Method and system for precise positioning of navigation marks based on machine vision

Abstract

This invention provides a machine vision-based method and system for precise navigation mark positioning, relating to the field of computer vision technology. First, it acquires a dataset of drone-captured images of navigation marks in the surrounding water. Then, it performs specular reflection component stripping processing on the water surface to obtain a diffuse reflection substrate image and a specular reflection intensity distribution map. Next, it performs light field consistency reconstruction processing to obtain a normalized illumination water image. Then, it constructs a multi-scale structural tensor field description map of the navigation mark, representing the principal structural direction and structural saliency intensity of each pixel position of the navigation mark. Based on this description map, it extracts the topology of the navigation mark's structural skeleton, generating a topological skeleton diagram of the navigation mark's structure containing a set of top skeleton nodes and a set of supporting skeleton chains. Finally, it calls a pre-set drone-to-ground spatial intersection positioning model to perform spatial intersection calculation processing, generating precise positioning information containing the navigation mark's geospatial coordinates and elevation information, achieving high-precision and high-reliability positioning of navigation marks in complex environments.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and more specifically, to a method and system for precise positioning of navigation marks based on machine vision. Background Technology

[0002] In the field of navigation mark positioning, traditional positioning methods mainly rely on manual on-site measurement or positioning based on a single sensor (such as GPS). Manual on-site measurement not only consumes a lot of manpower, material resources, and time, but also poses significant safety risks in some complex aquatic environments, such as those with rapid currents, wide waters, or dangerous obstacles. Furthermore, the measurement accuracy is easily affected by human factors.

[0003] Single-sensor-based positioning methods, such as relying solely on GPS for beacon location, while improving convenience to some extent, suffer from several drawbacks. GPS signals are easily interfered with in complex environments, such as urban waterways with tall buildings, mountainous waterways, or when affected by weather conditions (like heavy rain or dense fog). These conditions can lead to signal attenuation, loss, or significant positioning errors, compromising the accuracy and reliability of beacon location. Furthermore, existing image-based beacon location methods lack effective preprocessing of images acquired by drones. The specular reflections from water surfaces in these images can interfere with accurate beacon identification and location, resulting in substantial deviations in the positioning results. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for precise positioning of navigational aids based on machine vision, the method comprising: A set of aerial images of navigation marks in water areas taken by drones is acquired, wherein the set of aerial images of navigation marks in water areas taken by drones includes multiple continuously acquired water surface image frame units with acquisition time stamps from the perspective of drones; The image dataset of the navigation mark in the water area taken by the UAV is subjected to water surface specular reflection component stripping processing to obtain a diffuse reflection substrate image and a specular reflection intensity distribution map of the water area. Then, the diffuse reflection substrate image and the specular reflection intensity distribution map of the water area are subjected to light field consistency reconstruction processing to obtain an illumination-normalized water area image. The illumination-normalized water area image is processed to construct a multi-scale structural tensor field for navigation marks, resulting in a navigation mark structure tensor field description map. The navigation mark structure tensor field description map represents the main structural direction and structural saliency intensity of the navigation mark at each pixel position. Based on the tensor field description graph of the navigation beacon structure, the topology of the navigation beacon structure skeleton is extracted to generate a topological skeleton graph of the navigation beacon structure. The topological skeleton graph of the navigation beacon structure includes a set of skeleton nodes at the top of the navigation beacon and a set of skeleton chains supporting the navigation beacon. A preset UAV-to-ground spatial rendezvous and positioning model is invoked to perform spatial rendezvous calculation on the set of top skeleton nodes of the navigation beacon and the set of supporting skeleton chains of the navigation beacon, thereby generating precise positioning information of the navigation beacon, which includes the geographic spatial coordinates and elevation information of the navigation beacon.

[0005] Furthermore, embodiments of the present invention also provide a machine vision-based precise navigation mark positioning system, characterized in that it includes: A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to perform the above-described machine vision-based beacon positioning method by executing the machine-executable instructions.

[0006] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions, the machine-executable instructions being stored in a computer-readable storage medium, the processor of a computer device reading the machine-executable instructions from the computer-readable storage medium, the processor executing the machine-executable instructions, causing the computer device to execute the above-described machine vision-based beacon precise positioning method.

[0007] Based on the above, by acquiring a set of navigation mark water area image data containing multiple continuously acquired UAV-view water surface image frames with acquisition time stamps, the water surface reflection component of the navigation mark water area image data set is stripped to effectively eliminate the interference of mirror reflection on navigation mark identification. Combined with light field consistency reconstruction processing, a normalized water area image is obtained, which makes the image consistent under different lighting conditions, improving the image quality and usability. A multi-scale structural tensor field description map of the navigation mark is constructed, which can accurately characterize the main structural direction and structural saliency intensity of the navigation mark body at each pixel position. Based on the navigation mark structural tensor field description map, the topology extraction processing of the navigation mark structural skeleton is performed. The generated navigation mark structural topology skeleton map contains the set of top skeleton nodes of the navigation mark and the set of support skeleton chains of the navigation mark. Finally, the preset UAV-to-ground spatial intersection positioning model is called to perform spatial intersection solution processing on the navigation mark structural topology skeleton map to generate accurate positioning information containing the navigation mark geospatial coordinates and elevation information. This achieves high-precision and high-reliability positioning of navigation marks in complex environments, effectively solving the problems of low accuracy and poor reliability of traditional navigation mark positioning methods. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the execution flow of the machine vision-based precise positioning method for navigation marks provided in an embodiment of the present invention.

[0009] Figure 2 This is a schematic diagram of exemplary hardware and software components of the machine vision-based navigation mark precision positioning system provided in an embodiment of the present invention. Detailed Implementation

[0010] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a machine vision-based method for precise positioning of navigation marks, as provided in one embodiment of the present invention. The following is a detailed description of this machine vision-based method for precise positioning of navigation marks.

[0011] Step S110: Obtain a set of drone aerial images of navigation marks in water areas, wherein the set of drone aerial images of navigation marks in water areas includes multiple drone-view water surface image frame units that are continuously acquired and have acquisition time stamps.

[0012] In a scenario involving the inspection and positioning of navigation marks in inland waterways, a drone equipped with a visible light camera flies along a preset route to capture images of the navigation marks from multiple angles. Step S110 acquires a set of aerial image data of the navigation marks and surrounding waters taken by the drone. This set of aerial image data contains multiple continuously acquired drone-view water surface image frames with acquisition time stamps. Each drone-view water surface image frame corresponds to a snapshot at a fixed time interval, recording the image of the navigation mark and its surrounding waters at that moment. The acquisition of all image data has been authorized by the relevant airspace and waterway management departments and complies with data privacy and security regulations.

[0013] Step S120: Perform water surface specular reflection component stripping processing on the drone aerial image data set of navigation mark water area to obtain a water area diffuse reflection substrate image and a specular reflection intensity distribution map, and perform light field consistency reconstruction processing on the water area diffuse reflection substrate image and the specular reflection intensity distribution map to obtain an illumination normalized water area image.

[0014] Step S120 performs water surface reflective component stripping processing on the set of drone aerial image data of navigation marks obtained in step S110.

[0015] Step S121: Perform multi-color channel separation processing on the UAV-view water surface image frame units in the UAV aerial photography navigation mark water area image data set to obtain multiple single-color channel images. Select the single-color channel image with the highest water absorption rate as the strong water absorption channel image and the single-color channel image with the lowest water absorption rate as the weak water absorption channel image. Perform dark channel prior statistical processing on the weak water absorption channel image and extract the minimum pixel intensity value in each local neighborhood window of the weak water absorption channel image as the dark channel pixel value to generate a dark channel prior image.

[0016] For each frame of the water surface image from the UAV's perspective, its red, green, and blue color channels are separated to obtain red, green, and blue channel images. Based on the spectral absorption characteristics of water, red light attenuates fastest and has the highest absorption rate in water, while blue light attenuates slowest and has the lowest absorption rate. Therefore, the red channel image is selected as the strong absorption channel image for water, and the blue channel image is selected as the weak absorption channel image. For the blue channel image, a local neighborhood window size parameter is set, with the window side length being a fixed pixel unit value. For each pixel position in the image, a local neighborhood window is taken centered on that pixel position. The blue channel intensity values ​​of all pixels within this local neighborhood window are extracted, and the minimum value is found and used as the dark channel pixel value for that pixel position. The dark channel pixel values ​​of all pixel positions are arranged according to pixel coordinates to form the dark channel prior image.

[0017] Step S122: Perform atmospheric light component estimation processing on the dark channel prior image, extract the top 0.1% of pixels with the largest pixel intensity value in the dark channel prior image, and calculate the average pixel intensity value of the top 0.1% of pixels at the corresponding position in the weak absorption channel image of the water body as the global atmospheric light intensity value.

[0018] In the dark channel prior image, all pixels are sorted in descending order of their dark channel pixel values, and the pixels in the top 0.1% of the sorted positions are selected as the candidate atmospheric light pixel location set. These candidate atmospheric light pixel locations are then mapped back to the blue channel image, and the pixel intensity values ​​at these candidate atmospheric light pixel locations are extracted. The arithmetic mean of these pixel intensity values ​​is calculated, and this arithmetic mean is used as the global atmospheric light intensity value. This global atmospheric light intensity value represents the intensity of skylight reflected from the water surface.

