A method for optimizing selection of a center point of marine vessel positioning based on multi-source feature constraints

Abstract

The present application relates to the technical field of unmanned aerial vehicle vision positioning, in particular to a marine vessel positioning center point optimization selection method based on multi-source feature constraint, comprising the following steps: data input, obtaining unmanned aerial vehicle attitude information, camera parameters and initial detection frame of the vessel; the unmanned aerial vehicle attitude information includes longitude and latitude, height, pitch angle, roll angle, yaw angle and three-dimensional motion speed. The present application combines the unmanned aerial vehicle attitude, camera parameters and marine vessel positioning center point optimization selection method based on multi-source features, which is suitable for the longitude and latitude precise visual positioning of the marine vessel by the unmanned aerial vehicle under the complex sea conditions of sea waves, reflection, fog, side view, oblique view and vertical view, and solves the problem of traditional rectangular frame center view angle sensitivity and low precision by correcting the projection deformation and background interference through multi-source attitude constraint, which is suitable for the visual precise positioning of the marine vessel by the unmanned aerial vehicle and can be applied to maritime supervision, search and rescue, fishery management and other scenes.

Description

Technical Field

[0001] This invention relates to the field of UAV visual positioning technology, specifically to a method for optimizing the selection of the positioning center point of marine vessels based on multi-source feature constraints. Background Technology

[0002] UAV visual positioning technology calculates the geographic coordinates of a target by using the pixel coordinates of the target in an image, the UAV's attitude, and camera parameters, followed by perspective transformation. Determining the target's pixel coordinates is a crucial factor affecting the final positioning result. In existing technologies, the selection of pixel coordinates relies on the geometric center of the bounding box output by the target detection algorithm, which has the following drawbacks:

[0003] In ocean scenes, the rectangular frame of a ship often includes non-ship areas such as seawater, waves, and sunlight reflections, causing the center coordinates to deviate from the actual ship hull.

[0004] The optimal reference point for a ship differs depending on the viewing angle (top view / side view), but traditional methods do not dynamically adjust the detection strategy based on the ship's own characteristics (outline, motion), and therefore cannot adapt to changes in viewing angle.

[0005] The coordinates selected in a single frame are easily affected by image noise, and the coordinates in consecutive frames change significantly, resulting in large fluctuations in the positioning results.

[0006] Based on this, the present invention provides a method for optimizing the selection of the positioning center point of marine vessels based on multi-source feature constraints, so as to solve the above-mentioned technical problems. Summary of the Invention

[0007] The purpose of this invention is to provide a method for optimizing the selection of the positioning center point of marine vessels based on multi-source feature constraints, thereby solving the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] This invention proposes an optimized selection method for the positioning center point of marine vessels based on multi-source feature constraints, comprising the following steps:

[0010] S1: Data input, acquiring UAV attitude information, camera parameters, and initial vessel detection bounding box; the UAV attitude information includes the UAV's latitude and longitude (… , ),high Pitch angle Roll angle φ, yaw angle ψ, and three-dimensional motion speed The camera parameters include the camera intrinsic parameter matrix. and the camera's attitude angle relative to the drone ( , The initial detection box is the rectangular bounding box of the ship in the image. ;in( ) represents the pixel coordinates of the top-left corner of the rectangle. () represents the bottom right pixel coordinates;

[0011] S2: Mask optimization, calculating the theoretical ground projection range corresponding to the initial detection box B based on the UAV attitude information and camera parameters. The confidence level of the semantic segmentation results of ships within the initial detection box B is adjusted to obtain an optimized mask. ;

[0012] S3: Detection of key structural points;

[0013] S4: Weighted fusion, calculating the optimized mask. weighted center of mass Based on the pitch angle θ of the UAV and the confidence level of the key structural points Calculate dynamic weight λ and fuse The fusion center point is obtained with the key structural points. ;

[0014] S5: Timing filtering, combined with the drone's movement speed. A Kalman filter model is constructed using information such as the fusion center point of consecutive frames to perform smoothing processing, and the final center point is output. .

[0015] Preferably, step S3, the detection of key structural points, specifically includes the following steps:

[0016] S3.1. Viewpoint Judgment: If the camera's total pitch angle + ≥60° is considered a direct downward view; if <60°, determined to be a side view / oblique downward view angle;

[0017] S3.2. Ship Orientation Determination: For side / oblique top views, the bow and stern orientations of the ship are determined by the ship's static outline features, which are the narrow end shape of the bow, the wide end shape of the stern, and the mid-to-rear position of the superstructure.

