Marine radar polar environment target stable tracking system based on intelligent bow estimation and operation method thereof

Abstract

The invention discloses a marine radar polar environment target stable tracking system based on intelligent bow estimation and an operation method thereof, which adopt a mode of tracking while scanning and comprise a trace point preprocessing module, a data association module, a track maintenance module, a track initiation module and a track management module, and carrying out trace point preprocessing, track initiation, track combination, track deletion, data association and tracking maintenance processing on the target trace points in the work cycle. During operation, firstly, an echo image is obtained from a processed radar echo signal, the processed image is input into the SuperPoint network for feature extraction, key points and descriptors of the image are obtained, and finally the corrected final rotation angle is obtained. According to the method, the deep neural network is trained, so that the deep neural network can learn the angular point information of the target in the radar echo image and has a certain degree of target fault tolerance between the images, and the target corresponding relation between the adjacent radar frame images is extracted more robustly.

Description

Technical Field

[0001] The invention belongs to the technical field of marine radar target detection and tracking, and in particular relates to a marine radar polar environment target stable tracking system based on intelligent bow estimation and an operation method thereof. Background Art

[0002] Polar regions are hubs connecting the world's continents and have important sea transportation channels. As the main equipment for ships to perceive the outside world, marine radars provide ships with the ability to detect ice and dangerous targets in polar regions all-weather, and can provide safe navigation for ships when sailing in icy waters. Target positioning and tracking of marine radars require the input of navigation information, but when working in polar regions, some navigation equipment is restricted in use due to the unique environment of polar regions and cannot work properly. Among them, the most important compass information for stable target tracking of marine radars cannot be accurately obtained due to the obvious drift of the earth's magnetic poles in polar regions. In extreme cases, magnetic compass equipment may not even work properly. Once the compass fails to work properly, multiple equipment alarms will appear on the bridge. For example: automatic rudder failure, radar display compass signal loss, etc., there are great safety hazards in the process of ship collision avoidance. Therefore, for marine radars working in polar regions, the problem of stable target tracking in the case of compass loss is the primary problem that needs to be solved.

[0003] Current image matching algorithms such as SIFT, SURF, and ORB can achieve feature matching to a certain extent. However, maritime navigation is complex and is often disturbed by extreme weather conditions such as waves, fog, rain, and snow. These factors cause unstable changes in targets in radar echo maps, and targets may disappear or change in shape, making it difficult for traditional algorithms to adapt to various complex situations. In addition, these algorithms require manual parameter setting and are often unable to adapt to all situations, so the robustness of the algorithms is poor. Summary of the invention

[0004] The invention proposes a polar environment target stable tracking system for a marine radar based on intelligent bow estimation and an operation method thereof, the purpose of which is to estimate the heading of a ship in real time by using artificial intelligence technology when a compass fails, and to apply the estimated heading information to the marine radar target tracking system to ensure stable tracking of polar targets, thereby assisting navigation safety and solving the problem of stable target tracking of marine radars in polar regions.

[0005] The present invention is achieved through the following technical solutions:

[0006] The stable tracking system of polar environment targets of marine radar based on intelligent bow estimation adopts a scanning and tracking mode, including a point track preprocessing module, a data association module, a track maintenance module, a track starting module, and a track management module. The point track preprocessing module preprocesses the input data to obtain the point track information required for tracking; the measurement data generated by the point track preprocessing module enters the data association module, and is associated with the relevant wave gate formed after the position prediction of each track, and the associated target measurement data enters the track maintenance module to filter the target track; the remaining measurement data of the target without correlation enters the track starting module; the temporary track generated by the track starting module and the track updated after filtering enter the track management module together; the track management module sorts the targets, deletes the tracks that have not been updated for a specified time, merges the tracks confirmed to be generated from the same target, and outputs the updated track.

[0007] Furthermore, the data source of the target tracking system is point track information, in which the distance information is based on the center of the ship, and the azimuth information is based on the bow information of the ship. During the tracking process, since the direction of the bow is constantly changing, the data processing uses the azimuth information based on the true north, that is, the bow information of the ship (i.e., compass) and the azimuth based on the bow are used to obtain the azimuth information of the target point track based on the true north. When an alarm of the loss of the bow information is received, the compass value obtained by the bow estimation based on the deep neural network is used, and the filter parameters are adaptively adjusted so that it can continue to maintain stable tracking.