[0019] Step S123: Based on the global atmospheric light intensity value, perform transmittance distribution estimation processing on the strong absorption channel image and the weak absorption channel image of the water body to obtain a water body transmittance distribution map. The transmittance value of each pixel in the water body transmittance distribution map represents the proportion of light transmitted by the water body at that pixel location.

[0020] For each pixel location in the strong absorption channel image of the water body, the initial transmittance value is calculated using the transmittance calculation formula in the atmospheric scattering model, based on the red channel intensity value of that pixel location and the global atmospheric light intensity value obtained in step S122. Considering that the red channel is significantly affected by water absorption, this initial transmittance value needs to be corrected. Simultaneously, the transmittance estimation is constrained using the blue channel intensity value of the corresponding pixel location in the weak absorption channel image of the water body to avoid overestimation of transmittance. After calculating the transmittance values ​​for all pixel locations, a two-dimensional matrix with the same size as the original image is formed; this two-dimensional matrix is ​​the water body transmittance distribution map.

[0021] Step S124: Based on the water transmittance distribution map and the global atmospheric light intensity value, perform specular reflection component stripping processing on each single-color channel image of the UAV-view water surface image frame unit, remove the additive light component caused by specular reflection from each single-color channel image, obtain the diffuse reflection image of each single-color channel, and merge all single-color channel diffuse reflection images to generate a water area diffuse reflection base image. Perform pixel-by-pixel difference operation processing between the UAV-view water surface image frame unit and the water area diffuse reflection base image to obtain the specular reflection intensity distribution map.

[0022] For the red, green, and blue channels of the original UAV-view water surface image frame unit, specular reflection stripping is performed. According to the atmospheric scattering physics model, the total light intensity received by the sensor equals the diffuse reflection component plus the specular reflection component. Using the water transmittance distribution map obtained in step S123 and the global atmospheric light intensity value obtained in step S122, the diffuse reflection component is separated from the original intensity of each channel by solving the scattering equation. Specifically, the original intensity value of each pixel position in each channel image is subtracted from the component contributed by atmospheric light, and then divided by the transmittance value corresponding to that pixel position to obtain the diffuse reflection intensity value of that channel image at that pixel position. The diffuse reflection intensity values ​​of the three channel images are recombined to generate a diffuse reflection base image of the water area. At the same time, the intensity values ​​of the red, green, and blue channels of each pixel position in the original UAV-view water surface image frame unit are subtracted from the diffuse reflection intensity value of the corresponding channel in the diffuse reflection base image of the water area to obtain the specular reflection intensity value of each pixel position. The specular reflection intensity values ​​of all pixel positions are arranged according to pixel coordinates to form a specular reflection intensity distribution map.

[0023] Step S125: Perform spatial variation illumination field estimation processing on the specular reflection intensity distribution map, extract the illumination intensity variation trend surface of the specular reflection intensity distribution map in the spatial dimension, and generate the specular reflection spatial variation illumination field.

[0024] The specular reflection intensity values ​​in the specular reflection intensity distribution map are not uniformly distributed spatially, but vary with the water wave angle and the direction of solar incidence. A two-dimensional surface fitting method is used to analyze the spatial trend of the specular reflection intensity distribution map. By setting the fitting order parameter and solving the coefficients of the fitted surface using the least squares method, a continuous spatially varying illumination trend surface defined across the entire image plane is obtained. This spatially varying illumination trend surface reflects the macroscopic variation law of specular reflection intensity and is labeled as the spatially varying illumination field of specular reflection.

[0025] Step S126: Perform pixel-by-pixel illumination compensation processing on the specular reflection spatial variation illumination field and the diffuse reflection base image of the water area to eliminate local brightness differences caused by the unevenness of the specular reflection space in the diffuse reflection base image of the water area, thereby obtaining an illumination-normalized water area image. Perform water surface texture detail preservation filtering processing on the illumination-normalized water area image to suppress residual noise in the illumination-normalized water area image while maintaining the clarity of the navigation mark edge structure, thereby generating an illumination-normalized water area image with preserved texture details. Perform local contrast stretching processing on the illumination-normalized water area image with preserved texture details to enhance the visual difference between the navigation mark main area and the water background area, thereby obtaining an illumination-normalized water area image with enhanced contrast.

[0026] Because the spatial variation of specular reflection affects areas that should be uniform in the diffuse substrate image, compensation is needed. The intensity values ​​of the red, green, and blue channels at each pixel location in the diffuse substrate image of the water area are divided by the intensity value of the corresponding pixel location in the spatially varied specular reflection illumination field, resulting in a preliminary illumination-normalized image. Then, an edge-preserving filter is applied to this preliminary illumination-normalized image. The edge-preserving filter smooths noise while preserving the sharpness of the beacon edges. After filtering, an illumination-normalized water area image with preserved texture details is obtained. Finally, local contrast stretching is performed on this image: a local window is set centered on each pixel location, the maximum and minimum pixel intensity values ​​within the window are calculated, and the intensity value of that pixel location is mapped to the full dynamic range, enhancing the contrast between the beacon and the background, resulting in the final illumination-normalized water area image.

[0027] Step S130: Perform multi-scale structural tensor field construction processing on the illumination-normalized water area image to obtain a navigation mark structure tensor field description map. The navigation mark structure tensor field description map represents the main structural direction and structural saliency intensity of the navigation mark at each pixel position.

[0028] Step S131: Perform multi-scale smoothing processing on the illumination-normalized water area image. Use multiple smoothing kernels with different smoothing scales to perform convolution smoothing operation on the illumination-normalized water area image to generate a multi-scale smoothed image set. Each multi-scale smoothed image in the multi-scale smoothed image set corresponds to a smoothing scale.

[0029] Multiple smoothing scale parameters are defined, with each smoothing scale corresponding to the standard deviation of a Gaussian kernel. For each smoothing scale, the illumination-normalized water image is convolved using the Gaussian kernel corresponding to that scale to obtain the smoothed image at that scale. The smoothed images corresponding to all smoothing scales constitute a multi-scale smoothed image set.

[0030] Step S132: Perform horizontal gradient calculation and vertical gradient calculation on each multi-scale smoothed image in the multi-scale smoothed image set to obtain the horizontal gradient distribution map and vertical gradient distribution map corresponding to each multi-scale smoothed image. Based on the horizontal gradient distribution map and the vertical gradient distribution map corresponding to each multi-scale smoothed image, construct a local gradient structure tensor for each pixel position. The local gradient structure tensor includes a horizontal gradient squared term, a vertical gradient squared term, and a horizontal and vertical gradient product term.

[0031] For a smoothed image at each scale, the horizontal and vertical gradient values ​​are calculated for each pixel location. The horizontal gradient is obtained by dividing the difference in pixel intensity to the right and left of that pixel location by twice the pixel pitch; the vertical gradient is obtained by dividing the difference in pixel intensity below and above that pixel location by twice the pixel pitch. At each pixel location, the square of the horizontal gradient is used as the first component of the local gradient structure tensor, the square of the vertical gradient as the second component, and the product of the horizontal and vertical gradients as the third component, forming the local gradient structure tensor for that pixel location.

[0032] Step S133: Perform local neighborhood Gaussian weighted aggregation on the local gradient structure tensors of all pixel locations corresponding to each multi-scale smoothed image to obtain the aggregated structure tensor field corresponding to each multi-scale smoothed image. Perform multi-scale fusion processing on the aggregated structure tensor fields corresponding to each smoothing scale, select the aggregated structure tensor with the strongest structural saliency at each pixel location in each smoothing scale as the fused structure tensor at that pixel location, and generate a description map of the beacon structure tensor field.

[0033] At each scale, the local gradient structure tensor at each pixel location is averaged using a Gaussian weighted average within its neighborhood. The Gaussian weights decrease as the distance from the center pixel increases. The weighted average yields the aggregated structure tensor field for that scale. Then, the aggregated structure tensor fields at different scales are fused: for each pixel location, the largest eigenvalue of the aggregated structure tensor field at each scale is calculated. This largest eigenvalue characterizes the structural saliency strength at that scale, and the aggregated structure tensor corresponding to the scale with the largest eigenvalue is selected as the fused structure tensor for that pixel location. The fused structure tensors for all pixel locations are arranged according to pixel coordinates, forming a description map of the beacon structure tensor field.

[0034] Step S134: Perform feature decomposition on the fused structure tensor of each pixel position in the navigation mark structure tensor field description diagram to obtain feature value pairs and feature vector pairs for each pixel position. The larger feature value in the feature value pair represents the structural saliency intensity of the pixel position, and the feature vector corresponding to the larger feature value in the feature vector pair represents the structural principal direction of the pixel position.

[0035] For the fused structure tensor matrix at each pixel location, solve its characteristic equation to calculate the eigenvalues ​​and corresponding eigenvectors. The solution to the characteristic equation yields two eigenvalues, with the larger eigenvalue labeled as the first eigenvalue and the smaller as the second eigenvalue. The direction of the eigenvector corresponding to the first eigenvalue is the principal structural direction at that pixel location. The magnitude of the first eigenvalue reflects the saliency of the structure at that location; a larger value indicates a more pronounced directional structure at that location.

[0036] Step S135: Perform feature value difference analysis on the feature value pairs of each pixel position in the tensor field description diagram of the navigation mark structure, calculate the difference between the larger and smaller feature values ​​in the feature value pairs, and mark the pixel positions where the difference value exceeds the preset difference threshold as anisotropic structure pixels.

