking three consecutive frames of range-Doppler spectrum images as a group, for each target in the range-Doppler spectrum image of the current frame (the first frame), if:

[0140] The target is detected in the range-Doppler spectrum image of the first frame, and the target is detected in all previous frames of range-Doppler spectrum images, or,

[0141] The target is detected in the first-frame range-Doppler spectrum image, and the target is detected in any previous frame range-Doppler spectrum image, or,

[0142] The target is detected in the first-frame range-Doppler spectrum image, and the target is not detected in any previous frame range-Doppler spectrum image, or,

[0143] The target is not detected in the first-frame range-Doppler spectrum image, and the target is detected in all previous frame range-Doppler spectrum images;

[0144] Then, in the subsequent steps, the image block sequence of the target in the range-Doppler spectrum images of three consecutive frames is input into the second network; otherwise, the image block sequence of the target in the range-Doppler spectrum images is discarded.

[0145] As described above, the present invention sets the length of the image block sequence to 3. Therefore, in the step of associating targets, three consecutive frames of images are used for target association. Referring to Figure 3 , according to the different detection results of each frame, five different multi-frame modes A to E will be formed, and there is also a mode in which none of the three frames of images includes a detected target, and this mode is discarded. Among them, the black-filled rectangle represents the detected samples (including correctly detected samples and false alarm samples), the white-filled rectangle represents the undetected samples (including non-targets and missed alarm samples), and the one with a dotted frame annotation is the detection result of the current frame, and the subsequent ones are the detection results of the previous frame and the frame before the previous frame respectively.

[0146] Specifically:

[0147] Mode A: Targets are detected in all frames;

[0148] Mode B: A target is detected in the current frame, and some targets are detected in the subsequent frames; that is, a target is detected in the current frame, and among the frames other than the current frame, some frames detect the target and some frames do not detect the target;

[0149] Mode C: A target is detected in the current frame, and no target is detected in the subsequent frames; that is, only the current frame detects the target;

[0150] Mode D: No target is detected in the current frame, and targets are detected in all subsequent frames; that is, only the current frame does not detect the target;

[0151] Mode E: No target is detected in the current frame, and some targets are detected in the subsequent frames; that is, no target is detected in the current frame, and among the frames other than the current frame, some frames detect the target and some frames do not detect the target.

[0152] Through experimental tests, only samples that meet patterns A to D are screened out in the present invention, and the corresponding ROI regions of each frame are intercepted as a sequence of image patches of the same size and provided to the second network in Stage-2 for the next judgment. It should be noted that even if no ROI is detected in a certain frame, but it exists in the multi-frame pattern in the form of an undetected sample, then the image patch interception range of this frame will depend on the location of the ROI region in the previous frame.

[0153] S3: Input the sequence of image patches into the second network, flatten the image patches of each frame in the sequence of image patches into vectors, add positional encoding, and form a sequence of image patch tokens. Add a learnable class token to each sequence of image patches, and input the tokens and class tokens of each sequence of image patches into the second network; during training, the second network learns the input data to train the network model of the second network.

[0154] The second network in the second stage (Stage-2) is used for discriminating the object authenticity of multi-frame image patches. For the simultaneously input sequence of target image patches, the multi-frame classification network analyzes the image morphology and dynamic change features of the target between each frame to judge the target and output the final detection result.

[0155] Stage-2 adopts an image classification network. The input is a sequence of multi-frame image patches of consecutive batches in a certain ROI region, and the output is the discrimination result of whether each sequence of image patches contains an object. The main purpose of the network is to further eliminate the false alarm objects misjudged by the Stage-1 network while trying to retain the real objects as much as possible, and retrieve the missed alarm objects in the current frame based on the past frames.

[0156] To enable the Stage-2 network to receive the input of a sequence of multi-frame image patches and be able to extract the object features of each frame of image patches and the association information between them at the same time, the present invention improves the input of multi-frame image patches based on the Vision Transformer (ViT) network and proposes a multi-frame ViT (MFViT) network as the second network. The number of parameters of the network is about 22.8M, and the MFViT network structure is as Figure 4 shown. Benefiting from the lower resolution of the target image patches, similar to the processing method of the Transformer network in the field of natural language processing, each frame of image patches can be directly used as the token input of the Transformer network, giving full play to the advantages of the Transformer network in the sequence modeling mechanism.

