Optical guiding method for recovery of deep-sea unmanned underwater vehicle

Abstract

The invention provides an optical guiding method for recycling a deep-sea unmanned underwater vehicle, and belongs to the technical field of underwater vehicle recycling. Four blue-green LED guiding light sources are non-uniformly and annularly distributed and installed on the front end face of a recycling cage, the three-dimensional coordinates of the guiding light sources are calibrated in advance, and guiding light source images are collected in real time through a binocular camera; an adaptive exposure control technology is adopted to adjust image quality, statistical information calculation is performed on an identified light source contour, an effective guide light source target is determined through convex constraint configuration judgment and ellipse fitting screening, four guide light sources are constructed as a traveling salesman problem, and a nearest neighbor heuristic algorithm is adopted to complete number matching. A rotation matrix and a translation vector are solved through a singular value decomposition method according to the pre-measured coordinates and the real-time calculated coordinates, and collaborative optimization is carried out at the same time, so that the technical problems of insufficient optical guidance positioning precision and poor real-time performance in the recovery process of the unmanned underwater vehicle in the deep sea environment are solved.

Description

Technical Field

[0001] This invention belongs to the field of underwater vehicle recovery technology, and more specifically, relates to an optical guidance method for the recovery of deep-sea unmanned underwater vehicles. Background Technology

[0002] The recovery of deep-sea unmanned underwater vehicles (UUVs) is a key technological aspect of marine engineering. Traditional recovery guidance methods primarily employ acoustic and magnetic guidance technologies. These methods utilize acoustic beacons or magnetic markers on the recovery cage, leveraging sound wave propagation or magnetic field variations to achieve coarse positioning and guidance of the vehicle. These methods have been widely applied in the complex deep-sea environment, providing fundamental technical support for the safe recovery of UUVs. However, traditional acoustic guidance methods have significant limitations in close-range precise positioning. The propagation of acoustic signals in seawater is easily affected by factors such as ocean current disturbances, temperature layer changes, and seabed reflections, leading to decreased positioning accuracy. Furthermore, the relatively slow response speed of acoustic systems makes it difficult to meet the real-time control requirements of dynamic recovery processes. While magnetic guidance methods offer some anti-interference capabilities, their effective range is limited and they are susceptible to the influence of the seabed's geomagnetic environment. In existing technologies, when an unmanned underwater vehicle (UUV) approaches the recovery cage for final docking, the lack of high-precision close-range guidance makes it difficult for the UUV to accurately identify the precise position and attitude information of the recovery cage. This results in a low recovery success rate and prolonged recovery time. Furthermore, traditional guidance methods lack robustness in complex marine environments and are easily affected by factors such as water turbidity, changes in lighting, and ocean current disturbances, impacting the reliability and efficiency of the entire recovery system. In other words, existing technologies suffer from insufficient optical guidance and positioning accuracy and poor real-time performance during the recovery of UUVs in deep-sea environments. Summary of the Invention

[0003] In view of this, the present invention provides an optical guidance method for the recovery of deep-sea unmanned underwater vehicles, which can solve the technical problems of insufficient optical guidance and positioning accuracy and poor real-time performance in the recovery of unmanned underwater vehicles in deep-sea environments in the prior art.

[0004] This invention is implemented as follows: It provides an optical guidance method for recovering deep-sea unmanned underwater vehicles (UUVs). Four blue-green LED guidance light sources are installed on the front face of the recovery cage. The three-dimensional spatial coordinates of the guidance light sources are measured and stored using an industrial photogrammetry system. The intrinsic and extrinsic parameters of the binocular camera at the bow of the UUV are calibrated, and upper-level and lower-level game models are established. The UUV is guided to a distance of 15m from the recovery cage using acoustic guidance. The binocular camera image acquisition system is activated, and the guidance light source target is detected. An adaptive exposure control mathematical model is established, images are acquired, and the guidance light source target is segmented and its pixel size is calculated. The binocular camera synchronously acquires images of the guidance light sources, and a rolling shutter is used. The Ball method removes background light interference and generates a maximum value mask image and a spot contour mask through automatic contrast enhancement. Statistical information is calculated on the identified light source contours, and four effective guide light source targets are determined by convex constraint configuration judgment and ellipse fitting major-minor axis ratio screening. These four guide light source targets are constructed as traveling salesman problem nodes, and the nearest neighbor heuristic algorithm is used to complete the guide light source number matching. The centroid method is used to achieve precise positioning of the guide light source center, and the epipolar constraint principle is used to complete the same-name matching of the guide light source center point in the binocular image. The three-dimensional coordinates of the guide light source center are calculated based on the binocular vision measurement principle. Based on the pre-measured three-dimensional coordinate information of the guide light source and the calculated three-dimensional coordinates of the guide light source in the binocular coordinate system, the rotation matrix and translation vector are solved by the singular value decomposition method to calculate the relative pose relationship between the unmanned underwater vehicle and the recovery cage.

[0005] The four blue-green LED guide lights are installed in a non-uniform ring distribution. The upper-level game model aims to maximize the recovery success rate, while the lower-level game model aims to minimize system energy consumption. The lower-level game model is activated when the guide light is detected.

[0006] Specifically, the steps of the adaptive exposure control mathematical model are as follows: initialize the stereo camera exposure value to 25ms; if the pixel size is less than 25 pixels, adjust the exposure value to the original value multiplied by 1.20; if the pixel size is greater than 80 pixels, adjust the exposure value to the original value multiplied by 0.75.

[0007] The industrial photogrammetry system is specifically a high-precision three-dimensional coordinate measurement device based on the principle of stereo vision. It uses multiple calibrated cameras to photograph the target and calculates the three-dimensional coordinates of spatial points using the principle of triangulation, achieving a measurement accuracy of sub-millimeter level.

[0008] Specifically, the Rolling Ball method is an image background removal technique based on morphological operations that effectively separates foreground objects from unevenly lit background areas by simulating the rolling of a ball on the image surface.

[0009] Specifically, the automatic contrast enhancement process involves calculating the minimum grayscale value of the image. and the maximum gray value of the image Calculate the grayscale value of the enhanced image pixel by pixel. Where (x, y) are the pixel coordinates. This represents the grayscale of the original image.

[0010] Specifically, the steps for calculating the statistical information include obtaining the center coordinates, contour area, average gray level of the contour, length of the major axis of the ellipse fitting, and length of the minor axis of the ellipse fitting.

[0011] The step of calculating the three-dimensional coordinates of the guiding light source in the binocular coordinate system further includes defining the centroids of two sets of three-dimensional point sets and calculating the centroid-free coordinates. Specifically, the centroid-free coordinates are the relative coordinates obtained by subtracting the centroid coordinates of the corresponding point set from the three-dimensional point coordinates, calculated using the following formula: and ,in and These are the centroids of the two point sets, respectively.

[0012] Specifically, the singular value decomposition method is a mathematical method in linear algebra that decomposes a matrix into a product of three matrices, by constructing a matrix... Singular value decomposition of matrix W yields Where U and V are orthogonal matrices, For diagonal matrices, rotation matrices .

[0013] Specifically, the translation vector is a three-dimensional vector describing the translation relationship between two coordinate systems, and its calculation formula is as follows: .

[0014] Specifically, the maximum value mask image is initialized as an all-zero matrix with the same size as the original image. For each pixel, if its grayscale value is not less than the grayscale value of pixels within an 8-connected region, the pixel is considered a maximum value and marked as 1 at the corresponding position. The Traveling Salesman Problem is a classic combinatorial optimization problem that optimizes the matching order of guide light source numbers by finding the shortest path to all nodes. The nearest neighbor heuristic algorithm is a greedy algorithm for solving the Traveling Salesman Problem, starting from any node and selecting the nearest unvisited node as the next node to be visited.