[0037] For each pixel location, the difference between the first feature value and the second feature value is calculated, and this difference is used as the degree of difference value. A preset difference threshold value is set. When the degree of difference value is greater than the difference threshold, it indicates that the structure at that pixel location exhibits obvious anisotropy, and the pixel is marked as an anisotropic structure pixel. If the degree of difference value is less than or equal to the difference threshold, it indicates that the structure at that location is isotropic or has a weak structure.

[0038] Step S136: Perform directional consistency region growth processing based on the structural principal direction of the anisotropic structural pixels. Aggregate spatially adjacent anisotropic structural pixels with an angular deviation of less than a preset directional deviation threshold into directional consistency structural regions to generate a set of directional consistency structural regions. Perform regional geometric attribute description processing on each directional consistency structural region in the set of directional consistency structural regions. Calculate the regional area parameter and the regional principal direction parameter of each directional consistency structural region. Associate the regional area parameter and the regional principal direction parameter with each pixel position within the directional consistency structural region. Use the navigation beacon structure tensor field description map associated with the regional geometric attribute description information as the updated navigation beacon structure tensor field description map.

[0039] Using anisotropic structural pixels as seed points, a region growing algorithm is executed. Starting from the seed pixel, it checks whether its neighboring pixels are anisotropic structural pixels, and whether the angle difference between the principal structural direction of the neighboring pixels and the principal direction of the current region is less than a preset direction deviation threshold. If the condition is met, the neighboring pixel is added to the current region, and the principal direction of the region is updated to the average of the principal structural directions of all pixels within the region. This growth process is repeated until no new pixels are added. After growth, each connected region is a region with consistent orientation. For each region, the number of pixels it contains is counted as the region area parameter, and the average direction of the principal structural directions of all pixels within the region is calculated as the region principal direction parameter. The above region attribute information is appended to the structural tensor field description corresponding to each pixel within the region.

[0040] Step S140: Based on the tensor field description graph of the navigation beacon structure, perform topology extraction processing on the navigation beacon structure skeleton to generate a topology skeleton graph of the navigation beacon structure. The topology skeleton graph of the navigation beacon structure includes a set of skeleton nodes at the top of the navigation beacon and a set of skeleton chains supporting the navigation beacon.

[0041] Step S141: Perform non-maximum suppression processing on the structural saliency intensity of each pixel position in the navigation beacon structure tensor field description diagram. Compare the structural saliency intensity of each pixel position with the structural saliency intensity of its neighboring pixel positions along the main structural direction. Only retain pixels with local maxima as structural skeleton candidate pixels to generate a set of structural skeleton candidate pixels. Perform hysteresis double-threshold connection processing on the set of structural skeleton candidate pixels. Select structural skeleton candidate pixels with structural saliency intensity exceeding a preset high connection threshold as skeleton seed pixels. Starting from each skeleton seed pixel, trace structural skeleton candidate pixels with structural saliency intensity exceeding a preset low connection threshold in the neighborhood along the main structural direction. Use the traced pixel chain as a skeleton chain segment to generate a set of skeleton chain segments.

[0042] For each pixel location, extend one pixel forward and one pixel backward along its structural principal direction, comparing the structural saliency intensity of that pixel with that of its two adjacent pixels. If the pixel's intensity is greater than both, retain it as a local maximum; otherwise, suppress it. The pixels retained after non-maximum suppression form a set of candidate pixels for the structural skeleton. Then, perform hysteresis dual-threshold connections: set a higher connection threshold and a lower threshold. Starting from all candidate pixels with structural saliency intensity greater than the higher threshold, trace candidate pixels in the neighborhood with intensity greater than the lower threshold along the structural principal direction, and treat the traced continuous pixel chain as a skeleton chain segment. All skeleton chains form a set of skeleton chain segments.

[0043] Step S142: Perform segment endpoint detection processing on each skeleton segment in the skeleton segment set, extract the two terminating pixels at both ends of each skeleton segment as the segment endpoints of the skeleton segment, and record the pixel coordinates of each segment endpoint.

[0044] For each skeleton chain segment, its endpoint is the pixel in the chain that is connected to only one neighboring chain segment pixel. Traverse the chain segment starting from one end until you find an end with no following pixels, and record the coordinates of that end pixel. Similarly, find the other end. Record the coordinates of two endpoints for each skeleton chain segment.

[0045] Step S143: Perform endpoint connection processing on the endpoints of different skeleton chain segments in the skeleton chain segment set whose spatial distance is less than a preset endpoint connection distance threshold and whose directional deviation is less than a preset directional deviation threshold. Merge the endpoints of the chain segments that meet the connection conditions into a continuous skeleton chain to generate a continuous skeleton chain set. Perform topology node identification processing on each continuous skeleton chain in the continuous skeleton chain set. Extract the intersection pixels of at least three skeleton chain segments in the continuous skeleton chain as skeleton branch nodes. Extract the endpoint pixels of the continuous skeleton chain connected by only one skeleton chain segment as skeleton terminal nodes to generate a skeleton node set.

[0046] For any two distinct skeleton chain segments, calculate the Euclidean distance between them and the angle between the tangent directions at their respective endpoints. If the distance is less than the endpoint connection distance threshold and the angle between the directions is less than the direction deviation threshold, connect the two endpoints, merging the two skeleton chain segments. Repeat this process until no further connections are possible. The resulting connections form a longer, continuous skeleton chain. For each continuous skeleton chain, identify its topology: if a pixel connects to three or more skeleton chain branches, mark it as a skeleton branch node; if a pixel is the end of a chain segment and connects to only one skeleton chain segment, mark it as a skeleton terminal node. All branch nodes and terminal nodes constitute the skeleton node set.

[0047] Step S144: Perform spatial clustering analysis on each skeleton branch node and each skeleton terminal node in the skeleton node set. Merge skeleton nodes of the same type whose spatial distance is less than the preset node merging distance threshold into one skeleton node to obtain a merged skeleton node set. Perform curve simplification on the skeleton chain segments between skeleton nodes in the merged skeleton node set, retain the starting node and ending node of the skeleton chain segment, and simplify the middle pixel sequence of the skeleton chain segment into a straight line connecting the starting node and the ending node to generate a navigation beacon structure topology skeleton diagram.

[0048] For skeleton branch nodes, branch nodes whose spatial distance is less than the node merging distance threshold are merged into one node, and the coordinates of the new node are the average of the coordinates of all nodes before the merge. The same applies to skeleton terminal nodes. After merging, the nodes in the merged skeleton node set are organized into a graph structure according to their connections on the skeleton chain segments. For a skeleton chain segment connecting two nodes, the original curved curve is replaced with a straight line edge, and the endpoints of the straight line are the two nodes. The final graph structure is the navigation beacon structure topology skeleton graph.

[0049] Step S145: Perform spatial position sorting processing on the skeleton nodes in the topological skeleton diagram of the navigation mark structure. Sort the skeleton nodes in the image of the normalized water area according to the order of the vertical pixel coordinates of the skeleton nodes in the image of the normalized water area from small to large. Extract the first preset number of skeleton nodes with the smallest vertical pixel coordinates as the set of skeleton nodes at the top of the navigation mark.

[0050] In an image coordinate system, the vertical coordinate typically increases downwards, so pixels with smaller vertical coordinates are located at the top of the image. The top of a navigation beacon is usually located at the very top of the main body of the beacon in the image. The skeleton nodes are sorted in ascending order of their vertical coordinates, and the top-ranked nodes are selected as the top skeleton nodes of the navigation beacon; these nodes correspond to the top structure of the navigation beacon.

[0051] Step S146: Extract the straight connecting edges from the topology skeleton diagram of the navigation beacon structure, which connect each skeleton node in the set of skeleton nodes at the top of the navigation beacon with the skeleton node with the larger vertical pixel coordinate, and organize the extracted straight connecting edges into a set of navigation beacon support skeleton chains according to the vertical extension direction.

[0052] For each top skeleton node of a navigation beacon, search downwards along the topological skeleton graph to find the skeleton node connected to it with a larger vertical coordinate, and record the connection edges between them. Organize these connection edges into chains in a top-down order, with each chain corresponding to a navigation beacon support skeleton, and all chains constitute a set of navigation beacon support skeleton chains.

[0053] Step S150: Call the preset UAV ground-to-ground spatial intersection positioning model to perform spatial intersection calculation on the set of top skeleton nodes of the navigation beacon and the set of support skeleton chains of the navigation beacon to generate precise positioning information of the navigation beacon, which includes the geographic spatial coordinates and elevation information of the navigation beacon.

[0054] Step S151: Obtain the airborne positioning data of the UAV when it collects the water surface image frame unit from the perspective of the UAV. The airborne positioning data includes the longitude coordinates of the UAV, the latitude coordinates of the UAV, the altitude of the UAV, and the heading angle parameters of the UAV.

[0055] Read the longitude, latitude, and altitude values ​​of the image frame at the time of acquisition from the UAV's onboard GPS receiver. Read the heading angle value at that moment from the UAV's inertial navigation module, which is the angle between the UAV's nose and true north.

[0056] Step S152: Obtain the airborne imaging attitude data of the UAV when it is collecting the water surface image frame unit from the perspective of the UAV. The airborne imaging attitude data includes the gimbal pitch angle parameter and the gimbal roll angle parameter.

[0057] Read the gimbal pitch angle (the angle between the camera's optical axis and the horizontal plane) and gimbal roll angle (the angle of rotation of the camera around the optical axis) at the moment the image frame was captured from the drone's gimbal angle sensor.