[0018] S3.3. Type Adaptation Detection: A preset classification model is used to identify the vessel type, including cargo ships, fishing boats, and yachts. The vessel projection viewpoint is determined based on the UAV attitude information, and the key structural points and their confidence levels under the corresponding viewpoint are detected in combination with the vessel type. .

[0019] Preferably, in step S3.2, the rule for determining the static outline features of the vessel is as follows:

[0020] Bow identification: The end of the vessel in the image that has a "sharp and narrow" shape is the bow;

[0021] Stern identification: The end of the ship in the image that is "wide and flat" and has a superstructure is the stern;

[0022] Superstructure determination: Identify the superstructure region of the ship based on semantic segmentation results.

[0023] Preferably, in step S3.3, the detection rule for the key structural points is as follows:

[0024] A top-down view ( ≥60°): Detect the midpoint of the ship's centerline. The central axis is optimized by calculation using a mask. The midpoint of the longer side of the smallest bounding rectangle is determined;

[0025] Side view / oblique top view ( <60°: The geometric center of the superstructure at the stern is used as the center of the cockpit. The superstructure area is determined by matching the pre-stored structural features corresponding to the ship type. The pre-stored structural features are the relative size and positional relationship between the superstructure and the ship hull.

[0026] Preferably, in step S2, the theoretical range of ground projection... The calculation process is as follows:

[0027] Step S2.1: Calculate the total camera pitch angle = + Total yaw angle = + Total roll angle =φ+

[0028] Step S2.2: Based on UAV altitude , , , The camera intrinsic parameter matrix K is used to convert the pixel coordinates of the four vertices of the initial detection box B into coordinates in the ground coordinate system using the perspective projection formula. The origin is the ground point directly below the drone, the X-axis points in the direction of the drone's yaw angle, and the Y-axis is perpendicular to the X-axis, forming... ;

[0029] Step S2.3: The confidence correction method is as follows: adjust the initial mask of semantic segmentation. The ground coordinates of the medium pixel exceed The confidence score of each pixel is multiplied by a penalty coefficient α to obtain the optimized mask. .

[0030] Preferably, in step S4, the weighted centroid The calculation formula is:

[0031] ;

[0032] ;

[0033] in, To optimize the mask The set of all pixels in ( , () represents the coordinates of pixel p. For pixel p in Optimize confidence in =( );

[0034] Preferably, in step S4, the dynamic weights The calculation formula is:

[0035] ;

[0036] in, , For the preset proportional coefficient and + =1, The confidence level of the key structural points; the fusion center point The coordinates are:

[0037] ;

[0038] ;

[0039] in,( , () represents the pixel coordinates of the key structural points.

[0040] Preferably, in step S5, the state vector of the Kalman filter model is defined as ,in:( The points are the fusion center points of the current frame, respectively. Horizontal and vertical pixel coordinates; These represent the pixel motion speeds of the fusion center point in the horizontal and vertical directions, respectively, through consecutive frames. The coordinate difference and the time interval between two frames are calculated and initialized; the observed value of the model is the fusion center point of the current frame. pixel coordinates ( , The observation matrix is ​​used to correlate the state vector ( ) and observed values.

[0041] Preferably, in step S5, the control term of the Kalman filter model includes the UAV's absolute ground motion speed ( , and yaw angle change rate The control term is mapped to a correction amount for pixel motion through a control matrix, which is used to compensate for the interference of UAV motion on the pixel displacement of the ship, so that the state prediction conforms to the relative motion law of the UAV and the ship.