[0015] Specifically, the centroid method determines the center position of the target by calculating the weighted average of the coordinates of all pixels within the target area. The epipolar constraint principle is a fundamental geometric constraint relationship in binocular stereo vision, using epipolar geometry to constrain the search range of corresponding points in the left and right images. The binocular vision measurement principle involves using two cameras to observe the same target from different perspectives and calculating the target's three-dimensional coordinates using triangulation principles.

[0016] This invention addresses the technical problems of insufficient positioning accuracy and poor real-time performance in traditional guidance methods by establishing a binocular vision measurement system based on blue-green LED guidance light sources and employing a two-layer game theory model to optimize the recovery strategy. The invention utilizes a non-uniform ring distribution design of four blue-green LED guidance light sources, combined with pre-calibrated three-dimensional spatial coordinate information from an industrial photogrammetry system. Real-time image acquisition and precise identification and positioning of the guidance light sources are achieved using a binocular camera. Rollingball background removal technology and automatic contrast enhancement processing improve the accuracy and robustness of light source identification. Adaptive exposure control and centroid-based precise positioning technology significantly enhance the measurement accuracy of the guidance light source center coordinates. High-precision calculation of the three-dimensional coordinates of the guidance light sources is achieved using binocular stereo matching and triangulation based on the epipolar constraint principle. The relative pose relationship between the underwater vehicle and the recovery cage is solved using singular value decomposition, providing reliable spatial positioning information for accurate recovery. In summary, this invention solves the technical problems of insufficient optical guidance and positioning accuracy and poor real-time performance in the recovery of unmanned underwater vehicles in deep-sea environments, as mentioned in the background section. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention.

[0018] Figure 2 This is a schematic diagram of the overall scheme of the optical guidance system in Example 2.

[0019] Figure 3 The diagram shows the camera mounting structure in Example 2, which includes three sub-diagrams: sub-diagram A is a left view, sub-diagram B is a perspective view, and sub-diagram C is a top view.

[0020] Figure 4 This is a flowchart of the optical guidance method in Example 2.

[0021] Figure 5 This is a flowchart of the light source recognition process in Example 2. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0023] like Figure 1 The diagram shows a flowchart of an optical guidance method for recovering a deep-sea unmanned underwater vehicle provided by the present invention. This method includes the following steps:

[0024] An optical guidance method for recovering deep-sea unmanned underwater vehicles includes the following steps:

[0025] S01. Four blue-green LED guide light sources are installed in a non-uniform ring distribution on the front end of the recovery cage. The three-dimensional spatial coordinate information of the guide light sources in the recovery cage is measured by an industrial photogrammetry system and stored in the configuration file of the software processing system. At the same time, the internal and external parameters of the binocular camera installed on the bow of the unmanned underwater vehicle are calibrated, and an upper-level game model with the goal of maximizing the recovery success rate is established.

[0026] S02. Guide the unmanned underwater vehicle to a distance of 15m from the recovery cage using acoustic guidance, activate the binocular camera image acquisition system and detect the guide light source target. If the guide light source is detected, activate the lower-level game model with the goal of minimizing system energy consumption.

[0027] S03. Initialize the binocular camera exposure value to 25ms, acquire images and segment the guide light source target to calculate the pixel size. If the pixel size is less than 25 pixels, adjust the exposure value to the original value multiplied by 1.20. If the pixel size is greater than 80 pixels, adjust the exposure value to the original value multiplied by 0.75. Establish an adaptive exposure control mathematical model.

[0028] S04. Synchronously acquire guide light source images using a binocular camera, remove background light interference using the Rolling Ball method, and calculate image statistical information, including the minimum and maximum gray values ​​of the image, through automatic contrast enhancement processing to generate a maximum value mask image and a light spot contour mask.

[0029] S05. Statistical information calculation is performed on the identified light source contours to obtain the center coordinates, contour area, average gray level of the contour, length of the major axis of the ellipse fitting, and length of the minor axis of the ellipse fitting. Four effective guiding light source targets are determined by judging the convex constraint configuration and screening the ratio of the major and minor axes of the ellipse fitting. The four guiding light source targets are constructed as nodes of the traveling salesman problem and the nearest neighbor heuristic algorithm is used to solve the shortest loop to complete the matching of guiding light source numbers.

[0030] S06. Accurately locate the center of the guiding light source using the centroid method and record the center coordinates. Complete the same-name matching of the center point of the guiding light source in the binocular image through the epipolar constraint principle. Calculate the three-dimensional coordinates of the center of the guiding light source based on the binocular vision measurement principle and the pre-calibrated intrinsic and extrinsic parameters of the binocular camera.

[0031] S07. Based on the pre-measured three-dimensional coordinate information of the guidance light source and the calculated three-dimensional coordinates of the guidance light source in the binocular coordinate system, define the centroids of two sets of three-dimensional point sets and calculate the centroid coordinates. Solve the rotation matrix and translation vector through the singular value decomposition method to calculate the relative pose relationship between the unmanned underwater vehicle and the recovery cage.

[0032] The industrial photogrammetry system is a high-precision three-dimensional coordinate measurement device based on the principle of stereo vision. It uses multiple calibrated cameras to photograph the target and calculates the three-dimensional coordinates of spatial points using the principle of triangulation, achieving a measurement accuracy of sub-millimeter level.

[0033] The Rolling Ball method is an image background removal technique based on morphological operations. It effectively separates foreground objects from unevenly lit background areas by simulating the rolling of a ball on the image surface.

[0034] The automatic contrast enhancement process is achieved by calculating the minimum grayscale value of the image. and the maximum gray value of the image Calculate the grayscale value of the enhanced image pixel by pixel. ,in These are the pixel coordinates. This represents the grayscale of the original image.

[0035] The maximum value mask image is initialized as an all-zero matrix, with the same image size as the original image. For each pixel, if its gray value is not less than the gray value of pixels in the eight-connected region, the pixel is considered a maximum value point and marked as 1 at the corresponding position.

[0036] The spot contour mask is initialized to be consistent with the maxima mask. For the maxima points corresponding to the maxima mask image, the connected regions are searched based on the region growing method. Any maxima pixel is used as the initial seed point to start region growing. If the difference between the pixel gray level and the seed point gray level does not exceed a set threshold, it is added to the growing region.

[0037] The convex constraint configuration judgment refers to checking whether the four guiding light source targets form a convex quadrilateral geometric constraint condition, ensuring that the distribution of guiding light source targets meets the geometric requirements of spatial positioning.

[0038] The ratio of the major axis to the minor axis of the ellipse fitting refers to the ratio of the length of the major axis to the length of the minor axis of the ellipse fitting obtained after fitting the contour to an ellipse.

[0039] The Traveling Salesman Problem is a classic combinatorial optimization problem that optimizes the matching order of guide light source numbers by finding the shortest path to all nodes.

[0040] The nearest neighbor heuristic algorithm is a greedy algorithm for solving the traveling salesman problem. Starting from any node, it selects the nearest unvisited node to the current node as the next node to be visited.

[0041] The epipolar constraint principle is a fundamental geometric constraint relationship in binocular stereo vision, which uses epipolar geometry theory to constrain the search range of corresponding points in the left and right images.

[0042] The centroid method is a method for determining the center position of a target by calculating the weighted average of the coordinates of all pixels within the target area.

[0043] The binocular vision measurement principle is to use two cameras to observe the same target from different perspectives and calculate the target's three-dimensional coordinates through triangulation.

[0044] The centroid-free coordinates refer to the relative coordinates obtained by subtracting the centroid coordinates of the corresponding point set from the three-dimensional point coordinates. The calculation formula is as follows: and ,in and These are the centroids of the two point sets, respectively.