[0058] Step S153: Construct a spatial position reference point for the UAV based on the airborne positioning data, and construct an imaging line-of-sight direction vector of the airborne imaging device in space based on the airborne imaging attitude data and the UAV heading angle parameters, thereby generating a directional spatial imaging line of sight.

[0059] The longitude, latitude, and altitude values ​​of the UAV are used to construct a three-dimensional spatial point as the UAV's spatial position reference point. The spatial direction of the camera's optical axis is calculated based on the gimbal's pitch angle and the UAV's heading angle: first, the heading angle is converted into a horizontal direction angle, and then the components of the light direction in the horizontal and vertical planes are calculated by combining the pitch angle values, resulting in a unit direction vector in three-dimensional space. This unit direction vector is the imaging line-of-sight direction vector.

[0060] Step S154: Convert the image pixel coordinates of each buoy top skeleton node in the set of buoy top skeleton nodes into normalized coordinates of the imaging plane in the normalized water image through the intrinsic parameter mapping relationship of the airborne imaging device, and obtain the set of normalized coordinates of buoy top skeleton nodes.

[0061] Based on the camera's intrinsic parameter matrix (including parameters such as focal length and principal point coordinates), the pixel coordinates of the top skeleton node of each beacon are converted into normalized imaging plane coordinates. The conversion relationship is determined by the camera projection model.

[0062] Step S155: Based on the UAV spatial location reference point, the directional spatial imaging line of sight, and the normalized coordinate set of the top skeleton node of the buoy, construct an imaging ray, and introduce the average elevation plane of the water area where the buoy is located as the intersection constraint plane. Calculate the intersection point of each imaging ray with the constraint plane, solve for the three-dimensional spatial coordinates of the top skeleton node of the buoy in geographic space, generate the top spatial coordinate set of the buoy, extract the image pixel coordinates of the bottom skeleton node of each buoy support skeleton chain in the buoy support skeleton chain set, and convert the image pixel coordinates of the bottom skeleton node into normalized coordinates of the imaging plane through the intrinsic parameter mapping relationship to obtain the normalized coordinate set of the bottom support.

[0063] Starting from the UAV's spatial reference point, a ray is constructed along the imaging line of sight. The ray passes through the direction determined by the normalized coordinates of the top skeleton node of the beacon. By calculating the intersection of the ray with the average elevation plane (with a known constant elevation value), the three-dimensional coordinates of the top node of the beacon in the object space are obtained. Similarly, the same operation is performed on the bottom nodes of the beacon support skeleton chain to obtain the spatial coordinates of the bottom nodes.

[0064] Step S156: Based on the UAV spatial location reference point, the directional spatial imaging line of sight, and the normalized coordinate set of the support bottom, construct an imaging ray, and introduce the average elevation plane of the known water area where the navigation beacon is located as the intersection constraint plane. Calculate the intersection point of each imaging ray with the constraint plane, solve for the three-dimensional spatial coordinates of the bottom skeleton node of each navigation beacon support skeleton chain in geographic space, and generate the support bottom spatial coordinate set.

[0065] This step is based on the same principle as step S155, except that the object being processed is the supporting bottom node.

[0066] Step S157: Calculate the spatial vector of the buoy support skeleton chain in the object space based on the spatial coordinate set of the top of the buoy and the spatial coordinate set of the bottom of the support, and determine the buoy tilt state parameters based on the angle between the spatial vector and the direction of gravity.

[0067] For each support frame chain of the navigation beacon, calculate the vector pointing from its bottom spatial coordinates to its top spatial coordinates. Calculate the angle between this vector and the unit vector of gravity; this angle is the tilt angle of the navigation beacon. The tilt parameters include the tilt direction angle (the projection direction of the spatial vector onto the horizontal plane) and the tilt angle value.

[0068] Step S158: Perform spatial position fusion processing on the three-dimensional spatial coordinates of each top skeleton node of the navigation mark in the set of top spatial coordinates of the navigation mark to obtain the geographic spatial coordinates of the navigation mark, and perform tilt error compensation processing on the geographic spatial coordinates of the navigation mark according to the tilt state parameters of the navigation mark to generate precise positioning information of the navigation mark, which includes the geographic spatial coordinates of the navigation mark and the elevation information of the navigation mark.

[0069] The average spatial coordinates of multiple top skeleton nodes of the navigation beacon are taken to obtain the center coordinates of the top of the beacon, which are then used as the beacon's geospatial coordinates. Based on the beacon's tilt angle and tilt direction angle, the horizontal offset between the bottom and top of the beacon is calculated, and the geospatial coordinates are corrected to obtain accurate coordinates after eliminating the tilt effect. The beacon's elevation information is obtained by averaging the spatial coordinates of the top nodes of the beacon.

[0070] Step S210: Perform navigation mark time-series displacement tracking processing on the multi-frame UAV-view water surface image frame units in the UAV aerial image data set of navigation mark water area to obtain the navigation mark time-series displacement vector field, and perform time-series filtering processing on the navigation mark precise positioning information according to the navigation mark time-series displacement vector field to generate time-stable navigation mark precise positioning information.

[0071] Step S211: Extract the first UAV-view water surface image frame unit and the second UAV-view water surface image frame unit that are temporally adjacent from the UAV aerial image data set of navigation mark water area.

[0072] From the collection of drone aerial images of waterways and navigation marks, two adjacent frames are extracted in chronological order of acquisition time and denoted as the first drone-view water surface image frame unit and the second drone-view water surface image frame unit.

[0073] Step S212: Perform the water surface reflection component stripping process and the light field consistency reconstruction process on the first UAV view water surface image frame unit and the second UAV view water surface image frame unit respectively to obtain the first illumination normalized water surface image and the second illumination normalized water surface image.

[0074] Perform all the processing steps S121 to S126 on the first UAV-view water surface image frame unit to obtain the first illumination-normalized water surface image. Perform the same processing on the second UAV-view water surface image frame unit to obtain the second illumination-normalized water surface image.

[0075] Step S213: Perform the navigation mark multi-scale structure tensor field construction processing on the first illumination-normalized water area image and the second illumination-normalized water area image respectively to obtain the first navigation mark structure tensor field description map and the second navigation mark structure tensor field description map.

[0076] Perform all steps S131 to S135 on the first illumination-normalized water area image to obtain the tensor field description map of the first navigational aid structure. Perform the same processing on the second illumination-normalized water area image to obtain the tensor field description map of the second navigational aid structure.

[0077] Step S214: Based on the first and second navigation beacon structure tensor field description maps, perform temporal variation analysis of the structure tensor field, calculate the angular change of the main direction of the structure at the corresponding pixel position and the intensity change of the structural saliency intensity, and generate a temporal variation distribution map of the structure tensor field.

[0078] For each pixel location, the principal structural direction angle and structural saliency intensity value are extracted from the first beacon structure tensor field description map, and the corresponding principal structural direction angle and structural saliency intensity value are extracted from the second beacon structure tensor field description map. The absolute value of the difference between the two angle values ​​is calculated as the angle change, and the difference between the two intensity values ​​is calculated as the intensity change. The angle change and intensity change values ​​for all pixel locations are filled into two matrices of the same size as the image, respectively, to obtain the angle change distribution map and the intensity change distribution map. The two are combined and labeled to form the temporal change distribution map of the structure tensor field.

[0079] Step S215: Perform navigation beacon motion region segmentation processing on the temporal variation distribution map of the structural tensor field, and extract the pixel regions in the temporal variation distribution map of the structural tensor field where the change in structural saliency intensity exceeds a preset intensity change threshold as the navigation beacon motion region.

[0080] In the intensity variation distribution map, pixel locations where the intensity variation exceeds a preset intensity variation threshold are marked as candidate moving pixels. Connectivity analysis is performed on the marked candidate moving pixels, clustering all candidate moving pixels into several connected regions. These connected regions are then used as the beacon's moving region.

[0081] Step S216: Perform pixel-level displacement vector estimation processing within the movement area of ​​the navigation mark. For each pixel within the movement area of ​​the navigation mark, perform local image block matching between the first illumination-normalized water area image and the second illumination-normalized water area image to obtain the pixel displacement vector of each pixel and generate a set of pixel-level displacement vectors.

[0082] For each pixel location within the beacon's movement area, a fixed-size image block is extracted from the first illumination-normalized water image, centered on that pixel location, as a reference block. A target block with the highest matching degree is searched in the second illumination-normalized water image, with the search range limited to the neighborhood of the corresponding pixel location in the second illumination-normalized water image. The matching degree is measured by calculating the normalized cross-correlation value between the reference block and the candidate block. After finding the target block with the highest matching degree, the positional offset vector between the center of the target block and the center of the reference block is calculated as the pixel displacement vector for that pixel. The pixel displacement vectors of all pixels within the beacon's movement area constitute a set of pixel-level displacement vectors.

[0083] Step S217: Perform motion consistency region clustering processing on the pixel-level displacement vector set, and aggregate spatially adjacent pixels whose pixel displacement vector direction deviation and magnitude deviation are both within the preset motion consistency tolerance range into motion consistency region blocks to generate a set of motion consistency region blocks.