[0042] Compared with the prior art, the beneficial effects of the present invention are:

[0043] This invention provides an optimized method for selecting the center point of a marine vessel's positioning, combining UAV attitude, camera parameters, and multi-source features of the vessel. It is applicable to the precise visual positioning of marine vessels' latitude and longitude by UAVs in complex sea conditions such as waves, reflections, and fog, as well as under dynamic perspectives including side-view, oblique-view, and front-view. By correcting projection distortion and background interference through multi-source attitude constraints, it solves the problems of sensitivity and low accuracy of the center view of traditional rectangular frames. It is suitable for the precise visual positioning of marine vessels by UAVs and can be applied to scenarios such as maritime supervision, search and rescue, and fisheries management. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the method for optimizing the selection of the positioning center point of marine vessels based on multi-source feature constraints according to the present invention. Detailed Implementation

[0045] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0046] Example 1, please refer to Figure 1 This invention proposes an optimized selection method for the positioning center point of marine vessels based on multi-source feature constraints, comprising the following steps:

[0047] S1: Data input, acquiring UAV attitude information, camera parameters, and initial vessel detection bounding box; the UAV attitude information includes the UAV's latitude and longitude (… , ),high Pitch angle Roll angle φ, yaw angle ψ, and three-dimensional motion speed The camera parameters include the camera intrinsic parameter matrix. and the camera's attitude angle relative to the drone ( , The initial detection box is the rectangular bounding box of the ship in the image. ;in( ) represents the pixel coordinates of the top-left corner of the rectangle. () represents the bottom right pixel coordinates;

[0048] S2: Mask optimization, calculating the theoretical ground projection range corresponding to the initial detection box B based on the UAV attitude information and camera parameters. The confidence level of the semantic segmentation results of ships within the initial detection box B is adjusted to obtain an optimized mask. ;

[0049] S3: Detection of key structural points;

[0050] S4: Weighted fusion, calculating the optimized mask. weighted center of mass Based on the pitch angle θ of the UAV and the confidence level of the key structural points Calculate dynamic weight λ and fuse The fusion center point is obtained with the key structural points. ;

[0051] S5: Timing filtering, combined with the drone's movement speed. A Kalman filter model is constructed using information such as the fusion center point of consecutive frames to perform smoothing processing, and the final center point is output. .

[0052] In this embodiment, it should also be noted that step S3, key structural point detection, specifically includes the following steps:

[0053] S3.1. Viewpoint Judgment: If the camera's total pitch angle + ≥60° is considered a direct downward view; if <60°, determined to be a side view / oblique downward view angle;

[0054] S3.2. Ship Orientation Determination: For side / oblique top views, the bow and stern orientations of the ship are determined by the ship's static outline features, which are the narrow end shape of the bow, the wide end shape of the stern, and the mid-to-rear position of the superstructure.

[0055] S3.3. Type Adaptation Detection: A preset classification model is used to identify the vessel type, including cargo ships, fishing boats, and yachts. The vessel projection viewpoint is determined based on the UAV attitude information, and the key structural points and their confidence levels under the corresponding viewpoint are detected in combination with the vessel type. .

[0056] In this embodiment, it should also be noted that the judgment rule for the static outline features of the ship in step S3.2 is as follows:

[0057] Bow identification: The end of the vessel in the image that has a "sharp and narrow" shape is the bow;

[0058] Stern identification: The end of the ship in the image that is "wide and flat" and has a superstructure is the stern;

[0059] Superstructure determination: Identify the superstructure region of the ship based on semantic segmentation results.

[0060] In this embodiment, it should also be noted that the detection rule for the key structural points in step S3.3 is as follows:

[0061] A top-down view ( ≥60°): Detect the midpoint of the ship's centerline. The central axis is optimized by calculation using a mask. The midpoint of the longer side of the smallest bounding rectangle is determined;

[0062] Side view / oblique top view ( <60°: The geometric center of the superstructure at the stern is used as the center of the cockpit. The superstructure area is determined by matching the pre-stored structural features corresponding to the ship type. The pre-stored structural features are the relative size and positional relationship between the superstructure and the ship hull.

[0063] In this embodiment, it should also be noted that in step S2, the theoretical range of ground projection... The calculation process is as follows:

[0064] Step S2.1: Calculate the total camera pitch angle = + Total yaw angle = + Total roll angle =φ+

[0065] Step S2.2: Based on UAV altitude , , , The camera intrinsic parameter matrix K is used to convert the pixel coordinates of the four vertices of the initial detection box B into coordinates in the ground coordinate system using the perspective projection formula. The origin is the ground point directly below the drone, the X-axis points in the direction of the drone's yaw angle, and the Y-axis is perpendicular to the X-axis, forming... ;

[0066] Step S2.3: The confidence correction method is as follows: adjust the initial mask of semantic segmentation. The ground coordinates of the medium pixel exceed The confidence score of each pixel is multiplied by a penalty coefficient α to obtain the optimized mask. .