[0008] The operation method of the polar environment target stable tracking system of the marine radar based on intelligent bow estimation is to use a new ship heading value when the compass fails. The specific operation steps are as follows:

[0009] (1) Based on the newly acquired estimated heading value and time information, the heading value of the point track at the current moment is obtained by the difference principle, and the angle information of the point track relative to the heading of the ship is added to obtain the latest angle information based on true north;

[0010] (2) Adjust the azimuth noise estimate and disturbance covariance value of the Kalman filter to make it compatible with the change of angular noise;

[0011] (3) Adjust the filter parameters and other related logic of the track automatic start module to enhance its compatibility with angle noise;

[0012] (4) The ship heading value corresponding to the point track is output together with the track for display.

[0013] Furthermore, in step (1), a real-time ship heading estimation method is used to apply the SuperPoint algorithm to the application scenario of ship heading estimation. The specific processing steps are as follows:

[0014] (a) Processing the radar echo data to remove interference, noise and part of the sea clutter, and plotting the video lines into images in a Cartesian coordinate system to obtain a radar echo map;

[0015] (b) accumulating at least two radar scanning cycles to obtain radar echo images of adjacent frames;

[0016] (c) The target domain is cropped on two adjacent radar image frames to remove the pixels outside the inscribed circle in the image;

[0017] (d) Input the images into the trained SuperPoint network to obtain key points and corresponding descriptors;

[0018] (e) Calculate the mapping matrix corresponding to the key points and descriptors through the RANSAC algorithm;

[0019] (f) using the inverse tangent function to calculate the image rotation angle corresponding to the rotation part in the mapping matrix;

[0020] (g) The obtained angle is passed through an α-β filter and smoothed according to historical data to obtain a calibrated angle value.

[0021] Furthermore, in the specific operation of step (a), the radar saves an echo map for each scan, and obtains two adjacent images with a resolution of 1024 pixels × 1024 pixels. Since the image's rotation angle information is calculated, in the relative rotation change, the part outside the inscribed circle of the image may be lost in the adjacent frames. The processed image is input into the SuperPoint network for feature extraction to obtain the key points and descriptors of the image. The BF matching algorithm in the OpenCV tool library is used for feature matching, and the first 1000 (if the number of pairs is greater than 1000) groups of matched feature points are taken to calculate the affine transformation matrix.

[0022] Furthermore, in the specific operation of step (b), radar images of two adjacent frames are obtained, and a mask area is made with the center of the image as the center of the circle and half of the image width as the radius. Next, this image is input into the SuperPoint network, which uses a convolutional layer to extract image features, and finally obtains feature points and descriptors in the input image. Then, the BF matcher in OpenCV is used for feature matching to obtain a 3×3 two-dimensional affine transformation matrix to obtain an estimated angle. During the radar scanning process, there may be waves and other targets on the sea, which will cause noise on the echo map, so the calculated angle has a certain deviation. However, during the turning process of the ship, the angular velocity changes continuously, and the angle change law is also gentle. The α-β filter is used to smooth the turning angle of the bow to reduce error interference.

[0023] Furthermore, in the specific operation of step (c), a mask area is drawn with the image center position (512, 512) as the center and a radius of 512, the pixels in the mask area are retained, and the pixels in the remaining parts are removed, thereby achieving the effect of noise reduction and removal of redundant targets, retaining the pixel values ​​of the mask area, deleting the pixel values ​​of other areas, and obtaining a cropped and scaled image.

[0024] Furthermore, in the specific operation of step (d), the radar image feature extraction algorithm based on SuperPoint is composed of a convolutional neural network. The feature extraction part is composed of a traditional VGG network. Since the present invention focuses on operation efficiency and the radar image is not sensitive to changes in the color and texture of the target, the number of channels in the network is set to half of the original VGG network, which can not only operate efficiently, but also meet the feature extraction capabilities of the radar image.

[0025] The input shape of a single image is (N, C in , H in , W in ), the output shape after the convolution layer is (N, C out , H out , W out ), then the calculation formula of the convolutional layer is as follows:

[0026]

[0027] in," * " represents the cross product operation of the weight matrix W and the input layer, N represents the batch size, C represents the number of channels, and H represents the in and H out Represents the height of the input and output feature maps, W in and W out Indicates the width of the input and output feature maps.