[0045] The singular value decomposition method is a mathematical method in linear algebra that decomposes a matrix into a product of three matrices. This method involves constructing a matrix... Singular value decomposition of matrix W yields Where U and V are orthogonal matrices, For diagonal matrices, rotation matrices .

[0046] The rotation matrix is ​​a 3×3 orthogonal matrix that describes the rotational relationship between two coordinate systems.

[0047] The translation vector is a three-dimensional vector describing the translation relationship between two coordinate systems, and its calculation formula is as follows: .

[0048] The objective function of the upper-level game model is: .

[0049] The objective function of the lower-level game model is: .

[0050] The positioning accuracy parameter is derived from a comprehensive evaluation of the accuracy of the guide light source recognition, the binocular vision measurement error, and the pose calculation accuracy, and is used for the calculation of the objective function of the upper-level game model and the evaluation of the success rate of recovery.

[0051] The recovery time parameter is derived from the sum of image processing time, pose calculation time, vehicle adjustment time, and communication delay time, and is used for the calculation of the objective function of the upper-level game model and the optimization of recovery efficiency.

[0052] The energy consumption parameters are derived from the sum of the power consumption of the LED guiding light source, the power consumption of the binocular camera, the power consumption of computing and processing, and the power consumption of the communication system, and are used for the calculation of the objective function of the upper-level game model and the lower-level game model.

[0053] The maximum allowable time parameter is derived from the task time constraints and safety margin design, and is used to evaluate the time efficiency of the objective function of the upper-level game model.

[0054] The robustness parameters are derived from a comprehensive evaluation of ocean current disturbance intensity, illumination variation, water transparency, and target occlusion, and are used to assess the system stability of the objective function of the upper-level game model.

[0055] The power consumption parameter of the optical system is derived from the sum of the power consumption of the LED guiding light source and the power consumption of the binocular camera, and is used for the optical device energy consumption optimization of the objective function of the lower-level game model.

[0056] The processing time parameter is derived from the sum of the image processing algorithm execution time and the pose calculation time, and is used to optimize the computational efficiency of the objective function of the lower-level game model.

[0057] The iteration count parameter is derived from the iteration count of the guiding light source recognition algorithm and the pose optimization algorithm, and is used to control the algorithm complexity of the objective function of the lower-level game model.

[0058] The power consumption parameters are derived from the power consumption of the microcomputer processor and the operating power consumption of the software processing system, and are used for power consumption optimization of the computing equipment for the objective function of the lower-level game model.

[0059] The image quality parameters are derived from a comprehensive evaluation of image contrast, sharpness, and signal-to-noise ratio, and are used to ensure the imaging quality of the objective function of the lower-level game model and improve the robustness of the algorithm.

[0060] The coupling parameter represents the mutual influence between the upper-level game model and the lower-level game model. It is derived from the trade-off analysis between the recovery accuracy requirement and energy consumption limit and is used to coordinate the optimization objectives of the two game models.

[0061] The specific implementation methods of the above steps are described in detail below.

[0062] The specific implementation of step S01 involves completing the system initialization and modeling process through multiple sub-steps. First, four blue-green LED guide light sources are installed on the front face of the recovery cage in a non-uniform ring distribution. The blue-green wavelengths are selected within the range of 470–520 nm to obtain optimal transmission characteristics in seawater. The four light sources are installed in an irregular quadrilateral layout to ensure unique identification, and the spacing between adjacent light sources is controlled within the range of 0.5–1.5 m. Next, an industrial photogrammetry system is used to perform three-dimensional coordinate measurements on the guide light sources. This system, based on the principle of stereo vision, uses multiple calibration cameras to capture images of the light source positions from different angles. The three-dimensional coordinates of the spatial points are calculated using the principle of triangulation, achieving sub-millimeter level measurement accuracy. The measurement results are stored in the software processing system configuration file in Cartesian coordinate system form. Then, the intrinsic and extrinsic parameters of the binocular camera installed on the bow of the unmanned underwater vehicle are calibrated. The intrinsic parameter calibration includes determining the focal length, principal point coordinates, and distortion coefficients. The extrinsic parameter calibration includes determining the relative pose relationship between the two cameras. The calibration process is completed using the Zhang Zhengyou calibration method by capturing images of the calibration board from multiple angles. Finally, an upper-level game model is established with the goal of maximizing the success rate of recovery. This model comprehensively considers multiple factors such as positioning accuracy, recovery time, energy consumption control, system robustness and coupling constraints. The optimal system parameter configuration is found through mathematical optimization methods. In the model, the weight coefficient α for positioning accuracy is set to 0.4, the weight coefficient β for time efficiency is set to 0.3, the weight coefficient γ for energy consumption is set to 0.2, and the weight coefficient δ for robustness is set to 0.1.

[0063] The specific implementation of step S02 involves target acquisition through the coordinated operation of acoustic guidance and optical detection. First, the unmanned underwater vehicle (UUV) is guided to a distance of 15m from the recovery cage using an acoustic guidance system. Acoustic guidance employs ultra-short baseline positioning technology combined with an inertial navigation system for coarse positioning, with an accuracy controlled within the range of 1-3m. During guidance, position information and control commands are transmitted in real-time via an acoustic communication link. Once the UUV enters the optical guidance range, the binocular camera image acquisition system is activated. The binocular camera's frame rate is set to 120fps to meet real-time requirements, the image resolution is 1280×1024 pixels, and the initial exposure time is set to 25ms. Simultaneously, a guidance light source detection algorithm is activated during image acquisition. This algorithm uses a target detection method based on color and shape features, analyzing the distribution characteristics of blue-green light spots in the image to determine whether a guidance light source has been detected. If the guiding light source is successfully detected, a lower-level game model is initiated with the goal of minimizing system energy consumption. This model mainly optimizes parameters such as optical system power consumption, processing time, computational power consumption, and image quality. Energy consumption is minimized by dynamically adjusting the working state of each subsystem. The weight coefficient μ for optical system power consumption is set to 0.35, the weight coefficient ν for processing time is set to 0.25, the weight coefficient ω for computational power consumption is set to 0.25, and the weight coefficient η for image quality is set to 0.15.

[0064] The specific implementation of step S03 is to achieve the best imaging effect through adaptive exposure control. First, the binocular camera exposure value is initialized to 25ms, an initial value based on empirical settings of underwater lighting conditions and LED light source brightness. Then, an image is acquired under the current exposure parameters, and the guiding light source target is segmented. Image segmentation uses a method based on color space transformation and thresholding, converting the RGB image to the HSV color space and extracting the target using the blue-green channel information. Next, the pixel size of the guiding light source target is calculated. Pixel size calculation is achieved by counting the number of pixels in the segmented light source region, while considering the integrity of the light source shape and the clarity of its boundaries. The exposure value is adjusted according to the pixel size. If the pixel size is less than 25 pixels, it indicates that the light source is too dark and the exposure time needs to be increased; in this case, the exposure value is adjusted to be multiplied by 1.20. If the pixel size is greater than 80 pixels, it indicates that the light source is too bright and may cause saturation; in this case, the exposure value is adjusted to be multiplied by 0.75. An adaptive exposure control mathematical model is established through multiple adjustments. This model describes the nonlinear relationship between the exposure value and the pixel size of the light source. The model parameters are obtained by least squares fitting, with a fitting accuracy requirement of over 95%. Ultimately, the optimal exposure parameters are automatically set under different distances and lighting conditions.