[0084] The algorithm iterates through each pixel displacement vector in the set of pixel-level displacement vectors, using the pixel displacement vector as a feature for region growing. Starting from the seed pixel, it checks whether the angular deviation between the displacement vector direction of its neighboring pixels and the average displacement vector direction of the current region is less than a preset directional deviation tolerance, and also checks whether the magnitude deviation of the displacement vector is less than a preset magnitude deviation tolerance. If the conditions are met, the neighboring pixel is added to the current region, and the average displacement vector of the region is updated. This growth process is repeated until no new pixels are added. After growth, each connected region is a motion-consistent region block. All motion-consistent region blocks constitute a set of motion-consistent region blocks.

[0085] Step S218: Perform average displacement vector calculation on each motion-consistent region block in the set of motion-consistent region blocks to obtain a representative displacement vector for each motion-consistent region block, and assign the representative displacement vector to all pixels in the motion-consistent region block to generate a beacon temporal displacement vector field in the image pixel domain.

[0086] For each motion-consistent region block, calculate the arithmetic mean of the displacement vectors of all pixels within that region to obtain a representative displacement vector for that region block. This representative displacement vector is then assigned to all pixel positions within that region block. After traversing all motion-consistent region blocks, the displacement vector for each pixel position within the entire beacon motion region is obtained. These displacement vectors are arranged according to pixel coordinates, forming a beacon temporal displacement vector field.

[0087] Step S219: Perform time-series filtering on the precise positioning information of the navigation beacon based on the time-series displacement vector field to generate time-stable precise positioning information of the navigation beacon.

[0088] Step S2191: Extract the displacement vector of the pixel position of each top skeleton node in the top skeleton node set of the navigation beacon from the navigation beacon temporal displacement vector field, and use it as the top node displacement vector set of the navigation beacon.

[0089] In the set of top skeleton nodes of the beacon generated in step S145, the pixel coordinates of each top skeleton node of the beacon are obtained. In the beacon temporal displacement vector field generated in step S218, the displacement vectors corresponding to these pixel coordinate positions are found, and all the extracted displacement vectors are combined into a set of displacement vectors for the top nodes of the beacon.

[0090] Step S2192: Based on the changes in the set of displacement vectors of the top nodes of the beacon and the spatial position reference point of the UAV, perform temporal state prediction processing on each top spatial coordinate of the top nodes of the beacon in the current frame. According to the top spatial coordinates of the top nodes of the previous frame, the image motion trend reflected by the displacement vectors of the top nodes of the beacon and the position change of the UAV, predict the top spatial coordinates of the top nodes of the beacon in the current frame through the spatial forward intersection geometric relationship, and generate the set of predicted coordinates of the top nodes of the beacon.

[0091] Obtain the spatial coordinates of the beacon top obtained from the previous frame. Calculate the motion trend of the beacon top in the image plane based on the displacement vectors in the set of displacement vectors for the beacon top nodes. Convert this motion trend into an approximate displacement in object space. Considering the change in the UAV's position, extrapolate the spatial coordinates of the beacon top in the previous frame to obtain the predicted spatial coordinates of the beacon top in the current frame. Perform the same prediction process on all beacon top nodes to generate a set of predicted spatial coordinates for the beacon top.

[0092] Step S2193: Obtain the set of spatial coordinates of the top of the navigation beacon obtained by the spatial intersection solution of the current frame as the set of spatial observation coordinates of the top of the navigation beacon.

[0093] From the set of top spatial coordinates of the beacon generated in step S155, obtain the top spatial coordinates of the beacon calculated in the current frame, and use them as the set of top spatial observation coordinates of the beacon.

[0094] Step S2194: Perform weighted fusion processing on the predicted coordinate set and the observed coordinate set of the top space of the navigation beacon. According to the preset prediction confidence weight and observation confidence weight, perform weighted averaging on the predicted coordinates and observed coordinates to obtain the fused coordinate set of the top space of the navigation beacon.

[0095] For each beacon top node, a preset prediction confidence weight and observation confidence weight are established, with the sum of the two weights being 1. The predicted coordinates of the beacon top space are multiplied by the prediction confidence weight, and then the observed coordinates of the beacon top space are multiplied by the observation confidence weight to obtain the fused coordinates of the beacon top space. The same operation is performed on all nodes to obtain the set of fused coordinates of the beacon top space.

[0096] Step S2195: Extract the displacement vectors of the pixel positions of each bottom skeleton node in the navigation beacon support skeleton chain set from the navigation beacon temporal displacement vector field, and use them as the support bottom node displacement vector set.

[0097] In the beacon support skeleton chain set generated in step S146, the pixel coordinates of each bottom skeleton node are obtained. The displacement vectors corresponding to these pixel coordinate positions are found in the beacon temporal displacement vector field to form a set of displacement vectors for the bottom support nodes.

[0098] Step S2196: Based on the changes in the set of displacement vectors of the support bottom node and the spatial position reference point of the UAV, perform temporal state prediction processing on each support bottom spatial coordinate in the set of support bottom spatial coordinates of the current frame, predict the support bottom spatial coordinates of the current frame according to the support bottom spatial coordinates of the previous frame and the displacement vectors of the support bottom node, and generate a set of support bottom spatial prediction coordinates.

[0099] Obtain the spatial coordinates of the support bottom end obtained from the previous frame. Based on the displacement vectors in the set of displacement vectors of the support bottom end nodes, and combined with the UAV position change, extrapolate the spatial coordinates of the support bottom end from the previous frame to obtain the predicted spatial coordinates of the support bottom end in the current frame. Perform the same prediction process on all bottom end nodes to generate a set of predicted spatial coordinates of the support bottom end.

[0100] Step S2197: Obtain the set of spatial coordinates of the support bottom end obtained by the spatial intersection solution of the current frame as the set of spatial observation coordinates of the support bottom end.

[0101] From the support bottom spatial coordinate set generated in step S156, obtain the support bottom spatial coordinates calculated in the current frame, and use them as the support bottom spatial observation coordinate set.

[0102] Step S2198: Perform weighted fusion processing on the predicted coordinate set of the support bottom space and the observed coordinate set of the support bottom space. According to the preset prediction confidence weight and observation confidence weight, perform weighted averaging on the predicted coordinates and observed coordinates to obtain the fused coordinate set of the support bottom space.

[0103] For each bottom node, multiply the predicted coordinates of the supporting bottom space by the prediction confidence weight, and add the observed coordinates of the supporting bottom space multiplied by the observation confidence weight to obtain the fused coordinates of the supporting bottom space. Perform the same operation on all bottom nodes to obtain the set of fused coordinates of the supporting bottom space.

[0104] Step S2199: Recalculate the fusion spatial vector of the buoy support skeleton chain in the object space based on the fusion coordinate set of the top space of the buoy and the fusion coordinate set of the bottom space of the support, and determine the fusion buoy tilt state parameters based on the angle between the fusion spatial vector and the gravity direction.

[0105] For each beacon support skeleton chain, using the fused spatial coordinates of the top of the beacon obtained in step S2194 and the fused spatial coordinates of the bottom of the support obtained in step S2198, calculate the fused spatial vector from the bottom to the top. Calculate the angle between this fused spatial vector and the unit vector of gravity direction to obtain the fused beacon tilt angle, and combine this with the horizontal projection direction of the vector to obtain the fused beacon tilt state parameters.

[0106] Step S21910: Perform spatial position fusion processing on the three-dimensional spatial coordinates of each top skeleton node of the navigation beacon in the fused coordinate set to obtain fused navigation beacon geospatial coordinates, and perform tilt error compensation processing on the fused navigation beacon geospatial coordinates according to the tilt state parameters of the fused navigation beacon to generate time-stable and accurate navigation beacon positioning information.

[0107] The arithmetic mean of all coordinates in the fused coordinate set of the top space of the navigation beacon obtained in step S2194 is taken to obtain the fused navigation beacon geospatial coordinates. Based on the tilt angle and tilt direction of the fused navigation beacon obtained in step S2199, the horizontal offset between the bottom and top of the navigation beacon is calculated to correct the fused navigation beacon geospatial coordinates and eliminate the influence of tilt on positioning. The corrected coordinates and corresponding elevation information are output as time-stable and accurate navigation beacon positioning information.

[0108] Step S310: Perform navigation beacon type identification processing on the navigation beacon structure topology skeleton diagram to obtain navigation beacon type identification results, and perform type-specific correction processing on the navigation beacon precise positioning information based on the navigation beacon type identification results to generate type-corrected navigation beacon precise positioning information.

[0109] Step S311: Perform top geometric shape description processing on the set of top skeleton nodes of the navigation beacon in the topology skeleton diagram of the navigation beacon structure, calculate the spatial distribution convex hull of each skeleton node in the set of top skeleton nodes of the navigation beacon, and extract the shape description parameters of the spatial distribution convex hull, the shape description parameters including the convex hull area parameter and the convex hull roundness parameter.

[0110] Project all nodes from the top skeleton node set of the beacon onto the image plane and calculate the convex hull polygon of these points. The convex hull is the smallest convex polygon that contains all points. Calculate the area of ​​the convex hull polygon as the convex hull area parameter. Calculate the convex hull roundness parameter, which is equal to the convex hull area multiplied by four times pi and divided by the square of the convex hull perimeter. The closer the roundness value is to 1, the closer the shape is to a circle.

[0111] Step S312: Perform support structure topology description processing on the set of support skeleton chains in the navigation beacon structure topology skeleton diagram, count the number of support skeleton chains in the set of support skeleton chains, calculate the angle distribution characteristics between each support skeleton chain, and extract the length distribution characteristics of the support skeleton chains to obtain the support structure topology descriptor.