[0028] During the model training process, the size of the input image is set to 512 pixels × 512 pixels to save cache. However, in the actual application process, the image size is adjusted to 1024 pixels × 1024 pixels to better retain the detail information of the radar echo image. The image is subjected to feature extraction and dimensionality reduction through a shared coding network, so that the width and height of the feature map are reduced to 1 / 8 of the original. Subsequently, the feature map is processed by two decoders, namely the feature point detection decoder and the descriptor decoder. The output of the feature point detection decoder represents the probability that each pixel in the image is a key point, while the descriptor decoder outputs the semi-dense descriptor corresponding to the key point. This output is bilinearly interpolated to obtain a complete descriptor, and finally, a descriptor of unit length is obtained through L2 normalization.

[0029] This algorithm is suitable for detecting corner information of objects in images. Compared with traditional algorithms such as SIFT, ORB, and SURF, it is more robust and has a wider range of adaptability. Its performance depends on the data set and the parameter settings during the model training process.

[0030] Furthermore, the steps (e) and (f) extract the rotation part of the matrix, calculate the rotation angle by the inverse tangent function, use the BF matching algorithm to perform key point matching, and then use the RANSAC method to calculate the optimal single mapping transformation matrix M between multiple two-dimensional point pairs, as shown below:

[0031]

[0032] The 2×2 submatrix in the upper left corner is the rotation part. The inverse tangent trigonometric function is used to calculate the rotation angle between the two images. The obtained rotation angle is input into the α-β filter for correction to obtain the corrected final rotation angle.

[0033] Furthermore, in the specific operation of step (g), an α-β filter is used for state estimation and data smoothing. Assume that the residual between the measured value and the predicted value is The α-β filter selects the α and β parameters and uses the residual r k α times to correct the position estimate, using the residual r k β / T times is used to correct the speed estimate. Where β / T is the normalized β. To ensure system stability, the values ​​of α and β must be positive and small, and usually need to satisfy the following formula:

[0034] 0<α<1

[0035] 0<β≤2

[0036] 0<4-2α-β

[0037] This method requires experiments to adjust the values ​​of α and β, and has good stability.

[0038] The technical principle of the present invention is: the artificial intelligence methods such as adjacent frame radar echo information accumulation, target domain cropping, feature extraction and matching are adopted to realize the engineering method of real-time ship bow estimation, and it is combined with the navigation radar tracking system to solve the problem of stable target tracking of the navigation radar when the compass is lost. A ship bow estimation method based on the SuperPoint algorithm is proposed for real-time ship bow estimation during the navigation of Arctic ships. This method uses adjacent frame radar echoes for information accumulation, removes noise information around the image through the target domain cropping method, uses the SuperPoint algorithm to extract targets from radar images, matches the extracted targets through the BF matching algorithm, and finally calculates the rotation angle between two adjacent frames of images, corrects the rotation angle through an α-β filter, and finally obtains real-time rotation angle information.

[0039] Beneficial effects: The present invention trains a deep neural network so that it can learn the corner point information of the target in the radar echo image and has a certain degree of target tolerance between images, so as to more robustly extract the target correspondence between adjacent radar frame images. The deep neural network can automatically learn and adjust the parameters of feature extraction and matching, adapt to different environments and scene changes, and reduce the complexity of manual intervention and parameter setting. The method can effectively deal with the image matching problem in complex marine environments, improve the accuracy and robustness of matching, and obtain a better bow estimation. Combined with the estimation of the bow direction, the tracking system of the navigation radar is improved, which can effectively solve the problem of stable target tracking of the polar navigation radar, and help improve navigation safety and navigation efficiency. For the estimation of the bow direction, the navigation radar can continuously obtain radar echoes during the scanning process, and based on the subtle change information of the target position in the adjacent radar antenna frame data, the transformation matrix (including displacement and rotation angle) between the two frames of images is obtained through feature matching, and then the rotation angle information between the two frames of images can be calculated by the inverse tangent function, so as to estimate the bow direction of the ship. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 The figure is a flowchart of a ship heading estimation algorithm in one embodiment of the present invention.

[0041] Figure 2 The SuperPoint network structure diagram of an embodiment of the present invention.

[0042] Figure 3 This is a comparison diagram of radar echoes of two original adjacent frames in one embodiment of the present invention.