[0157] Specifically, each image patch has its own propagation path in the Transformer network, but these propagation paths are not completely independent of each other. First, each image patch will be flattened into a vector and positional encoding information will be added; subsequently, a learnable class token is additionally added before the tokens composed of the image patch sequence, and all tokens are used as the input to the Transformer encoder network. The encoder network is composed of a stack of one or more encoder modules, and each encoder module contains a self-attention module and an MLP module. In the self-attention module, the tokens corresponding to each frame input will calculate attention scores with the corresponding tokens of other frames, enabling the network to obtain the ability to analyze the temporal correlation and dynamic change relationship between the tokens in the image patch sequence. Compared with the original ViT network, since there is no longer a need for the embedding process of image patch segmentation for a single image and there is no spatial relationship between the image patches of each frame at the same moment, MFViT can avoid the limitations of inductive biases such as locality and translational invariance of the original ViT network, and theoretically can collect more comprehensive feature information than traditional CNN networks.

[0158] Considering that all the input data of the second network in Stage-2 comes from the object detection results of the first network in Stage-1, a dataset is constructed by finding positive and negative samples based on the output results of Stage-1, and the network is trained in a zero-training manner.

[0159] In step S2, in step S1, using the dataset already generated for the first network, inputting the first network that has been trained based on this dataset, and then finding real objects and false objects from the output results of the first network to generate the dataset for the second network includes: dividing the image patch sequence output by the first network into positive samples and negative samples; the positive sample data includes the objects correctly detected in the current frame and the objects that have correct detection results in any previous frame except the current frame and have no false detection results; the negative sample data includes the objects that are falsely detected in the current frame and have no correct detection results in any previous frame; combining the positive samples and negative samples to obtain the training and validation dataset for the second network.

[0160] Reference Figure 9 , the generation of the dataset for the Stage-2 network is mainly divided into two parts: positive sample generation and negative sample generation. In the present invention, both positive and negative samples come from the samples in the Stage-1 detection results that meet the specified multi-frame pattern. If the relative order of the detection results of other frames except the current frame is not considered, all possible multi-frame patterns are as Figure 9 shown. Among them, the green filled rectangles represent the correctly detected samples of Stage-1, and the red filled rectangles represent the false alarm samples of Stage-1.

[0161] For positive samples, on the basis of retaining the modes correctly detected in the current frame of Stage-1, modes that are not detected in the current frame but are all or partially correctly detected in the remaining frames, and modes that are incorrectly detected in the current frame but are all correctly detected in the remaining frames should be additionally added to supplement the potential undetected targets in the current frame. Therefore, modes A (A-1 to 3), modes B (B-1 to 2), mode C, mode D-1, mode E-1, and mode F3 are selected as the positive samples for the training of the Stage-2 network. For negative samples, it should be ensured that the image block sequences input to Stage-2 do not contain correctly detected samples to avoid the tendency of the Stage-2 network to generate additional undetected targets. Therefore, modes F1, F4, and F6 are selected as the negative samples for the training of the Stage-2 network. 3744 batches of RD spectrum images except for the test dataset are input to the Stage-1 network. When the YOLO threshold is set at 8e-4, 269753 groups of image block sequences can be obtained, among which 99960 groups are positive samples and 169793 groups are negative samples. The ratio of the dataset used for training and validation is approximately 8:2.

[0162] S4: Use the trained first network model and second network model for target detection.

[0163] Next, qualitative analysis and quantitative analysis methods are used to illustrate the target detection effect of the target detection method provided by the present invention.

[0164] The complete training / validation pipeline of the method provided by the present invention is as Figure 10 shown. First, the measured data of the high-frequency ground wave radar is input to the dataset generation program, and the YOLO dataset is generated by combining the CFAR detection results and AIS target information to complete the training of the Stage-1 network; subsequently, the YOLO dataset is input to the Stage-1 network, and the image block sequences that meet the specified modes are collected as the positive and negative samples of the image block dataset to complete the training of the Stage-2 network.