[0065] The specific implementation of step S04 involves accurate extraction of the light source target through image preprocessing and mask generation. First, a binocular camera is used to synchronously acquire images of the guiding light source, with the synchronization error controlled within 10 ns to ensure accurate stereo matching. Hardware triggering is used during image acquisition to ensure the timing consistency of the two cameras. Then, the Rolling Ball method is used to remove background light interference. This method, based on morphological operation principles, separates the foreground target and background region by simulating the rolling of a ball on the image surface. The rolling ball radius parameter is set to 1.5–2.0 times the expected size of the light source, typically ranging from 30 to 50 pixels. Next, automatic contrast enhancement processing is performed. This processing optimizes global contrast by calculating the minimum and maximum gray values ​​of the image. The enhanced image gray values ​​are mapped to the range of 0–255 through a linear transformation, maintaining the relative gray value relationship of the image. Then, a local maximum mask image is generated. This mask is created by searching for local maxima points. For each pixel, it is checked whether it is the maximum value within an 8-connected region. If the condition is met, it is marked as 1 in the mask; otherwise, it is marked as 0. Finally, a light spot contour mask is generated. This mask is constructed based on the region growing method. The growth starts with the maximum point as the seed point. If the gray difference between the neighboring pixel and the seed point is less than a set threshold, it is added to the growth region. The threshold is set to 15% of the gray value of the seed point. The growth process continues until there are no more pixels that meet the conditions.

[0066] The specific implementation of step S05 is to achieve accurate identification of the guide light source through contour analysis and optimized matching. First, statistical information is calculated for the identified light source contour, including parameters such as contour center coordinates, contour area, average contour grayscale, ellipse fitting major axis length, and ellipse fitting minor axis length. The contour center coordinates are calculated using geometric moments, the contour area is obtained by counting pixels, and the average grayscale is calculated by the arithmetic mean of the grayscale values ​​of the pixels within the contour. Then, effective guide light source targets are screened. Screening conditions include a contour area greater than 30 pixels, an average contour grayscale greater than 200, and an ellipse fitting major-minor axis ratio less than 3. These conditions are determined based on the physical characteristics and imaging features of the LED light source. Next, the effectiveness of the four light sources is verified through convex constraint configuration judgment. The convex constraint configuration requires the four light source points to form a convex quadrilateral. This judgment is achieved by calculating the convex hull of the four points and verifying its consistency with the origin set. Then, the effectiveness of the light source targets is further verified through ellipse fitting major-minor axis ratio screening. The major-minor axis ratio threshold is set to 3.0; targets exceeding this threshold are considered too elongated and are excluded. Finally, the four guiding light sources are constructed as nodes of the Traveling Salesman Problem, and the shortest loop is solved using the nearest neighbor heuristic algorithm. This algorithm starts from any node and selects the nearest unvisited node to the current node as the next visited node. By constructing the shortest loop, the uniqueness of the guiding light source numbers is achieved. The matching process considers the geometric relationship and spatial distribution characteristics between the light sources.

[0067] The specific implementation of step S06 involves achieving three-dimensional coordinate measurement through precise positioning and stereo matching. First, the centroid method is used to precisely locate the center of the guide light source. This method determines the center position by calculating the weighted average of the coordinates of all pixels within the light source area. The weights are pixel grayscale values, and the calculation process considers the brightness distribution characteristics of the light source to improve positioning accuracy, achieving sub-pixel level accuracy. Then, the center coordinates of each guide light source are recorded, and a coordinate index table is established. The coordinate recording accuracy is maintained to three decimal places to meet subsequent calculation requirements. Next, the epipolar constraint principle is used to complete the matching of the center points of the guide light sources in the binocular images. The epipolar constraint, based on epipolar geometry theory, constrains points in the left image to the corresponding epipolar lines in the right image, significantly narrowing the search range and improving matching accuracy. The matching process uses a feature correlation-based method, with a correlation coefficient threshold set above 0.8. Then, the three-dimensional coordinates of the center of the guiding light source are calculated based on the principle of binocular vision measurement. This principle utilizes the geometric relationship of two cameras observing the same target from different perspectives, and solves the three-dimensional position of the spatial point through triangulation. The calculation process requires the use of pre-calibrated intrinsic and extrinsic parameters of the binocular camera, including focal length, principal point coordinates, distortion coefficient, and relative pose relationship between the cameras. The accuracy of the three-dimensional coordinate calculation is affected by the combined effect of camera calibration accuracy and image measurement accuracy, and can usually achieve a measurement accuracy at the millimeter level.

[0068] The specific implementation of step S07 involves achieving accurate calculation of relative pose relationships through coordinate system transformation and pose optimization. First, based on the pre-measured 3D coordinate information of the guiding light source and the calculated 3D coordinates of the guiding light source in the binocular coordinate system, two corresponding 3D point sets are constructed. Each point set contains four corresponding light source positions, and each point includes X, Y, and Z coordinate components. Then, the centroids of the two sets of 3D point sets are defined, and the centroid-free coordinates are calculated. The centroid calculation is achieved by the arithmetic mean of all point coordinates, and the centroid-free coordinates are obtained by subtracting the corresponding centroid coordinates from the original coordinates. The purpose of this step is to eliminate the influence of translation components to facilitate the solution of the rotation matrix. Next, a covariance matrix for pose calculation is constructed. This matrix is ​​formed by the summation of the outer coordinates of the two sets of centroid-free coordinates, and has a dimension of 3×3, containing rotation information between the two coordinate systems. Then, the covariance matrix is ​​decomposed using the singular value decomposition (SVD) method. SVD represents the matrix as a product of three matrices, including two orthogonal matrices and one diagonal matrix. The decomposition process is implemented using numerical calculation methods, with a required accuracy at the double-precision floating-point level. Finally, the rotation matrix and translation vector are solved based on the singular value decomposition results. The rotation matrix is ​​obtained by multiplying two orthogonal matrices, and the translation vector is calculated by the linear relationship between the centroid coordinates and the rotation matrix. The calculation results represent the relative pose relationship between the unmanned underwater vehicle and the recovery cage, including three rotation angles and three translation components. The pose calculation accuracy is affected by the accuracy of point coordinate measurement and the numerical stability of the algorithm. It can usually achieve an accuracy level of less than 0.1 degrees for angle error and less than 5 mm for position error.

[0069] The key technical ideas of this invention are mainly reflected in the following aspects. First, the optimization strategy of the two-layer game model: by constructing an upper-layer game model with the goal of maximizing the recovery success rate, and a lower-layer game model with the goal of minimizing system energy consumption, coordinated optimization of recovery performance and energy consumption control is achieved. Compared with traditional single-objective optimization methods, the two-layer game model can significantly reduce system energy consumption while ensuring the recovery success rate, and at the same time improve the system's adaptability and robustness. Second, the fusion technology of adaptive exposure control and traveling salesman problem solution: adaptive exposure control is achieved by establishing a mathematical relationship model between exposure value and light source pixel size, and unique identification matching of guide light sources is achieved by combining the shortest loop solution of the traveling salesman problem. Compared with traditional fixed exposure and simple matching methods, this technology can maintain stable recognition performance under different distances and lighting conditions, greatly improving the accuracy and reliability of light source identification. Thirdly, the synergistic processing technique of Rolling Ball background removal and region growing is employed. The Rolling Ball method effectively separates the foreground target from unevenly illuminated background areas, while the region growing method achieves precise extraction of the light spot contour. Compared to traditional threshold segmentation methods, this synergistic processing technique is better adapted to complex underwater lighting environments, significantly improving the robustness and accuracy of light source target detection. The synergistic effect of these key technological approaches forms a complete optical guidance system solution. Through the organic combination of multi-level optimization, adaptive control, and intelligent processing, it achieves high-precision, high-reliability, and low-energy-consumption target recovery for deep-sea unmanned underwater vehicles, demonstrating significant comprehensive performance advantages and practical value compared to existing technologies.