[0043] Figure 4 This is a feature matching effect diagram of two adjacent frames after applying the algorithm in one embodiment of the present invention.

[0044] Figure 5 A target tracking system incorporating real-time ship heading estimation information in one embodiment of the present invention. DETAILED DESCRIPTION

[0045] The following is a detailed description of an embodiment of the present invention in conjunction with the accompanying drawings: This embodiment is implemented on the premise of the technical solution of the present invention, and a detailed implementation method and a specific operation process are given, but the protection scope of the present invention is not limited to the following embodiments.

[0046] The stable tracking system of polar environment targets of marine radar based on intelligent bow estimation adopts a scanning and tracking mode, including a point track preprocessing module, a data association module, a track maintenance module, a track starting module, and a track management module. The point track preprocessing module preprocesses the input data to obtain the point track information required for tracking; the measurement data generated by the point track preprocessing module enters the data association module, and is associated with the relevant wave gate formed after the position prediction of each track, and the associated target measurement data enters the track maintenance module to filter the target track; the remaining measurement data of the target without correlation enters the track starting module; the temporary track generated by the track starting module and the track updated after filtering enter the track management module together; the track management module sorts the targets, deletes the tracks that have not been updated for a specified time, merges the tracks confirmed to be generated from the same target, and outputs the updated track. The data source of the target tracking system is point information. In the point information, the distance information is based on the center of the ship, and the azimuth information is based on the bow information of the ship. During the tracking process, since the direction of the bow is constantly changing, the data processing uses the azimuth information based on the true north, that is, the bow information of the ship (i.e., compass) and the azimuth based on the bow are used to obtain the azimuth information of the target point based on the true north. When an alarm is received that the bow information is lost, the compass value obtained by the deep neural network-based bow estimation is used, and the filter parameters are adaptively adjusted to enable it to continue to maintain stable tracking.

[0047] The operation method of the polar environment target stable tracking system of the marine radar based on intelligent bow estimation is to use a new ship heading value when the compass fails. The specific operation steps are as follows:

[0048] (1) Based on the newly acquired estimated heading value and time information, the heading value of the point track at the current moment is obtained by the difference principle, and the angle information of the point track relative to the heading of the ship is added to obtain the latest angle information based on true north;

[0049] (2) Adjust the azimuth noise estimate and disturbance covariance value of the Kalman filter to make it compatible with the change of angular noise;

[0050] (3) Adjust the filter parameters and other related logic of the track automatic start module to enhance its compatibility with angle noise;

[0051] (4) The ship heading value corresponding to the point track is output together with the track for display.

[0052] The real-time ship heading estimation based on radar echo includes the following processes:

[0053] 1. Establish a target tracking system that combines real-time ship heading estimation information:

[0054] The target tracking system of the marine radar adopts the scanning and tracking method to perform point track preprocessing, track initiation, track merging, track deletion, data association and tracking maintenance on the target points within the working cycle.

[0055] The point track preprocessing module preprocesses the input data to obtain the point track information required for tracking; the measurement data generated by the preprocessing module enters the data association module, and is associated with the relevant gates formed after the position prediction of each track. The associated target measurement data enters the track maintenance module to filter the target track; the remaining measurement data of the target without correlation enters the track starting module; the temporary track generated by the track starting module and the track updated after filtering enter the track management module together; the track management module sorts the targets, deletes the tracks that have not been updated for a specified time, merges the tracks confirmed to be generated from the same target, and outputs the updated track.

[0056] The main data source of the target tracking system is the point information. In the point information, the distance information is based on the center of the ship, and the azimuth information is based on the bow information of the ship. During the tracking process, since the direction of the bow is constantly changing, the data processing uses the azimuth information based on the true north, that is, the bow information of the ship (i.e., compass) and the azimuth based on the bow are used to obtain the azimuth information of the target point based on the true north. When an alarm is received that the bow information is lost, the compass value obtained by the deep neural network-based bow estimation is used, and the filter parameters are adaptively adjusted to enable it to continue to maintain stable tracking.