[0165] The complete inference / test pipeline of the method provided by the present invention is as Figure 11 shown. The test dataset is input to the input end of the Stage-1 network. After 3 steps including ROI candidate extraction in Stage-1, target association and mode screening, and multi-frame target fine screening in Stage-2, the intermediate detection results of Stage-1 for each frame and the final detection results output by Stage-2 can be obtained. To minimize the undetected rate of Stage-1 and provide operating space for the subsequent Stage-2 network, a lower YOLO threshold setting is required.

[0166] (I) Qualitative analysis

[0167] The test data set consists of 142 frames of RD spectrum images. The final test results of 3 consecutive frames from frame 103 to frame 105 are as follows. Figures 12a to 12c As shown, the green squares represent the normally detected targets, the orange boxes represent the false alarm targets, and the blue boxes represent the false alarm targets eliminated in Stage-2.

[0168] From Figures 12a to 12c it can be seen that the method proposed by the present invention can reduce the target missing alarm while effectively filtering out the false alarm targets generated in each frame according to the target image characteristics of the current frame and the previous frame.

[0169] Using the same RD spectrum image data, the method provided by the present invention (MFTDF) is used to perform target recognition respectively with the CA-CFAR method, the OES-ELM method and the S3D method. The recognition comparison results are as follows. Figure 13 As shown.

[0170] Among them:

[0171] ① CA-CFAR: It is a cell averaging constant false alarm rate (CFAR) detection method with a fixed false alarm rate threshold and is one of the most commonly used radar target detection methods.

[0172] ② OES-ELM: It is a two-stage cascaded target detection method based on the optimal error self-correcting extreme learning machine, mainly composed of a linear classifier and a self-correcting ELM network. In the initialization stage, the self-correcting ELM network uses the L1 / 2 regularization operator to ensure the sparsity of the network and cuts off the useless hidden layer neurons.

[0173] ③ S3D: It is a cascaded target detection method using semi-supervised self-distillation learning, mainly composed of an improved peak CFAR and a self-distillation ResNet network (CNN network). In the training process of the ResNet network, both labeled samples and unlabeled samples are used, and the pseudo-labels generated for the unlabeled samples are only retained when the output confidence is high.

[0174] The final detection results of the above methods will be subjected to clustering processing, and whether there is a part of the target clusters obtained covering the area where the real target is located will be used as the discrimination condition for whether the target is correctly detected. In addition, the detection results within sea clutter and ground clutter are excluded from the detection results of all the above methods, and the detection results within large clutter such as ionospheric clutter are additionally excluded from the CFAR method and the ELM method. The comparison of the detection results for the complete RD spectrum images is as follows. Figure 13 As shown, the green boxes represent the normally detected targets, the red boxes represent the missing alarm targets, and the orange boxes represent the false alarm targets.

[0175] From Figure 13It can be seen that the MFTDF method proposed by the present invention can overall generate fewer missed alarm and false alarm targets.

[0176] To further evaluate the detection effect of the MFTDF method proposed by the present invention on targets with different dynamic characteristics, the moving trajectories and the detection results of each frame are analyzed respectively under three scenarios of target characteristic changes: the target crossing ground clutter, the target being near ground clutter, and the target signal-to-noise ratio being unstable. The detection result diagrams of each frame of the target of the comparison method are shown in Figure 14 and Figure 15 as shown. Among them, the image background of each sample is drawn by superimposing the RD spectrum images of the frames where the target is located. It should be noted that the target trajectories and each point marker shown in the figure are not the tracks formed by network tracking, but only represent the independent detection results of each frame of the target.

[0177] From Figure 14 and Figure 15 it can be seen that the MFTDF method proposed by the present invention can achieve stable detection of the target under different scenarios of target characteristic changes. At the same time, in the case where the signal-to-noise ratio of some targets is extremely weak, the Stage-2 network of MFTDF can also supplement the detection of the current frame according to the previous frames, further reducing the situation of target loss.