[0070] It should be noted that this invention also solves the following technical problem: the unstable accuracy of guide light source recognition in complex deep-sea environments. Traditional optical guidance methods often experience significant fluctuations in target recognition accuracy when facing complex conditions such as turbid water, uneven illumination, and turbulent currents in the deep sea, affecting the reliability of the entire recovery system. This invention effectively separates foreground targets from unevenly illuminated background areas using Rolling Ball background removal technology, improves image visual quality by combining it with automatic contrast enhancement processing technology, and achieves accurate target segmentation by establishing a maximum value mask image and a spot contour mask. It ensures the geometric effectiveness of the identified target through convex constraint configuration judgment and ellipse fitting major-minor axis ratio screening mechanism, and solves the combined optimization problem of guide light source number matching using traveling salesman problem modeling and nearest neighbor heuristic algorithms. Therefore, this invention significantly improves the accuracy and stability of guide light source recognition in complex marine environments. Furthermore, this invention also solves the technical problem of balancing energy consumption control and positioning accuracy in unmanned underwater vehicle recovery systems. In actual deep-sea recovery operations, improving positioning accuracy typically requires increasing light source power, extending exposure time, and increasing computational complexity. These measures significantly increase system energy consumption, and under the constraint of limited battery capacity, excessive energy consumption can affect mission execution time and system reliability. This invention constructs a two-layer game theory model architecture. The upper-layer model comprehensively optimizes macroscopic indicators such as positioning accuracy, recovery time, and system robustness, while the lower-layer model focuses on microscopic energy consumption control such as optical system power consumption, processing time, and computational power consumption. The two layers achieve dynamic coordination through coupling parameters, maximizing system energy efficiency while ensuring recovery accuracy requirements. Simultaneously, adaptive exposure control technology dynamically adjusts camera parameters based on the pixel size of the guide light source, avoiding energy waste caused by fixed parameter settings, and achieving an optimal balance between positioning accuracy and energy consumption control.

[0071] Specifically, the principle of this invention is as follows: The fundamental reason why this invention can solve the problems of insufficient accuracy and poor real-time performance of optical guidance and positioning in deep-sea environments lies in the adoption of a multi-level collaborative optimization technical architecture and a precise visual measurement method. First, the selection of blue-green LED guidance light sources is based on the optical characteristic that blue-green light wavelengths have good water penetration capabilities in deep-sea environments. The non-uniform ring distribution design of the four guidance light sources ensures that sufficient spatial constraint information can be provided at different observation angles, avoiding the fuzzy solution problem caused by traditional single-point or symmetrical distribution guidance methods under certain attitudes. The design logic of the two-layer game model lies in decomposing the retrieval problem into two levels: macro-strategy optimization and micro-execution optimization. The upper-layer game model aims to maximize the retrieval success rate, comprehensively considering multiple constraints such as positioning accuracy, retrieval time, energy consumption control, and system robustness. The lower-layer game model aims to minimize system energy consumption, focusing on the optimization of execution-level aspects such as optical system power consumption, processing time, and computational complexity. The two-layer models achieve coordination and cooperation through coupling parameters, ensuring that the overall system achieves an optimal balance between accuracy and efficiency. Adaptive exposure control technology addresses the impact of varying lighting conditions on image quality in the deep-sea environment by real-time monitoring of the guide light source pixel size and dynamically adjusting camera exposure parameters. RollingBall background removal and automatic contrast enhancement technologies effectively separate the guide light source from the complex seabed background, improving target recognition accuracy. The centroid-based positioning method combined with epipolar constraint binocular stereo matching technology achieves sub-pixel-level center positioning accuracy for the light source. Based on pre-calibrated 3D coordinates of the guide light source and real-time calculated binocular coordinates, the rigid body transformation matrix obtained through singular value decomposition accurately describes the relative pose between the vehicle and the recovery cage, providing high-precision spatial positioning information for the real-time control system. This enables high-precision, real-time optical guidance and recovery.

[0072] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0073] In this embodiment, the specific implementation of steps S01-S02 is the same as described above, and will not be repeated in detail here.

[0074] The specific implementation of step S03 is to achieve the best imaging effect through an adaptive exposure control algorithm, and the exposure value adjustment rule is expressed as follows:

[0075] When the target pixel size of the light source At that time, the exposure value adjustment formula is:

[0076] ;

[0077] When the target pixel size of the light source At that time, the exposure value adjustment formula is:

[0078] ;

[0079] In the formula, The adjusted exposure value is in milliseconds (ms). The exposure value before adjustment, in milliseconds (ms). The pixel size of the light source target, expressed in pixels.

[0080] The parameter acquisition method is as follows: The initial value was set to 25ms and obtained by controlling the camera's exposure time. The number of pixels in the light source region is calculated using an image segmentation algorithm, including: Step 1: Converting the acquired image from RGB to HSV color space; Step 2: Using blue-green channel information for threshold segmentation to extract the target light source region; Step 3: Counting the number of pixels in the segmented light source region. value.

[0081] The specific implementation of step S04 is to achieve accurate extraction of the light source target through image enhancement algorithm and background removal algorithm. The mathematical expression of its automatic contrast enhancement processing is as follows:

[0082] ;

[0083] In the formula, The grayscale value of the image after automatic contrast enhancement; The grayscale values ​​of the image after processing by the Rolling Ball method; The minimum grayscale value of the image; The maximum grayscale value of the image; These are the pixel coordinates. Represents the horizontal coordinate. Represents the vertical coordinate.

[0084] The parameter acquisition method is as follows: Background was removed using a Rolling Ball background removal algorithm, which employs morphological manipulation principles and sets the rolling ball radius to 1.5 times the expected size of the light source. and By traversal Obtain all pixels of the image. , .

[0085] The specific implementation of step S05 is to achieve accurate identification and matching of the guiding light source through contour analysis and combination optimization. The mathematical expression for the generation of candidate light spot combinations is as follows:

[0086] ;

[0087] In the formula, The number of candidate spot combinations generated; The number of candidate light spots; From The number of combinations of selecting 4 from 3 candidate light spots.

[0088] The formula for calculating the center coordinates of the contour is:

[0089] , ;

[0090] The final pixel grayscale and calculation formula for the selected light spot combination are as follows:

[0091] ;

[0092] The formula for calculating the distance between guide light sources is:

[0093] ;

[0094] In the formula, and These are the x and y coordinates of the center of the contour, respectively, in pixels; This represents the area of ​​the outline, which is the number of pixels within the outline. and For the first within the outline The coordinates of each pixel; The sum of the total pixel grayscale values ​​of the light spot combination; For the first The outline area of ​​each light spot; For the first The average gray level of the outline of each light spot; For the first The first light source and the first The Euclidean distance between the centers of each light source, in pixels; , , , The first The and the first The coordinates of the center of each light source.

[0095] The parameter acquisition method is as follows: The selection criteria obtained through the aforementioned screening steps include an outline area greater than 30 pixels, an average outline gray level greater than 200, and an ellipse fitting major-minor axis ratio less than 3. It is obtained by counting the number of pixels marked as 1 in the contour mask; Obtain the coordinates of all pixels within the contour by traversing the contour mask; and It is obtained by calculating through contour statistics, where .

[0096] The specific implementation of step S06 is to achieve three-dimensional coordinate measurement of the guiding light source through the principle of binocular stereo vision. The mathematical expression of the center positioning using the centroid method is as follows:

[0097] , ;

[0098] In the formula, and These are the precise coordinates of the center of the light source, in pixels. This represents the total number of pixels within the light source area; and The first in the light source area The coordinates of each pixel; For pixels The grayscale value at that location is used as a weighting coefficient.

[0099] The parameter acquisition method is as follows: This is obtained by counting the number of pixels in the mask of the light source area; Read the grayscale value of the corresponding pixel directly from the enhanced image; The coordinates of all pixels within the light source region are obtained by traversing the mask of the light source region.