[0057] 2. Real-time ship heading estimation based on radar echo:

[0058] For targets in sea clutter, firstly, an echo image is obtained from the processed radar echo signal. Specifically, the radar saves an echo image for each scan, and obtains two adjacent images with a resolution of 1024 pixels × 1024 pixels. Since the image's rotation angle information is calculated, in the relative rotation change, the part outside the inscribed circle of the image may have the phenomenon of target loss in adjacent frames. Therefore, a mask area is drawn with the center position (512,512) of the image as the center and a radius of 512. The pixels in the mask area are retained and the pixels in the rest are removed, thereby achieving the effect of noise reduction and removal of redundant targets. The processed image is input into the SuperPoint network for feature extraction to obtain the key points and descriptors of the image. The BF matching algorithm in the OpenCV tool library is used for feature matching. The first 1000 (if the number of pairs is greater than 1000) groups of matched feature points are taken to calculate the affine transformation matrix. Subsequently, the rotation part of the matrix is ​​extracted, and the rotation angle is calculated by the inverse tangent function. The obtained rotation angle is input into the α-β filter for correction to obtain the corrected final rotation angle. As shown Figure 2 shown.

[0059] 3. Ship heading estimation algorithm:

[0060] The flowchart of the ship heading estimation algorithm is as follows: Figure 1-5 As shown. First, obtain two radar images of adjacent frames, and make a mask area with the center of the image as the center and half the image width as the radius. Keep the pixel values ​​of the mask area, delete the pixel values ​​of other areas, and get the cropped and scaled image. Next, input this image into the SuperPoint network, which uses a convolutional layer to extract image features, and finally obtains the feature points and descriptors in the input image. Then, the BF matcher in OpenCV is used for feature matching to obtain a 3×3 two-dimensional affine transformation matrix to obtain the estimated angle. During the radar scanning process, there may be waves and other targets on the sea, which will cause noise on the echo map, so the calculated angle has a certain deviation. However, during the turning process of the ship, the angular velocity changes continuously, and the angle change law is also gentle. Therefore, the α-β filter is used to smooth the turning angle of the bow to reduce error interference.

[0061] 4. Feature extraction algorithm based on SuperPoint:

[0062] The radar image feature extraction algorithm based on SuperPoint is composed of convolutional neural networks. The feature extraction part is composed of the traditional VGG network. Since the operation efficiency needs to be focused and the radar image is not sensitive to changes in the color and texture of the target, the number of channels in the network is set to half of the original VGG network, which can run efficiently and also meet the feature extraction capabilities of radar images.

[0063] The input shape of a single image is (N, C in , H in , W in ), the output shape after the convolution layer is (N, C out , H out , W out ), then the calculation formula of the convolutional layer is as follows:

[0064]

[0065] in," * " represents the cross product operation of the weight matrix W and the input layer, N represents the batch size, C represents the number of channels, and H represents the in and H out Represents the height of the input and output feature maps, W in and W out Indicates the width of the input and output feature maps.

[0066] During the model training process, the size of the input image is set to 512 pixels × 512 pixels to save cache. However, in the actual application process, the image size is adjusted to 1024 pixels × 1024 pixels to better retain the detail information of the radar echo image. The image is subjected to feature extraction and dimensionality reduction through a shared coding network, so that the width and height of the feature map are reduced to 1 / 8 of the original. Subsequently, the feature map is processed by two decoders, namely the feature point detection decoder and the descriptor decoder. The output of the feature point detection decoder represents the probability that each pixel in the image is a key point, while the descriptor decoder outputs the semi-dense descriptor corresponding to the key point. This output is bilinearly interpolated to obtain a complete descriptor. Finally, a descriptor of unit length is obtained through L2 normalization.

[0067] This algorithm is suitable for detecting corner information of objects in images. Compared with traditional algorithms such as SIFT, ORB, and SURF, it is more robust and has a wider range of adaptability. Its performance depends on the data set and the parameter settings during the model training process.

[0068] 5. Feature matching and angle calculation:

[0069] The BF matching algorithm is used for key point matching. Subsequently, the RANSAC method is used to calculate the optimal single mapping transformation matrix M between multiple two-dimensional point pairs, as shown below:

[0070]

[0071] The 2×2 submatrix in the upper left corner is the rotation part, and the inverse tangent trigonometric function is used to calculate the rotation angle between the two images. Figure 3 shown.