[0112] The number of support skeleton chains in the set of navigation beacon support skeleton chains is counted. For every two adjacent support skeleton chains, the angle between them at the connection point is calculated, resulting in an angle sequence. The length of each support skeleton chain is calculated, resulting in a length sequence. The number of support skeleton chains, the statistical characteristics of the angle sequence (such as mean and variance), and the statistical characteristics of the length sequence are combined to form a support structure topology descriptor.

[0113] Step S313: Perform statistical processing on the structural saliency intensity of each skeleton node in the set of skeleton nodes at the top of the navigation mark in the navigation mark structure tensor field description diagram to generate the distribution characteristics of the top structural saliency intensity.

[0114] The structural saliency intensity values ​​of each skeleton node in the top skeleton node set of the navigation beacon are extracted from the tensor field description diagram of the navigation beacon structure. The maximum, minimum, mean and standard deviation of these intensity values ​​are calculated as the distribution characteristics of the top structural saliency intensity.

[0115] Step S314: Combine the shape description parameters, the topology descriptor of the support structure, and the salience intensity distribution features of the top structure into a navigation mark structure feature vector.

[0116] The convex hull area parameter, convex hull roundness parameter, number of supporting skeleton chains, statistical characteristics of the included angle of the supporting skeleton chains, statistical characteristics of the length of the supporting skeleton chains, and the salience intensity distribution characteristics of the top structure are concatenated into a multi-dimensional vector, which is the navigation mark structure feature vector.

[0117] Step S315: Input the navigation mark structure feature vector into a preset navigation mark type classifier for type discrimination processing. The navigation mark type classifier outputs the navigation mark type label corresponding to the navigation mark structure topology skeleton diagram based on the mapping relationship between the structure feature vector and the navigation mark type label, as the navigation mark type recognition result.

[0118] The preset navigation mark type classifier is a support vector machine (SVM) classifier. This SVM classifier is pre-trained using a large number of navigation mark structure feature vector samples with labeled navigation mark types. The navigation mark structure feature vector generated in step S314 is input into this classifier, and the classifier outputs a navigation mark type label, such as "conical navigation mark", "spherical navigation mark", "cylindrical navigation mark", etc., as the navigation mark type recognition result.

[0119] Step S316: Determine the standard navigation mark geometric parameters for the navigation mark type based on the navigation mark type identification result. The standard navigation mark geometric parameters include the standard top diameter parameter and the standard support height parameter.

[0120] Based on the navigation mark type identification results, the geometric parameters of this type of navigation mark are retrieved from the preset standard navigation mark parameter database, including the standard top diameter value and the standard support height value.

[0121] Step S317: Obtain the top circle radius parameter from the top shape fitting parameters of the navigation mark and the vertical distance parameter between the end point coordinates and the start point coordinates of the support structure from the vertical support shape fitting parameters of the navigation mark, and use them as the measured top radius of the navigation mark and the measured support height of the navigation mark.

[0122] The radius of the fitted circle of the top skeleton node set extracted from the topological skeleton diagram of the navigation beacon is used as the measured top radius of the navigation beacon. The vertical distance between the bottom and top skeleton nodes is calculated from the set of navigation beacon support skeleton chains as the measured support height of the navigation beacon.

[0123] Step S318: Calculate the dimensional conformity of the beacon top based on the ratio of the standard top diameter parameter to the measured top radius of the beacon, and calculate the beacon top identification confidence weight based on the conformity; calculate the dimensional conformity of the support structure based on the ratio of the standard support height parameter to the measured support height of the beacon, and calculate the support structure identification confidence weight based on the conformity.

[0124] The ratio of the standard top diameter parameter to the measured top radius of the navigation mark multiplied by two is used as the navigation mark top dimensional conformity. This conformity is then mapped to a confidence weight for navigation mark top identification; the closer the conformity is to 1, the higher the confidence weight. Similarly, the conformity of the supporting structure dimensions is calculated and mapped to a confidence weight for supporting structure identification.

[0125] Step S319: Use the confidence weight of the top of the navigation mark and the confidence weight of the support structure as weighting factors for each positioning coordinate in the precise positioning information of the navigation mark, and perform weight adjustment processing on the geographic spatial coordinates of the navigation mark and the elevation information of the navigation mark to generate the precise positioning information of the navigation mark after type correction.

[0126] The navigation mark geospatial coordinates and elevation information generated in step S158 are weighted and averaged using the confidence weights of the top of the navigation mark and the support structure. This yields the accurate positioning information of the navigation mark after type correction.

[0127] Step S410: Perform multi-navigation mark instance detection processing on the same frame of the UAV-view water surface image unit in the UAV aerial image data set of navigation mark water area, extract all navigation mark instances existing in the UAV-view water surface image frame unit, and obtain the navigation mark structure topology skeleton diagram corresponding to each navigation mark instance. Extract the navigation mark top skeleton node set from the navigation mark structure topology skeleton diagram corresponding to each navigation mark instance, calculate the navigation mark top center pixel coordinates of each navigation mark instance, and obtain the multi-navigation mark top center coordinate set.

[0128] In step S410, steps S120 to S144 are performed on a frame of UAV-view water surface image to obtain a topological skeleton diagram of the navigation beacon structure for all navigation beacon instances in the frame. For each navigation beacon instance, a set of top skeleton nodes is extracted from the corresponding topological skeleton diagram, and the arithmetic mean of the pixel coordinates of all skeleton nodes in the set is calculated as the center pixel coordinate of the top of the navigation beacon for that instance. The center pixel coordinates of the top of all navigation beacon instances constitute a set of multiple navigation beacon top center coordinates.

[0129] Step S411: Perform spatial distance calculation processing on the set of top center coordinates of the multiple navigation marks, calculate the pixel Euclidean distance between the top center pixel coordinates of any two navigation mark instances, and generate a navigation mark spacing matrix.

[0130] For each pair of beacon instances in the set of top center coordinates of multiple beacons, calculate the Euclidean distance between their top center pixel coordinates. Organize all distance values ​​into a two-dimensional matrix according to the index of the beacon instances. This two-dimensional matrix is ​​the beacon spacing matrix, where the element in the i-th row and j-th column represents the pixel distance between the i-th beacon and the j-th beacon.

[0131] Step S412: Perform relative orientation calculation on the set of top center coordinates of the multiple navigation marks, calculate the direction angle of the line connecting the top center pixel coordinates of any two navigation mark instances, and generate a relative orientation matrix between navigation marks.

[0132] For each pair of beacon instances, calculate the direction angle of the line connecting the center pixel coordinates of the tops of the two beacons (with true north as a reference). Organize all direction angle values ​​into a two-dimensional matrix according to the index of the beacon instances; this two-dimensional matrix is ​​the relative bearing matrix between beacons.

[0133] Step S413: Based on the beacon spacing matrix, the relative orientation matrix between beacons, and the initial spatial distance estimate calculated from the beacon geospatial coordinate set, construct a spatial association description graph between beacons. The weight of the graph edges is jointly determined by the spatial association strength in the image domain and the initial distance estimate in the object space. The spatial association description graph between beacons uses each beacon instance as a graph node and beacon pairs whose spatial distance between beacons is less than a preset beacon association distance threshold as graph edges. The weight of the graph edges is jointly determined by the spatial distance between beacons and the relative orientation between beacons.

[0134] Each navigational aid instance is treated as a node in the graph. For navigational aid pairs in the navigational aid spacing matrix that are less than a preset navigational aid association distance threshold, an edge is added between the two nodes. The weight of the edge is calculated based on the spatial distance and relative orientation between the navigational aids; the closer the distance and the more consistent the relative orientation, the greater the weight. Simultaneously, the initial spatial distance estimate calculated from the navigational aid geospatial coordinate set is incorporated as additional information into the edge weights, resulting in a description graph of the spatial association relationships between navigational aids.

[0135] Step S414: Perform temporal fusion processing on the spatial relationship description graph between navigation marks corresponding to the frame units of the multi-frame UAV-view water surface image. Retain graph nodes and graph edges that appear stably in multiple consecutive frames, and remove isolated graph nodes that appear only in a single frame to generate a spatial relationship topology graph between navigation marks. Extract navigation mark instances corresponding to all graph nodes from the spatial relationship topology graph between navigation marks, obtain the navigation mark geospatial coordinates of each navigation mark instance through the spatial intersection solution processing, and generate a set of navigation mark geospatial coordinates.

[0136] Temporal fusion is performed on the spatial association description graphs of navigation marks corresponding to multiple frames of images. Only graph nodes and edges that appear in multiple consecutive frames (e.g., three or more consecutive frames) are retained, while isolated nodes that appear only in a single frame are removed. The resulting graph is a topological graph of spatial associations between navigation marks. From this topological graph, navigation mark instances corresponding to all nodes are extracted, and the geospatial coordinates of each navigation mark instance are obtained through step S150 to generate a set of navigation mark geospatial coordinates.

[0137] Step S415: Extract the two navigation mark instances connected by each graph edge from the spatial association topology graph between navigation marks, and obtain the spatial distance between the two navigation mark instances and the relative orientation between the navigation marks as graph edge constraint conditions.

[0138] Traverse each edge in the spatial association topology graph between navigation marks, find the two navigation mark instances connected by the edge, and read the spatial distance and relative orientation between the two navigation mark instances from the navigation mark spacing matrix and the navigation mark relative orientation matrix, as the constraint conditions of the graph edge.