[0067] In this embodiment, it should also be noted that in step S4, the weighted centroid... The calculation formula is:

[0068] ;

[0069] ;

[0070] in, To optimize the mask The set of all pixels in ( , () represents the coordinates of pixel p. For pixel p in Optimize confidence in =( );

[0071] In this embodiment, it should also be noted that in step S4, the dynamic weights The calculation formula is:

[0072] ;

[0073] in, , For the preset proportional coefficient and + =1, The confidence level of the key structural points; the fusion center point The coordinates are:

[0074] ;

[0075] ;

[0076] in,( , () represents the pixel coordinates of the key structural points.

[0077] In this embodiment, it should also be noted that in step S5, the state vector of the Kalman filter model is defined as follows: ,in:( The points are the fusion center points of the current frame, respectively. Horizontal and vertical pixel coordinates; These represent the pixel motion speeds of the fusion center point in the horizontal and vertical directions, respectively, through consecutive frames. The coordinate difference and the time interval between two frames are calculated and initialized; the observed value of the model is the fusion center point of the current frame. pixel coordinates ( , The observation matrix is ​​used to correlate the state vector ( ) and observed values.

[0078] In this embodiment, it should also be noted that in step S5, the control term of the Kalman filter model includes the UAV's absolute ground motion speed ( , and yaw angle change rate The control term is mapped to a correction amount for pixel motion through a control matrix, which is used to compensate for the interference of UAV motion on the pixel displacement of the ship, so that the state prediction conforms to the relative motion law of the UAV and the ship.

[0079] Example 2, please refer to Figure 1 In practical applications, this embodiment is based on a conventional marine vessel UAV monitoring scenario, adopting industry-standard hardware and algorithm frameworks, requiring no special customized equipment. The specific configuration is as follows:

[0080] Unmanned Aerial Vehicle System: Commercial multi-rotor drone equipped with an RTK-GPS module (to meet the positioning accuracy requirements of the ocean) and an IMU module (which can output attitude angle and speed in real time).

[0081] Camera module: Industrial-grade camera, with a valid intrinsic parameter matrix (including focal length and principal point coordinates) obtained through factory calibration, supporting frame rates of 30fps and above, and image resolution adapted to the processing capabilities of the UAV's onboard computing unit.

[0082] Algorithm deployment: Lightweight object detection, semantic segmentation, ship classification and key point detection models are deployed on the UAV's onboard embedded computing unit to ensure real-time inference (inference speed is no less than the camera frame rate).

[0083] Application scenarios: Open ocean areas, targeting conventional commercial vessels (such as cargo ships and fishing boats), without severe obstructions (such as dense fog or large waves), with drone flight altitude coverage of 50~200m, and projection angles including direct overhead view, oblique overhead view, and side view.

[0084] Step S1: Multi-source data input:

[0085] The drone and camera system are activated. The drone's latitude and longitude and relative sea level are obtained through RTK-GPS, and the pitch angle, roll angle, yaw angle and three-dimensional motion speed are obtained through IMU. All data are transmitted to the onboard computing unit in real time.

[0086] Call the pre-stored camera intrinsic parameter matrix and the camera's installation attitude angle relative to the UAV to calculate the camera's total attitude angles (total pitch angle, total roll angle, and total yaw angle).

[0087] Start the target detection algorithm to detect ships in the images captured by the camera in real time, and output the initial rectangular detection box of the ship;

[0088] All input data is validated to remove abnormal data (such as lost GPS signal or incomplete detection frame). If the data is abnormal, the valid data from the previous frame is used to ensure the continuity of the process.

[0089] Step S2: Mask optimization:

[0090] Based on the drone's altitude, the camera's total attitude angle, and the intrinsic parameter matrix, the pixel coordinates of the four vertices of the initial detection box are converted into coordinates in the ground coordinate system using the principle of perspective projection, thus forming the theoretical range of ground projection (reflecting the actual range of the ship).

[0091] Semantic segmentation is performed on the image region within the initial detection box, and an initial mask is output. The confidence score of each pixel in the mask represents the probability that the pixel belongs to the hull.