[0072] The α-β filter is mainly used for state estimation and data smoothing. Assume that the residual between the measured value and the predicted value is The α-β filter selects the α and β parameters and uses the residual r k α times to correct the position estimate, using the residual r k β / T times is used to correct the speed estimate. Where β / T is the normalized β. To ensure system stability, the values ​​of α and β must be positive and small, and usually need to satisfy the following formula:

[0073] 0<α<1

[0074] 0<β≤2

[0075] 0<4-2α-β

[0076] This method requires experiments to adjust the values ​​of α and β, and has good stability.

[0077] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A stable target tracking system for marine radar in polar environment based on intelligent bow estimation, characterized in that: The scanning and tracking method is adopted, including a point track preprocessing module, a data association module, a track maintenance module, a track initiation module, and a track management module, which performs point track preprocessing, track initiation, track merging, track deletion, data association, and tracking maintenance processing on the target point tracks within the working cycle. The point track preprocessing module preprocesses the input data to obtain the point track information required for tracking; The measurement data generated by the point track preprocessing module enters the data association module and is associated with the relevant gates formed after the position prediction of each track. The associated target measurement data enters the track maintenance module to filter the target track; the remaining measurement data of the target without correlation enters the track starting module; the temporary track generated by the track starting module and the track updated after filtering enter the track management module together; the track management module sorts the targets, deletes the tracks that have not been updated for a specified time, merges the tracks confirmed to be generated from the same target, and outputs the updated track.

2. The marine radar polar environment target stable tracking system based on intelligent bow estimation according to claim 1 is characterized in that: The data source of the target tracking system is point track information. In the point track information, the distance information is based on the center of the ship, and the azimuth information is based on the bow information of the ship. When tracking, since the direction of the bow is constantly changing, the data processing uses the azimuth information based on due north, that is, the bow direction information of the ship and the azimuth based on the bow are used to obtain the azimuth information of the target point track based on due north. When an alarm of lost bow direction information is received, the compass value obtained by using the deep neural network-based bow direction estimation is used, and the filter parameters are adaptively adjusted to enable it to continue to maintain stable tracking.

3. The operating method of the marine radar polar environment target stable tracking system based on intelligent bow estimation as claimed in claim 1, characterized in that: When the compass fails, a new heading value is adopted. The specific operation steps are as follows: (1) Based on the newly acquired estimated heading value and time information, the heading value of the point track at the current moment is obtained by the difference principle, and the angle information of the point track relative to the heading of the ship is added to obtain the latest angle information based on true north; (2) Adjust the azimuth noise estimate and disturbance covariance value of the Kalman filter to make it compatible with the change of angular noise; (3) Adjust the filter parameters and other related logic of the track automatic start module to enhance its compatibility with angle noise; (4) The ship heading value corresponding to the point track is output together with the track for display.

4. The operating method of the marine radar polar environment target stable tracking system based on intelligent bow estimation according to claim 3 is characterized in that: The real-time ship heading estimation method used in step (1) is to apply the SuperPoint algorithm to the application scenario of ship heading estimation. The specific processing steps are as follows: (a) Processing the radar echo data to remove interference, noise and part of the sea clutter, and plotting the video lines into images in a Cartesian coordinate system to obtain a radar echo map; (b) accumulating at least two radar scanning cycles to obtain radar echo images of adjacent frames; (c) The target domain is cropped on two adjacent radar image frames to remove the pixels outside the inscribed circle in the image; (d) Input the images into the trained SuperPoint network to obtain key points and corresponding descriptors; (e) Calculate the mapping matrix corresponding to the key points and descriptors through the RANSAC algorithm; (f) using the inverse tangent function to calculate the image rotation angle corresponding to the rotation part in the mapping matrix; (g) The obtained angle is passed through an α-β filter and smoothed according to historical data to obtain a calibrated angle value.

5. The operating method of the marine radar polar environment target stable tracking system based on intelligent bow estimation according to claim 4 is characterized in that: In the specific operation of step (a), the radar saves an echo image every time it scans a circle, and obtains two adjacent images with a resolution of 1024 pixels×1024 pixels. Since the image rotation angle information is calculated, during the relative rotation change, the part outside the inscribed circle of the image may be lost in the adjacent frames.