[0139] Step S416: Based on the set of geospatial coordinates of the navigation marks, the map edge constraints, and the imaging geometric relationships of each navigation mark instance, construct a spatial consistency optimization objective function. The spatial consistency optimization objective function represents the cumulative deviation between the pixel spacing and relative orientation of the images formed after back-projecting the currently optimized geospatial coordinates of each navigation mark to each original image coordinate system and the image domain observations recorded in the map edge constraints, while also incorporating the smoothness constraint of the spatial position of the navigation marks.

[0140] A spatial consistency optimization objective function is constructed. The first part of this objective function is calculated by back-projecting the optimized beacon geospatial coordinates onto the original image frames, obtaining back-projected pixel coordinates, and then calculating the sum of squares of the distances between these back-projected pixel coordinates and the deviations between their relative orientations and the observations recorded in the map edge constraints. The second part of the objective function is a beacon spatial position smoothness constraint, limiting the variation in beacon coordinates between adjacent frames to a minimum. The entire objective function is a weighted sum of these two parts.

[0141] Step S417: Iteratively minimize the spatial consistency optimization objective function. In each iteration, adjust the geospatial coordinates of each beacon instance to gradually reduce the cumulative deviation until the cumulative deviation is less than the preset convergence threshold, and obtain the optimized set of geospatial coordinates of the beacon.

[0142] The spatial consistency optimization objective function is solved iteratively using either gradient descent or the Gauss-Newton method. In each iteration, the partial derivative of the objective function with respect to the geospatial coordinates of each navigation mark is calculated, and the coordinate values ​​are adjusted along the negative gradient direction. This iteration is repeated until the objective function value is less than a preset convergence threshold, resulting in the optimized set of navigation mark geospatial coordinates.

[0143] Step S418: Obtain the navigation beacon tilt state parameters for each navigation beacon instance, associate the navigation beacon tilt state parameters with the corresponding optimized navigation beacon geospatial coordinates, and perform elevation direction correction processing on the optimized navigation beacon geospatial coordinates according to the navigation beacon tilt state parameters to obtain the optimized navigation beacon elevation information for each navigation beacon instance. Combine the optimized navigation beacon geospatial coordinate set and the optimized navigation beacon elevation information into globally optimized precise navigation beacon positioning information.

[0144] Obtain the navigation mark tilt state parameters for each navigation mark instance from step S157. Based on the tilt angle value in the tilt state parameters, correct the elevation component of the optimized navigation mark geospatial coordinates to eliminate the influence of tilt on elevation measurement. Combine the corrected navigation mark geospatial coordinates and navigation mark elevation information to form globally optimized precise navigation mark positioning information.

[0145] For example, the method may further include: step S510: performing dynamic response analysis of water surface waves on the tensor field description diagram of the navigation mark structure to obtain a wave-modulated structure tensor field, and performing wave disturbance compensation on the topology extraction process of the navigation mark structure skeleton based on the wave-modulated structure tensor field to generate a wave-compensated navigation mark structure topology skeleton diagram.

[0146] Step S511: Select multiple consecutive UAV-view water surface image frame units with a time interval shorter than the preset wave sensing time interval from the UAV aerial photography navigation mark water area image data set, and construct a short-time frame sequence image set.

[0147] From the collection of drone aerial images of navigation marks in water areas, consecutive frames of water surface images from the drone's perspective are selected in chronological order of acquisition time markers, with the time interval between adjacent frames being less than a preset wave sensing time interval (e.g., 0.1 seconds). These frame units are then combined into a short-time frame sequence image set.

[0148] Step S512: Perform the water surface reflection component stripping process and the light field consistency reconstruction process on each frame of the UAV perspective water surface image unit in the short-time frame sequence image set to obtain a short-time illumination normalized water surface image sequence.

[0149] For each frame of the UAV-view water surface image unit in the short-time frame sequence image set, all processing steps S121 to S126 are performed sequentially to obtain the illumination-normalized water surface image corresponding to each frame. All the illumination-normalized water surface images of all frames are arranged in chronological order to form a short-time illumination-normalized water surface image sequence.

[0150] Step S513: Perform the navigation mark multi-scale structure tensor field construction processing on each frame of the short-time illumination normalized water area image sequence to obtain a short-time navigation mark structure tensor field description map sequence.

[0151] For each frame of the short-time illumination-normalized water area image sequence, perform all the processing steps S131 to S135 to obtain the navigation aid structure tensor field description map corresponding to each frame. Arrange the navigation aid structure tensor field description maps of all frames in chronological order to form a short-time navigation aid structure tensor field description map sequence.

[0152] Step S514: Extract the structural saliency intensity value sequence along the time dimension for each pixel position in the short-time navigation beacon structure tensor field description map sequence, and perform spectral decomposition processing on the structural saliency intensity value sequence to obtain the wave frequency domain response spectrum of each pixel position.

[0153] For each pixel location in the short-time beacon structure tensor field description sequence, the structural saliency intensity value of that pixel location is extracted along the time axis (frame sequence direction) to form a time series. A Fast Fourier Transform is performed on this time series to transform it from the time domain to the frequency domain, obtaining the wave frequency response spectrum for that pixel location. The wave frequency response spectrum reflects the energy distribution of the wave at different frequencies.

[0154] Step S515: Extract the amplitude response value and phase response value corresponding to the wave fundamental frequency component from the wave frequency domain response spectrum of each pixel location, use the amplitude response value as the wave modulation intensity parameter of the pixel location, and use the phase response value as the wave modulation phase parameter of the pixel location.

[0155] In the wave frequency response spectrum at each pixel location, find the frequency component corresponding to the main wave frequency (preset wave frequency range), extract the amplitude of the component as the wave modulation intensity parameter, and extract the phase of the component as the wave modulation phase parameter.

[0156] Step S516: Combine the wave modulation intensity parameters of all pixel locations to generate a wave modulation intensity distribution map, and combine the wave modulation phase parameters of all pixel locations to generate a wave modulation phase distribution map.

[0157] The wave modulation intensity parameters at each pixel location are filled into a two-dimensional matrix of the same size as the image according to the pixel coordinates to obtain the wave modulation intensity distribution map. Similarly, the wave modulation phase parameters at each pixel location are filled into another two-dimensional matrix to obtain the wave modulation phase distribution map.

[0158] Step S517: Perform field-to-field fusion processing on the wave modulation intensity distribution map and the wave modulation phase distribution map with the navigation beacon structure tensor field description map. At each pixel position of the navigation beacon structure tensor field description map, the wave modulation intensity parameter and the wave modulation phase parameter are used as the additional modulation dimension of the structure tensor at that pixel position to obtain the wave modulation structure tensor field.

[0159] For each pixel location in the navigation beacon structure tensor field description diagram, its original structure tensor is fused with the wave modulation intensity parameters and wave modulation phase parameters obtained in step S515. The fusion method involves adding the wave modulation intensity parameters and wave modulation phase parameters as additional dimensions to the structure tensor, forming an expanded tensor, labeled as the wave modulation structure tensor. The wave modulation structure tensors at all pixel locations constitute the wave modulation structure tensor field.

[0160] Step S518: Perform wave modulation inversion processing on the structural tensor at each pixel position in the wave-modulated structural tensor field, and separate the periodic structural response component caused by wave motion and the static structural response component caused by the beacon body from the structural tensor.

[0161] For each pixel position in the wave-modulated structure tensor field, the structure tensor is phase-demodulated using the wave-modulated phase parameter, and the time-varying structure response is separated into periodic wave response components and static beacon body response components.

[0162] Step S519: Perform structural saliency intensity reconstruction processing on the static structural response components, map the static structural response components back to the standard structural tensor space, and generate a wave-interference-free structural tensor field description diagram.

[0163] The static structural response components separated in step S518 are reorganized into a standard structural tensor form, i.e., each pixel location contains a horizontal gradient squared term, a vertical gradient squared term, and a horizontal-vertical gradient product term. Arranging these structural tensors according to pixel coordinates yields a description map of the wave-interference-free structural tensor field.

[0164] Step S520: Perform feature decomposition on the de-wave structure tensor at each pixel position in the de-wave interference structure tensor field description diagram to obtain de-wave eigenvalue pairs and de-wave eigenvector pairs at each pixel position. The larger eigenvalue in the de-wave eigenvalue pair represents the saliency intensity of the de-wave structure at that pixel position.

[0165] For the wave-removed structure tensor matrix at each pixel location in the wave-removed structure tensor field description diagram, solve its characteristic equation and calculate the eigenvalues ​​and corresponding eigenvectors. The larger eigenvalue is labeled as the first eigenvalue, representing the structural saliency intensity at that pixel location after wave removal; the smaller eigenvalue is labeled as the second eigenvalue; the eigenvector corresponding to the first eigenvalue represents the principal direction of the structure at that pixel location after wave removal.

[0166] Step S521: Perform non-maximum suppression processing on the saliency intensity of the de-wavy structure. Compare the saliency intensity of the de-wavy structure at each pixel position with the saliency intensity of the de-wavy structure at its neighboring pixel positions along the main direction of the de-wavy structure. Only retain pixels with local maxima of the de-wavy structure saliency intensity as candidate pixels for wave compensation skeleton, and generate a set of candidate pixels for wave compensation skeleton.