[0178] (2) Quantitative analysis

[0179] Quantitative analysis is carried out on the final detection results. The results of Stage-1 and the final detection results are compared with three types of previous methods. The performance indicators and inference times of each method are shown in Table 1. Among them, the detection range of 0 - 300 km means detecting the entire RD spectrum image, the detection range of 0 - 150 km means detecting the lower half RD spectrum image where all targets are located, and the detection range of 0 - 100 km means detecting the lower one-third RD spectrum image where there is no ionospheric clutter.

[0180] Table 2 Performance indicators and inference times of each method

[0181]

[0182]

[0183] It can be seen from the above table that the method proposed by the present invention can utilize multi-frame information to significantly improve the detection performance, and the final detection results obtain the best precision and recall rates at different detection distances.

[0184] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principle of the present invention by those of ordinary skill in the art shall be included within the protection scope of the present invention. Therefore, the protection scope of this patent application shall be subject to the protection scope of the appended claims.

Claims

1. A compact high-frequency ground wave radar target detection method based on multi-frame information, characterized in that: The following steps are involved: S1: Generate radar range-Doppler spectrum image based on radar frequency domain data and construct range-Doppler spectrum image dataset; S2: inputting the range-Doppler spectrum image into the first network, the first network performs target detection on each frame of the range-Doppler spectrum image, and uses the area suspected to contain the target as the image area of ​​interest of each frame of the image; Taking the image region of interest as the center, demarcating image blocks of each frame of the image, and composing image blocks of multiple frames of the image into an image block sequence; during training, the first network learns input data and trains the network model of the first network; S3: Input the image block sequence into the second network, flatten the image blocks of each frame image in the image block sequence into vectors, add position codes to form image block sequence tokens, add a learnable class token to each image block sequence, and input the token and class token of each image block sequence into the second network; during training, the second network learns the input data and trains the network model of the second network; S4: Use the trained first network model and the second network model to perform target detection.

2. The compact high-frequency ground wave radar target detection method based on multi-frame information as claimed in claim 1, characterized in that: The step S2 also includes: associating the targets in each frame of the range-Doppler spectrum image; Determine the position corresponding to each of the image regions of interest in the current frame, determine the distance of the target corresponding to the center of each of the image regions of interest relative to the radar station, and the radial speed of the target relative to the radar station; set the maximum motion state of the target, the maximum motion state includes the maximum turning rate and the maximum acceleration, and obtain the position range of the corresponding target in the next frame adjacent to the current frame based on the stationary state and the maximum motion state of the target; match the position range of the target in the current frame with the position range of the target in the next frame, associate the targets at adjacent positions, intercept the image region of interest corresponding to the target as an image block, and obtain an image block sequence; If the corresponding target cannot be found in the next frame adjacent to the current frame, the position of the target in the current frame is retained as the imaginary position of the target in the next frame, which is used for target association between the next frame and the next next frame; In step S3, the image block sequence after the associated target processing is input into the second network.

3. The compact high-frequency ground wave radar target detection method based on multi-frame information as claimed in claim 2, characterized in that: The step of associating the targets in each frame of the range-Doppler spectrum image also includes: Take three consecutive frames of range-Doppler spectrum images as a group. For each target in the first frame of range-Doppler spectrum image, if: The first frame of range-Doppler spectrum image detects the target, and all the previous frames of range-Doppler spectrum images detect the target, or, The first frame of the range-Doppler spectrum image detects the target, and any of the previous frames of the range-Doppler spectrum image detects the target, or, The first frame of range-Doppler spectrum image detects the target, and all previous frames of range-Doppler spectrum image do not detect the target, or, The first frame of range-Doppler spectrum image does not detect the target, but all the previous frames of range-Doppler spectrum images detect the target; Then in step S3, the image block sequence of the target in three consecutive frames of range-Doppler spectrum images is input into the second network, otherwise, the image block sequence of the target in the range-Doppler spectrum image is abandoned.