[0100] The specific implementation of step S07 is to calculate the relative pose relationship between the unmanned underwater vehicle and the recovery cage using a 3D-3D pose estimation algorithm. The mathematical model is established as follows:

[0101] Let the matched 3D point set in the two coordinate systems be:

[0102] , ;

[0103] The pose transformation relationship is as follows:

[0104] ;

[0105] In the formula, The set of three-dimensional coordinate points of the guiding light source in the coordinate system of the recovery cage; This is the set of three-dimensional coordinate points of the guiding light source in the camera coordinate system; For the first time in the coordinate system of the recovery cage The three-dimensional coordinate vector of a guiding light source; In the camera coordinate system, the first The three-dimensional coordinate vector of a guiding light source; It is a 3×3 rotation matrix; It is a three-dimensional translation vector; To guide the number of light sources, in this invention .

[0106] Definition of the first The error term for the point is:

[0107] ;

[0108] In the formula, For the first The pose transformation error vector for the matching point.

[0109] The least squares optimization problem is expressed as:

[0110] ;

[0111] In the formula, The objective function is... This represents the 2-norm of a vector.

[0112] Define the centroids of the two sets of points as:

[0113] , ;

[0114] In the formula, Let the centroid of the point set in the coordinate system of the recovery cage be defined. Let be the centroid of the point set in the camera coordinate system.

[0115] Calculate the centroid coordinates:

[0116] , ;

[0117] In the formula, For the first time in the coordinate system of the recovery cage Centroid coordinates of the points; In the camera coordinate system, the first Centroid coordinates of the points.

[0118] Construct the covariance matrix:

[0119] ;

[0120] In the formula, It is a 3×3 covariance matrix that contains rotation information between the two coordinate systems.

[0121] For matrix Perform singular value decomposition:

[0122] ;

[0123] In the formula, and It is a 3×3 orthogonal matrix; It is a 3×3 diagonal matrix containing singular values.

[0124] when When the value is full rank, the rotation matrix is:

[0125] ;

[0126] The translation vector is:

[0127] .

[0128] The parameter acquisition method is as follows: The three-dimensional spatial coordinates of the guide light source in the recovery cage are obtained by measuring the industrial photogrammetry system. The calculations are performed based on the principle of binocular stereo vision measurement, requiring the use of pre-calibrated intrinsic and extrinsic parameters of a binocular camera; singular value decomposition is achieved using numerical calculation methods, requiring a calculation accuracy at the double-precision floating-point level.

[0129] The principles and effects of the above formulas are analyzed as follows: Adaptive adjustment formula for exposure value and By establishing a mathematical relationship between exposure parameters and light source pixel size, optimal imaging results were achieved under different distances and lighting conditions. Compared to traditional fixed exposure methods, this significantly improved image quality and the stability of light source recognition. The formula employs a piecewise linear adjustment strategy to effectively avoid overexposure and underexposure, enhancing the system's adaptability to complex marine environments. Automatic contrast enhancement formula. By mapping image grayscale values ​​to the full dynamic range through linear transformation, the image's contrast and detail information are maximized. Compared to traditional global thresholding methods, this approach better handles uneven illumination. The formula maintains the relative grayscale relationships of the image, ensuring the stability and accuracy of subsequent image processing algorithms. Combinatorial number calculation formula. This provides a mathematical foundation for the unique identification of guided light sources. By enumerating all possible combinations of four light sources and combining them with geometric constraints, accurate light source matching in complex environments is achieved, exhibiting higher reliability and robustness compared to traditional heuristic matching methods. Pixel grayscale and formula The quality of light spot combinations is evaluated by comprehensively considering the product of the light spot area and the average gray level. Compared with single feature screening methods, this approach can more accurately identify genuine guide light source combinations. The area term in this formula... Reflecting the geometric characteristics of the light source, the average grayscale term Reflecting the brightness characteristics of the light source, the combination of these two factors improves the accuracy of light source identification and its anti-interference capability. Distance calculation formula. The Euclidean distance is used to measure the spatial relationship between light sources, providing geometric constraints for guiding light source numbering and matching. This formula achieves unique numbering by finding the shortest distance between adjacent light source pairs, exhibiting better numerical stability and computational efficiency compared to traditional angle or topological matching methods. (Center of gravity method center positioning formula) and A gray-scale weighted centroid calculation method is employed, fully utilizing the brightness distribution information of the light source to achieve sub-pixel-level positioning accuracy. Compared to traditional geometric center calculation methods, this significantly improves positioning accuracy. The formula effectively suppresses the influence of noise and background interference through a weighting mechanism. The 3D-3D pose estimation formula set is optimized using least squares. Singular Value Decomposition This method transforms the complex pose estimation problem into a linear algebra problem, achieving high-precision relative pose calculation. Compared with traditional iterative optimization methods, it exhibits better numerical stability and computational efficiency. This mathematical model utilizes centroid removal operations... and and covariance matrix construction This method effectively separates rotation and translation components, simplifies the solution process, and improves the robustness of the algorithm. The singular value decomposition method guarantees the rotation matrix... The orthogonality constraint ensures the physical meaning and mathematical correctness of the pose estimation results, and the translation vector... The calculation utilizes the linear properties of the centroid transformation, avoiding a complex nonlinear optimization process.

[0130] To better understand and implement this invention, a specific application scenario is provided as Example 2: A marine technology team, during a deep-sea resource exploration mission, needed to accurately recover an autonomous underwater vehicle (AUV) operating at a depth of 1800m. Traditional acoustic-guided recovery methods suffer from insufficient positioning accuracy and signal interference in complex seabed environments, resulting in a recovery success rate of only 67% and a long recovery time, averaging 42 minutes. To solve this technical challenge, the team decided to improve the recovery system using the optical guidance method of this invention. Figure 2 The overall scheme of the optical guidance system shown includes the overall layout and connection relationship of core components such as the guidance light source, binocular camera, electronic control system and software processing system.

[0131] Specifically, the guidance light source is mounted on the recovery cage, while the binocular camera, electronic control system, and software processing system are all located within the AUV's sealed control compartment. During operation, the guidance light source should be within the common field of view of the binocular camera. The electronic control system executes hardware triggers, causing the binocular camera to acquire image data of the guidance light source target. This image data is then transmitted to a microcomputer via a data transmission line. Subsequently, the computer processing software calculates the three-dimensional measurement results and relative positioning information of the target, which are finally uploaded to the AUV's control system.

[0132] Four circular LED guide light sources 1 are installed on the AUV recovery cage. They are non-uniformly distributed in a ring on the front face of the recovery cage 2. The guide light sources consist of four light sources 1 (including light source 1, light source 2, light source 3, and light source 4) and a light source control and drive circuit 3. They are pressure-resistant packaged, with the body made of TC4 material and the window glass made of sapphire. The structural sealing method adopts O-ring sealing, and the pressure resistance depth is designed to be 2000m.

[0133] The binocular camera 4 is installed in the bow dry compartment of the AUV. The mounting interface at the bottom of the camera has a positioning slot and positioning, and the positioning connection hole with the AUV is achieved through an adapter plate. Figure 2 A 520nm single bandpass filter is installed in front of the lens to reduce the impact of stray light on the image quality of the light source. The camera's frame rate is 1280×1024@120fps, which meets the requirement of attitude data ≥10Hz. The binocular camera has a gigabit network interface and features a hard synchronization trigger mode. The synchronization trigger pulse error between the two cameras is extremely small, much smaller than the exposure time (10ns).