6. The operating method of the marine radar polar environment target stable tracking system based on intelligent bow estimation according to claim 4 is characterized in that: In the specific operation of step (b), two radar images of adjacent frames are obtained, and a mask area is made with the center of the image as the center of the circle and half of the image width as the radius. Next, the image is input into the SuperPoint network, which uses a convolutional layer to extract image features, and finally obtains feature points and descriptors in the input image. Then, the BF matcher in OpenCV is used for feature matching to obtain a 3×3 two-dimensional affine transformation matrix, thereby obtaining an estimated angle. During the radar scanning process, there may be waves and other targets on the sea, which may cause noise on the echo map, so the calculated angle has a certain deviation. However, during the turning process of the ship, the angular velocity changes continuously, and the angle change law is also gentle. The α-β filter is used to smooth the turning angle of the bow to reduce error interference.

7. The operating method of the marine radar polar environment target stable tracking system based on intelligent bow estimation according to claim 4 is characterized in that: In the specific operation of step (c), a mask area is drawn with the center of the image as the center and a radius of 512, the pixels in the mask area are retained, and the pixels in the rest are removed, so as to achieve the effect of noise reduction and removal of redundant targets, retain the pixel values ​​of the mask area, delete the pixel values ​​of other areas, and obtain the cropped and scaled image.

8. The operating method of the marine radar polar environment target stable tracking system based on intelligent bow estimation according to claim 4 is characterized in that: In the specific operation of step (d), the radar image feature extraction algorithm based on SuperPoint is composed of a convolutional neural network, wherein the feature extraction part is composed of a traditional VGG network. Since the radar image is insensitive to changes in the color and texture of the target, the number of channels in the network is set to half of the original VGG network, which can not only run efficiently, but also meet the feature extraction capability of the radar image. The processed image is input into the SuperPoint network for feature extraction to obtain the key points and descriptors of the image. The BF matching algorithm in the OpenCV tool library is used for feature matching, and the first 1000 sets of matched feature points are taken to calculate the affine transformation matrix; The input shape of a single image is (N, C in , H in , W in ), the output shape after the convolution layer is (N, C out , H out , W out ), then the calculation formula of the convolutional layer is as follows: Among them, "*" represents the cross product operation of the weight matrix W and the input layer, N represents the batch size, C represents the number of channels, and H in and H out Represents the height of the input and output feature maps, W in and W out Indicates the width of the input and output feature maps; During the model training process, in order to save cache, the size of the input image is set to 512 pixels × 512 pixels. However, in the actual application process, the image size is adjusted to 1024 pixels × 1024 pixels to better retain the detail information of the radar echo image. The image is subjected to feature extraction and dimensionality reduction through a shared coding network, so that the width and height of the feature map are reduced to 1 / 8 of the original. Subsequently, the feature map is processed by two decoders, namely, a feature point detection decoder and a descriptor decoder. The output of the feature point detection decoder represents the probability that each pixel in the image is a key point, while the output of the descriptor decoder is a semi-dense descriptor corresponding to the key point. This output is bilinearly interpolated to obtain a complete descriptor. Finally, a descriptor of unit length is obtained through L2 normalization. This algorithm is suitable for detecting corner information of targets in images. Compared with traditional algorithms such as SIFT, ORB, and SURF, it is more robust and has a wider adaptability. Its performance depends on the parameter settings during the data set and model training process.

9. The operating method of the marine radar polar environment target stable tracking system based on intelligent bow estimation according to claim 4 is characterized in that: The steps (e) and (f) extract the rotation part of the matrix, calculate the rotation angle by the inverse tangent function, use the BF matching algorithm to perform key point matching, and then use the RANSAC method to calculate the optimal single mapping transformation matrix M between multiple two-dimensional point pairs, as shown below: The 2×2 submatrix in the upper left corner is the rotation part. The inverse tangent trigonometric function is used to calculate the rotation angle between the two images. The obtained rotation angle is input into the α-β filter for correction to obtain the corrected final rotation angle.

10. The operating method of the marine radar polar environment target stable tracking system based on intelligent bow estimation according to claim 4, characterized in that: In the specific operation of step (g), an α-β filter is used for state estimation and data smoothing. Assuming that the residual between the measured value and the predicted value is The α-β filter selects the α and β parameters and uses the residual r k α times to correct the position estimate, using the residual r k The speed estimation is corrected by β / T times, where β / T is the normalized β. To ensure system stability, the values ​​of α and β must be positive and small, and usually need to satisfy the following formula: 0<α< 0<β≤2 0<4-2α-β This method requires experiments to adjust the values ​​of α and β, and has good stability.

Images