4. The compact high-frequency ground wave radar target detection method based on multi-frame information as claimed in claim 1, characterized in that: In step S1, the step of constructing a range-Doppler spectrum image data set further includes: Target extraction is performed on each frame of the range-Doppler spectrum image, and the real target position area of ​​each frame of the image is delineated, including: Perform constant false alarm rate detection on each frame of range-Doppler spectrum image to extract all visible targets; The constant false alarm rate detection process is corrected by using one of the following methods or a combination of the following methods: Reference unit bilateral trimming method: For each frame of range-Doppler spectrum image, constant false alarm rate detection is performed for each minimum resolution unit, and the image units are sorted according to the amplitude values ​​of each image unit in the reference unit of the detector, and the image units with the maximum amplitude value and the minimum amplitude value are removed according to the set ratio. The average amplitude value of each image unit in the reference unit is calculated to obtain the trimmed average value of the reference unit, which is used to calculate the detection threshold of the constant false alarm rate detector; Upper reference unit detection method: set a distance threshold, which is used to characterize the distance of the radar detection blind spot. The distance-Doppler spectrum image area less than the distance threshold is regarded as a short-range blind spot. The area above the short-range blind spot is selected as the target area for using the upper reference unit detection method. The average value of the amplitude value of each image unit in the upper half of the reference unit is calculated, which is used to calculate the detection threshold of the constant false alarm rate detector.

5. The compact high-frequency ground wave radar target detection method based on multi-frame information as claimed in claim 1, characterized in that: In step S1, the step of constructing a range-Doppler spectrum image data set further includes: Extracting a target from each frame of the range-Doppler spectrum image and defining a real target position area for each frame of the image also includes: Map the ship target information received by AIS to the range-Doppler spectrum image to obtain the AIS target; Eliminate the targets in the non-radar detection area and the clutter interference area from the AIS targets to obtain the real AIS targets; Match all visible targets extracted by constant false alarm rate detection with all real AIS targets mapped by AIS information, and take the matching targets as the final targets; Modify the final goal; In step S2, the corrected data is input into the first network.

6. The compact high-frequency ground wave radar target detection method based on multi-frame information as claimed in claim 5, characterized in that: Methods for eliminating targets outside the radar detection area from AIS targets include: Calculate the angle between the line connecting the AIS target and the radar and the normal line of the radar antenna array to eliminate targets that are not within the radar detection area; According to the ship length of the AIS target, targets with incorrect length information are eliminated; Estimate the ratio of the radar cross section of the AIS target to the theoretical maximum radar cross section, and eliminate targets whose radar cross section ratio is not within the set range.

7. The compact high-frequency ground wave radar target detection method based on multi-frame information as claimed in claim 5, characterized in that: Methods for eliminating targets in the clutter interference area from AIS targets include: The theoretical positions of sea clutter and ground clutter in the range-Doppler spectrum image are calculated, and the constant false alarm rate detection results are clustered using the region growing algorithm. The positions of large clutter such as ionospheric clutter and radio frequency interference are searched from the clustering results, and the targets located in various clutter and interference areas are eliminated.

8. The compact high-frequency ground wave radar target detection method based on multi-frame information as claimed in claim 1, characterized in that: In step S2, the range-Doppler spectrum image data set is input into the trained first network, the real target and the false target are found from the output result of the first network, and the image block is generated and input into the second network, including: Divide the image block sequence output by the first network into positive samples and negative samples; The positive sample data includes targets correctly detected in the current frame, targets not detected in the current frame but correctly detected in any previous frame, and targets incorrectly detected in the current frame but correctly detected in all previous frames; The negative sample data includes targets for which the current frame is detected incorrectly and any of the previous frames do not have a correct detection result; In step S3, the positive samples and negative samples are combined and input into the second network.

9. The compact high-frequency ground wave radar target detection method based on multi-frame information as claimed in claim 1, 4, 5 or 8, characterized in that: The first network is a YOLOv8 network.

10. The compact high-frequency ground wave radar target detection method based on multi-frame information according to claim 1, 2, 3 or 8, characterized in that: The second network is a Vision Transformer network.

Images