[0092] For each pixel in the initial mask, its corresponding ground coordinates are calculated through inverse perspective transformation;

[0093] If the ground coordinates are outside the theoretical range of ground projection, they are determined to be background pixels, and their confidence level is multiplied by a preset penalty coefficient (the coefficient range is 0 < α ≤ 0.3, and the specific value is adjusted according to the distance between the pixel and the boundary of the projection range).

[0094] If the result is within the range, the original confidence level is retained, and an optimized mask is generated.

[0095] Step S3: Detection of key structural points:

[0096] Based on the total pitch angle of the camera, if the total pitch angle is greater than the preset threshold (60°), it is determined to be a top-down view.

[0097] If the total pitch angle is less than the preset threshold, it is determined to be a side view / oblique downward view.

[0098] The ship's outline is extracted using an edge detection algorithm, and the end with a "sharp and narrow" shape is identified as the bow, while the end with a "wide and flat shape and a superstructure" is identified as the stern.

[0099] The superstructure is identified using semantic segmentation results. Ship types (such as cargo ships, fishing boats, and yachts) are identified through a ship classification model. Combining pre-stored structural features of the ship type (such as the relative position and size relationship between the cockpit and the hull) and the current projection viewpoint, key structural points are detected at the corresponding viewpoints (the midpoint of the hull's centerline is detected from a top-down viewpoint, and the center of the cockpit of the superstructure at the stern is detected from a side / oblique top-down viewpoint). The pixel coordinates and detection confidence of the key structural points are then output.

[0100] Step S4: Weighted fusion:

[0101] Weighted centroid calculation: Taking all pixels in the optimized mask as the calculation object, and using the corrected confidence of each pixel as the weight, the pixel coordinates are weighted and averaged to obtain the weighted centroid.

[0102] Dynamic weight calculation: Based on the total camera pitch angle and the detection confidence of key structural points, dynamic weights are calculated according to a preset formula (weight range 0~1).

[0103] The larger the total pitch angle (the closer it is to a direct top view), the higher the weighting percentage of the weighted centroid.

[0104] The higher the confidence level of a key structural point, the higher its weight percentage.

[0105] Fusion center point calculation: The pixel coordinates of the weighted centroid and key structural points are fused according to dynamic weights to obtain the fusion center point. This center point combines the anti-background interference capability of the weighted centroid with the physical correlation of the key structural points.

[0106] Step S5: Timing Filtering

[0107] Initialize the state vector (including the pixel coordinates of the fusion center point and the pixel motion velocity), state covariance matrix, process noise covariance matrix and observation noise covariance matrix of the Kalman filter model;

[0108] The velocity parameters in the state vector are initialized using the coordinate difference of the fusion center point of consecutive frames and the frame interval time;

[0109] The absolute ground motion speed and yaw rate of change of the UAV are introduced as control terms. The motion parameters of the UAV are mapped into pixel motion corrections through a preset control matrix to compensate for the interference of the UAV motion on the pixel displacement of the ship.

[0110] For each frame fusion center point, perform state prediction and observation update iteration: predict the current frame state based on the optimal state and control term of the previous frame, and correct the predicted state by combining the observation value of the current frame fusion center point to obtain the optimal state of the current frame;

[0111] The observation noise covariance matrix is ​​dynamically adjusted based on the detection confidence level of the fusion center point.

[0112] The position parameters are extracted from the optimal state of the current frame to obtain the final positioning center point, which is used for subsequent ship latitude and longitude calculations.

[0113] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0114] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for optimizing the selection of the positioning center point of marine vessels based on multi-source feature constraints, characterized in that, Includes the following steps: S1: Data input, acquiring UAV attitude information, camera parameters, and initial vessel detection bounding box; the UAV attitude information includes the UAV's latitude and longitude (… , ),high Pitch angle Roll angle φ, yaw angle ψ, and three-dimensional motion speed The camera parameters include the camera intrinsic parameter matrix. and the camera's attitude angle relative to the drone ( , The initial detection box is the rectangular bounding box of the ship in the image. ;in( ) represents the pixel coordinates of the top-left corner of the rectangle. () represents the bottom right pixel coordinates; S2: Mask optimization, calculating the theoretical ground projection range corresponding to the initial detection box B based on the UAV attitude information and camera parameters. The confidence level of the semantic segmentation results of ships within the initial detection box B is adjusted to obtain an optimized mask. ; S3: Detection of key structural points; S4: Weighted fusion, calculating the optimized mask. weighted center of mass Based on the pitch angle θ of the UAV and the confidence level of the key structural points Calculate dynamic weight λ and fuse The fusion center point is obtained with the key structural points. ; S5: Timing filtering, combined with the drone's movement speed. A Kalman filter model is constructed using information such as the fusion center point of consecutive frames to perform smoothing processing, and the final center point is output. .