[0134] The electronic control system functions to drive and control the cooperative light source, control camera synchronization triggering, and acquire and process image data. The microcomputer 5, power supply system 6, and triggering system 7 are enclosed in a metal box and connected to the outside via a 5-core power cable 8. Three cores are used for RS485 serial communication to ensure timely data transmission to the AUV for appropriate attitude adjustments, and two cores are used for power supply.

[0135] The software processing system includes cooperative light source target detection, localization and matching, 3D reconstruction of the target light source, and relative pose calculation for the two images obtained by the binocular camera. Data output is mainly achieved through the serial port, including steps such as opening the serial port, configuring the serial port, reading and writing the serial port, and closing the serial port.

[0136] The implementation process begins with system deployment and initial configuration. The technical team installed four blue-green LED guide lights on the front face of the recovery cage using a non-uniform ring distribution. The light wavelength was chosen to be 485nm to obtain optimal seawater transmission characteristics. The four lights were installed in an irregular quadrilateral layout: Light source 1 was located at the upper left center of the front face of the recovery cage, with coordinates (-0.6, 0.8, 0); Light source 2 was located at the upper right, with coordinates (0.7, 0.9, 0); Light source 3 was located at the lower right, with coordinates (0.8, -0.7, 0); and Light source 4 was located at the lower left, with coordinates (-0.5, -0.8, 0). The units are in meters. After the light source installation was completed, an industrial photogrammetry system was used to accurately measure the three-dimensional coordinates of the guide lights, achieving a measurement accuracy of 0.3mm. The measurement results were stored in the software processing system configuration file. Subsequently, the internal and external parameters of the binocular camera installed on the bow of the unmanned underwater vehicle were calibrated, such as... Figure 3 The camera mounting structure shown achieves precise positioning connection with the vehicle via an adapter plate. The binocular camera has a baseline distance of 120mm, a focal length of 8mm, an image resolution of 1280×1024 pixels, and a frame rate of 120fps. The calibration process uses the Zhang Zhengyou calibration method, determining parameters through 15 calibration plate images from different angles. The calibration accuracy verification results show a reprojection error of less than 0.4 pixels. When establishing the upper-level game theory model, the positioning accuracy weight coefficient α is set to 0.4, the time efficiency weight coefficient β is set to 0.3, the energy consumption weight coefficient γ is set to 0.2, and the robustness weight coefficient δ is set to 0.1. The model convergence accuracy requirement is [not specified]. .

[0137] In the acoustic guidance phase, ultra-short baseline positioning technology was used to guide the unmanned underwater vehicle (UUV) to within 15m of the recovery cage. At approximately 18m from the recovery cage, the acoustic guidance system achieved a positioning accuracy of 2.3m, meeting the activation conditions for optical guidance. Once the UUV entered the optical guidance operating range, the binocular camera image acquisition system was immediately activated, simultaneously initiating the guidance light source detection algorithm. Initial image acquisition detected three suspected light source targets, triggering the lower-level game theory model. The optical system power consumption weight coefficient μ was set to 0.35, the processing time weight coefficient ν to 0.25, the computational power consumption weight coefficient ω to 0.25, and the image quality weight coefficient η to 0.15. Algorithm analysis showed that two targets met the geometric and brightness feature requirements, while one target was excluded because its ellipse fitting major-minor axis ratio exceeded 3.2.

[0138] The adaptive exposure control phase first initializes the stereo camera exposure values. The system acquires an image under the current parameters and performs guided light source target segmentation. It converts the RGB image to the HSV color space using color space transformation, extracts the light source target region using threshold segmentation based on blue-green channel information, and counts the number of pixels in the segmented light source region to obtain the pixel size. These are 18, 22, 156, and 178 pixels respectively. Based on the pixel size determination criteria, when... At this time, the first two light source targets need to have their exposure time increased, according to the formula. Calculate and adjust the exposure value to ms. When At that time, the exposure time for the latter two light source targets needs to be reduced to avoid oversaturation, according to the formula. Calculate and adjust the exposure value to After three iterations of adjustment, the pixel sizes of the four light source targets stabilized within the range of 32, 38, 67, and 71 pixels, meeting the optimal imaging requirements. The established adaptive exposure control mathematical model achieved a fitting accuracy of 97.3%, realizing the automatic setting of optimal exposure parameters under different distance conditions.

[0139] In the image preprocessing and target extraction stages, images of the guide light source were simultaneously acquired using a binocular camera, with the synchronization error controlled within 8 ns. The Rolling Ball method was employed to remove background light interference, with the rolling ball radius set to 35 pixels, effectively separating the foreground light source target from the scattered light from the seawater background. Automatic contrast enhancement was performed to calculate the minimum grayscale value of the image. Maximum grayscale value Through the linear transformation formula Achieve global contrast optimization, where These are the pixel coordinates. The image grayscale values ​​after processing by the Rolling Ball method. The image grayscale values ​​are those after automatic contrast enhancement. During the generation of the local maxima mask image, 67 local maxima points were identified through 8-connected region search. A light spot contour mask was generated based on the region growing method, using the maxima points as seed points. The grayscale threshold was set to 15% of the seed point's grayscale value, ultimately extracting 8 candidate light source contours.

[0140] In the guided light source identification and matching stage, statistical information is calculated for the extracted contours of the eight candidate light sources. The entire light source identification process strictly follows... Figure 5 The light source identification flowchart shown is executed. The contour statistics results shown in Table 1 display the key parameters of each candidate light source.

[0141] Table 1. Statistical information on candidate light source profiles

[0142] After verification based on the screening criteria, five candidate light sources were identified: 1, 2, 3, 4, and 7. These candidates had an outline area greater than 30 pixels, an average grayscale value greater than 200, and an aspect ratio less than 3. Candidate 5 was eliminated due to insufficient area, candidate 6 due to an excessive aspect ratio, and candidate 8 due to both insufficient area and grayscale value. Four of these five valid candidate light sources were selected to form a combination, resulting in a total of [number missing] candidate combinations. The coordinates of the contour center are calculated using a formula, taking light source number 1 as an example: , ,in Let be the contour area. Based on the convex constraint configuration, all five combinations constitute a valid convex quadrilateral. Calculate the pixel grayscale sum for each combination; the grayscale sum for combination {1,2,3,4} is calculated using the formula... Calculated as The value that is the maximum is selected as the final light spot combination. A nearest neighbor heuristic algorithm is used to match the guide light source numbers and calculate the distance between each light source. The distance between light source 1 and light source 4 is calculated according to the formula... Calculated as The minimum pixel value is used to identify adjacent light source pairs, and unique number matching is completed by sorting them counterclockwise.

[0143] In the precise positioning and 3D coordinate measurement stage, the centroid method is used to achieve precise positioning of the center of the guiding light source. Taking light source No. 1 as an example, the total number of pixels in the light source area... Calculated using the gray-scale weighted formula and The precise center coordinates are obtained as follows , ,in and The first in the light source area The coordinates of each pixel Using the grayscale value at each pixel as a weighting coefficient, sub-pixel-level positioning accuracy was achieved. Epipolar constraint principle was used to perform same-name matching of the center points of the guide light sources in the binocular images; the matching correlation coefficients all exceeded 0.85, meeting the matching reliability requirements. Based on the binocular vision measurement principle, the three-dimensional coordinates of the centers of each guide light source were calculated. The measurement results shown in Table 2 display the light source position information in the camera coordinate system.