2. The method according to claim 1, characterized in that, The key structural point detection in step S3 specifically includes the following steps: S3.

1. Viewpoint Judgment: If the camera's total pitch angle + ≥60° is considered a direct downward view; if <60°, determined to be a side view / oblique downward view angle; S3.

2. Ship Orientation Determination: For side / oblique top views, the bow and stern orientations of the ship are determined by the ship's static outline features, which are the narrow end shape of the bow, the wide end shape of the stern, and the mid-to-rear position of the superstructure. S3.

3. Type Adaptation Detection: A preset classification model is used to identify the vessel type, including cargo ships, fishing boats, and yachts. The vessel projection viewpoint is determined based on the UAV attitude information, and the key structural points and their confidence levels under the corresponding viewpoint are detected in combination with the vessel type. .

3. The method according to claim 2, characterized in that, In step S3.2, the rule for determining the static outline features of the vessel is as follows: Bow identification: The end of the ship in the image that has a "sharp and narrow" shape is the bow; Stern identification: The end of the ship in the image that is "wide and flat" and has a superstructure is the stern; Superstructure determination: Identify the superstructure region of the ship based on semantic segmentation results.

4. The method according to claim 3, characterized in that, In step S3.3, the detection rule for the key structural points is as follows: A top-down view ( ≥60°): Detect the midpoint of the ship's centerline. The central axis is optimized by calculation using a mask. The midpoint of the longer side of the smallest bounding rectangle is determined; Side view / oblique top view ( <60°: The geometric center of the superstructure at the stern is used as the center of the cockpit. The superstructure area is determined by matching the pre-stored structural features corresponding to the ship type. The pre-stored structural features are the relative size and positional relationship between the superstructure and the ship hull.

5. The method according to claim 4, characterized in that, In step S2, the theoretical range of ground projection The calculation process is as follows: Step S2.1: Calculate the total camera pitch angle = + Total yaw angle = + Total roll angle =φ+ Step S2.2: Based on UAV altitude , , , The camera intrinsic parameter matrix K is used to convert the pixel coordinates of the four vertices of the initial detection box B into coordinates in the ground coordinate system using the perspective projection formula. The origin is the ground point directly below the drone, the X-axis points in the direction of the drone's yaw angle, and the Y-axis is perpendicular to the X-axis, forming... ; Step S2.3: The confidence correction method is as follows: adjust the initial mask of semantic segmentation. The ground coordinates of the medium pixel exceed The confidence score of each pixel is multiplied by a penalty coefficient α to obtain the optimized mask. .

6. The method according to claim 5, characterized in that, In step S4, the weighted centroid The calculation formula is: ； ； in, To optimize the mask The set of all pixels in ( , () represents the coordinates of pixel p. For pixel p in Optimize confidence in =( ).

7. The method according to claim 6, characterized in that, In step S4, the dynamic weights The calculation formula is: ； in, , For the preset proportional coefficient and + =1, The confidence level of the key structural points; the fusion center point The coordinates are: ； ； in,( , () represents the pixel coordinates of the key structural points.

8. The method according to claim 7, characterized in that, In step S5, the state vector of the Kalman filter model is defined as follows: ,in:( The points are the fusion center points of the current frame, respectively. Horizontal and vertical pixel coordinates; These represent the pixel motion speeds of the fusion center point in the horizontal and vertical directions, respectively, through consecutive frames. The coordinate difference and the time interval between two frames are calculated and initialized; the observed value of the model is the fusion center point of the current frame. pixel coordinates ( , The observation matrix is ​​used to correlate the state vector ( ) and observed values.

9. The method according to claim 8, characterized in that, In step S5, the control term of the Kalman filter model includes the UAV's absolute ground motion velocity ( , and yaw angle change rate The control term is mapped to a correction amount for pixel motion through a control matrix, which is used to compensate for the interference of UAV motion on the pixel displacement of the ship, so that the state prediction conforms to the relative motion law of the UAV and the ship.

Images