[0144] Table 2. Three-dimensional coordinate measurement results of the guiding light source

[0145] In the relative pose relationship calculation stage, two sets of corresponding points are constructed based on the pre-measured 3D coordinates of the guide light source and the calculated light source coordinates in the camera coordinate system. Let the 3D coordinate point set of the guide light source in the retrieval cage coordinate system be... The set of three-dimensional coordinate points of the guiding light source in the camera coordinate system is The pose transformation relationship is as follows ,in It is a 3×3 rotation matrix. Let be a three-dimensional translation vector. Define the first... The error term for the point is Establish the least squares optimization problem Calculate the centroids of the two sets of point sets as follows: and ,in The number of guiding light sources is calculated through a centroid removal operation. and Eliminate the influence of translation components. After expanding and simplifying the error function, the least squares problem becomes... Pose calculation consists of three steps: first, solving the optimization problem. Expand on The error term is obtained .because The objective function simplifies to Construct the covariance matrix And perform singular value decomposition. The singular value decomposition results show that the singular values ​​are 2847.3, 2651.8, and 2398.7, all positive and the matrix is ​​of full rank, satisfying the solution conditions. When When the rotation matrix is ​​full rank Translation vector The calculated Euler angles of the rotation matrix are: pitch -2.3°, yaw 1.7°, roll -0.8°, and the translation vector is... mm. The pose calculation accuracy verification results show that the angle error is less than 0.08° and the position error is less than 3.2mm, which meets the requirements for high-precision recovery guidance.

[0146] The performance parameters of the entire optical guidance process are shown in Table 3, which demonstrates the technical specifications of the system in practical applications.

[0147] Table 3 Performance parameters of the optical guidance system

[0148] After 15 consecutive recovery tests, the entire optical guidance process strictly followed the guidelines. Figure 4The flowchart of the optical guidance method shown demonstrates the significant technical advantages of the recovery system employing this invention's optical guidance method in complex marine environments. Compared to traditional acoustic guidance recovery methods, the light source recognition accuracy has increased from 82.1% to 96.3%, primarily due to the synergistic effect of adaptive exposure control and Rolling Ball background removal technology, effectively adapting to imaging requirements under different distances and lighting conditions. Pose calculation accuracy has improved from 15.6 mm to 3.2 mm, a 79.5% increase. This significant improvement stems from the application of centroid-based subpixel localization technology and singular value decomposition pose estimation algorithm, achieving millimeter-level relative pose measurement accuracy. The total recovery time has been reduced from 42.0 minutes to 28.5 minutes, an efficiency improvement of 32.1%, mainly because the two-layer game model optimizes system parameter configuration, reducing search and adjustment time. System energy consumption has decreased from 3.4 W·h to 2.8 W·h, achieving an energy saving effect of 17.6%. Through the energy consumption optimization strategy of the lower-layer game model, the power consumption of the optical system and computing equipment has been significantly reduced while ensuring performance. The most critical recovery success rate improved from 67.0% to 89.2%, a success rate increase of 33.1%. This significant improvement in this core indicator demonstrates the practical value and reliability of the technical solution of this invention. Traditional methods are prone to positioning failures and signal interruptions in complex seabed topography and strong ocean current environments. However, the optical guidance method of this invention, through multi-level optimized control and intelligent image processing technology, effectively overcomes environmental interference factors, achieving stable and reliable high-precision recovery guidance, and providing important technical support for the safe recovery of deep-sea unmanned underwater vehicles.

[0149] It should be noted that the variables involved in this invention are explained in detail in Table 4.

[0150] Table 4. Variable Explanation Table

[0151] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. An optical guidance method for recovering deep-sea unmanned underwater vehicles, characterized in that, Four blue-green LED guide lights are installed on the front face of the recovery cage. The three-dimensional spatial coordinate information of the guide lights is measured and stored using an industrial photogrammetry system. The internal and external parameters of the binocular camera at the bow of the unmanned underwater vehicle are calibrated, and upper-level and lower-level game models are established. The unmanned underwater vehicle is guided to a distance of 15m from the recovery cage using acoustic guidance. The binocular camera image acquisition system is activated and the guide light target is detected. An adaptive exposure control mathematical model is established, images are acquired, and the guide light target is segmented and the pixel size is calculated. Images of the guidance light source are simultaneously acquired using a binocular camera. The Rolling Ball method is employed to remove background light interference, and a maximum value mask image and a light spot contour mask are generated through automatic contrast enhancement. Statistical information is calculated on the identified light source contours. Four valid guidance light source targets are determined through convex constraint configuration judgment and ellipse fitting with major and minor axis ratios. These four targets are constructed as nodes for the Traveling Salesman Problem, and the nearest neighbor heuristic algorithm is used to match the guidance light source numbers. The centroid method is used to accurately locate the center of the guidance light source. The epipolar constraint principle is used to match the center points of the guidance light source in the binocular images. The three-dimensional coordinates of the guidance light source center are calculated based on the binocular vision measurement principle. Based on the pre-measured three-dimensional coordinates of the guidance light source and the calculated three-dimensional coordinates in the binocular coordinate system, the rotation matrix and translation vector are solved using singular value decomposition to calculate the relative pose relationship between the unmanned underwater vehicle and the recovery cage.

2. The optical guidance method for recovering deep-sea unmanned underwater vehicles according to claim 1, characterized in that, Four blue-green LED guide lights are installed in a non-uniform ring distribution. The upper-level game model aims to maximize the recovery success rate, while the lower-level game model aims to minimize system energy consumption. The lower-level game model is activated when the guide light is detected.

3. The optical guidance method for recovering deep-sea unmanned underwater vehicles according to claim 2, characterized in that, The steps of the adaptive exposure control mathematical model are as follows: initialize the stereo camera exposure value to 25ms; if the pixel size is less than 25 pixels, adjust the exposure value to the original value multiplied by 1.20; if the pixel size is greater than 80 pixels, adjust the exposure value to the original value multiplied by 0.

75.

4. The optical guidance method for recovering deep-sea unmanned underwater vehicles according to claim 3, characterized in that, The industrial photogrammetry system is a high-precision three-dimensional coordinate measurement device based on the principle of stereo vision. It takes pictures of the target with multiple calibrated cameras and calculates the three-dimensional coordinates of spatial points using the principle of triangulation, achieving a measurement accuracy of sub-millimeter level.

5. The optical guidance method for recovering deep-sea unmanned underwater vehicles according to claim 4, characterized in that, The Rolling Ball method is specifically an image background removal technique based on morphological operations. It effectively separates foreground objects from unevenly lit background areas by simulating the rolling of a ball on the image surface.

6. The optical guidance method for recovering deep-sea unmanned underwater vehicles according to claim 5, characterized in that, The automatic contrast enhancement process specifically involves calculating the minimum and maximum grayscale values ​​of the image, and then calculating the enhanced image grayscale value pixel by pixel.

7. The optical guidance method for recovering deep-sea unmanned underwater vehicles according to claim 6, characterized in that, The steps for calculating the statistical information specifically involve obtaining the center coordinates, contour area, average gray level of the contour, length of the major axis of the ellipse fitting, and length of the minor axis of the ellipse fitting.

8. The optical guidance method for recovering deep-sea unmanned underwater vehicles according to claim 7, characterized in that, The step of calculating the three-dimensional coordinates of the guiding light source in the binocular coordinate system also includes defining the centroids of two sets of three-dimensional point sets and calculating the centroid-free coordinates; the centroid-free coordinates are specifically the relative coordinates obtained by subtracting the centroid coordinates of the corresponding point set from the three-dimensional point coordinates.

9. The optical guidance method for recovering deep-sea unmanned underwater vehicles according to claim 8, characterized in that, The maximum value mask image is specifically initialized as an all-zero matrix, with the image size being the same as the original image. For each pixel, if its gray value is not less than the gray value of pixels within the eight-connected region, the pixel is considered a maximum value point and marked as 1 at the corresponding position.

10. The optical guidance method for recovering a deep-sea unmanned underwater vehicle according to claim 9, characterized in that, The translation vector is specifically a three-dimensional vector that describes the translation relationship between two coordinate systems.

Images