[0167] For each pixel location, extend one pixel forward and one pixel backward along the main direction of its de-wavy structure. Compare the saliency intensity of the de-wavy structure of this pixel with that of its two adjacent pixels. If the intensity of this pixel is greater than both, retain this pixel as a candidate pixel for the wave compensation skeleton; otherwise, suppress it. All retained pixels constitute the set of candidate pixels for the wave compensation skeleton.

[0168] Step S522: Perform hysteresis dual-threshold connection processing on the wave compensation skeleton candidate pixel set, select wave compensation skeleton candidate pixels whose wave structure saliency intensity exceeds the preset high connection threshold as wave compensation skeleton seed pixels, start from each wave compensation skeleton seed pixel and track wave compensation skeleton candidate pixels in the neighborhood whose wave structure saliency intensity exceeds the preset low connection threshold along the main direction of wave structure, and use the tracked pixel chain as wave compensation skeleton chain segment to generate a wave compensation skeleton chain segment set.

[0169] Two thresholds are set: a high connectivity threshold and a low connectivity threshold. First, a candidate pixel for wave compensation skeleton with a de-wavy structure saliency intensity greater than the high connectivity threshold is used as a seed point. Then, along the main de-wavy structure direction of the seed point, candidate pixels for wave compensation skeleton with a de-wavy structure saliency intensity greater than the low connectivity threshold within the neighborhood are traced. The traced continuous pixel chain is considered as a wave compensation skeleton chain segment. All chains constitute the wave compensation skeleton chain segment set.

[0170] Step S523: Perform segment endpoint detection processing on each wave compensation skeleton segment in the wave compensation skeleton segment set, extract the two terminating pixels of each wave compensation skeleton segment as the wave compensation segment endpoints of the wave compensation skeleton segment, and record the pixel coordinates of each wave compensation segment endpoint.

[0171] For each wave-compensated skeleton chain segment, find its two endpoints: an endpoint is a pixel in the chain segment that is connected to only one neighboring pixel. Record the pixel coordinates of each endpoint.

[0172] Step S524: Perform endpoint connection processing on the endpoints of different wave compensation skeleton chain segments in the wave compensation skeleton chain segment set whose spatial distance is less than the preset endpoint connection distance threshold and whose directional deviation is less than the preset directional deviation threshold. Merge the endpoints of wave compensation chain segments that meet the connection conditions into a continuous wave compensation skeleton chain to generate a continuous wave compensation skeleton chain set.

[0173] For any two distinct wave compensation skeleton chain segments, calculate the Euclidean distance between them and the angle between the tangent directions at the two endpoints. If the distance is less than the endpoint connection distance threshold and the angle is less than the direction deviation threshold, connect the two endpoints and merge the two chain segments. Repeat this process until no further connections are possible, resulting in a set of continuous wave compensation skeleton chains.

[0174] Step S525: Perform topology node identification processing on each wave compensation continuous skeleton chain in the wave compensation continuous skeleton chain set, extract the intersection pixels of at least three wave compensation skeleton chain segments in the wave compensation continuous skeleton chain as wave compensation skeleton branch nodes, extract the endpoint pixels of only one wave compensation skeleton chain segment in the wave compensation continuous skeleton chain as wave compensation skeleton terminal nodes, and generate a wave compensation skeleton node set.

[0175] For each wave-compensated continuous skeleton chain, traverse every pixel on the chain. If a pixel is connected to three or more wave-compensated skeleton chain branches, mark it as a wave-compensated skeleton branch node; if a pixel is the end of a chain and is connected to only one chain segment, mark it as a wave-compensated skeleton terminal node. All branch nodes and terminal nodes constitute the wave-compensated skeleton node set.

[0176] Step S526: Perform spatial clustering analysis on the wave compensation skeleton branch nodes and wave compensation skeleton terminal nodes in the wave compensation skeleton node set, merge wave compensation skeleton nodes of the same type whose spatial distance is less than the preset node merging distance threshold into one skeleton node, and perform curve simplification processing on the wave compensation skeleton chain segments between the merged skeleton nodes to generate a wave compensation navigation beacon structure topology skeleton diagram.

[0177] For the branch nodes of the wave compensation skeleton, branch nodes whose spatial distance is less than the merging distance threshold are merged into one node, and the coordinates of the new node are taken as the arithmetic mean of the coordinates of the nodes before merging. The terminal nodes of the wave compensation skeleton are processed in the same way. After merging, the wave compensation skeleton chain segments between nodes are replaced with straight lines, and the starting and ending nodes are retained. The final graph structure is the topology skeleton diagram of the wave compensation navigation beacon structure.

[0178] Step S610: Obtain a set of multi-view drone aerial images of the same navigation mark from multiple different acquisition locations. Perform cross-view navigation mark topology consistency matching processing on the set of multi-view drone aerial images of the navigation mark to obtain multi-view navigation mark topology matching relationships. Based on the multi-view navigation mark topology matching relationships, perform multi-view intersection constraint optimization processing on the navigation mark precise positioning information to generate multi-view fused navigation mark precise positioning information.

[0179] Step S611: Perform the water surface reflection component stripping process and the light field consistency reconstruction process on the water surface image frame unit corresponding to each acquisition location from the perspective of the UAV, respectively, to obtain the water surface image with normalized illumination from each perspective.

[0180] For each UAV-view water surface image frame unit obtained at each acquisition location, all processing steps S121 to S126 are executed sequentially to obtain the illumination-normalized water area image under that viewpoint.

[0181] Step S612: Perform the multi-scale structure tensor field construction processing on the illumination-normalized water area images from each viewpoint to obtain the tensor field description map of the navigation mark structure from each viewpoint.

[0182] For each viewpoint, the illumination-normalized water area image is processed using steps S131 to S135 to obtain the navigation mark structure tensor field description map under that viewpoint.

[0183] Step S613: Based on the tensor field description diagram of the navigation beacon structure from each perspective, perform the topology extraction process of the navigation beacon structure skeleton to generate the topology skeleton diagram of the navigation beacon structure from each perspective.

[0184] For each viewpoint, perform all the processing steps S141 to S146 on the tensor field description diagram of the beacon structure to obtain the topological skeleton diagram of the beacon structure under that viewpoint.

[0185] Step S614: Extract the set of top skeleton nodes of the navigation beacon from the top skeleton diagram of the navigation beacon structure from each viewpoint, and calculate the topological connection relationship descriptor between each skeleton node in the set of top skeleton nodes of the navigation beacon under that viewpoint. The topological connection relationship descriptor includes the number of adjacent edges and the distribution of the included angle between adjacent edges between the top skeleton nodes of the navigation beacon.

[0186] For each viewpoint of the navigation beacon's topological skeleton diagram, extract the set of skeleton nodes at the top of the beacon. Count the number of edges directly connecting these nodes as the number of adjacent edges. Calculate the angle between each adjacent edge and the horizontal direction to obtain an angle sequence. Combine the statistical characteristics (mean, variance) of the number of adjacent edges and the angle sequence to form a topological connectivity descriptor.

[0187] Step S615: Extract the set of navigation beacon support skeleton chains from the topological skeleton diagram of the navigation beacon structure from each viewpoint, and calculate the skeleton chain length sequence and the angle sequence between skeleton chains in the set of navigation beacon support skeleton chains under that viewpoint, as the topological descriptor of the support structure.

[0188] For each viewpoint of the beacon structure topology skeleton diagram, extract the set of beacon support skeleton chains. Calculate the length of each support skeleton chain to obtain a length sequence. Calculate the angle between adjacent support skeleton chains to obtain an angle sequence. Combine the statistical features of the length sequences and the angle sequences to form a support structure topology descriptor.

[0189] Step S616: Combine the topology connection relationship descriptor and the support structure topology descriptor from the same perspective into a navigation beacon topology feature descriptor for that perspective.

[0190] The topological connection relationship descriptor obtained in step S614 and the support structure topological descriptor obtained in step S615 are concatenated into a multi-dimensional vector, which serves as the navigation mark topological feature descriptor from this perspective.

[0191] Step S617: Perform topological graph matching processing on any two navigation beacon topological feature descriptors from different perspectives, calculate the graph edit distance between the two perspectives, and construct a navigation beacon topological similarity matrix based on the graph edit distance.

[0192] For two navigational aid topology feature descriptors from different perspectives, they are treated as graph structures. The minimum number of edit operations required to transform one graph into the other is calculated as the graph edit distance. The smaller the graph edit distance, the more similar the navigational aid topologies of the two perspectives. The graph edit distances between all pairs of perspectives are organized into a square matrix to obtain the navigational aid topology similarity matrix between perspectives.

[0193] Step S618: Based on the inter-view navigation mark topology similarity matrix, perform cross-view navigation mark instance correspondence inference processing, mark navigation mark instances in view pairs whose navigation mark topology similarity exceeds the preset topology matching threshold as cross-view correspondences of the same navigation mark instance, and generate multi-view navigation mark topology matching relationships.

[0194] In the inter-viewpoint beacon topological similarity matrix, for each pair of views, if their graph edit distance is less than a preset topological matching threshold, then the beacon instances in these two views are considered to correspond to the same physical beacon. These cross-viewpoint correspondences are recorded to generate multi-view beacon topological matching relationships.

[0195] Step S619: Extract the set of top skeleton nodes and the set of support skeleton chains of the same buoy instance under all matching views based on the multi-view buoy topology matching relationship.

[0196] For each beacon instance, based on the multi-view beacon topology matching relationship, collect the beacon top skeleton node set and beacon support skeleton chain set of the beacon instance under all matching views.

[0197] Step S620: Obtain the UAV spatial position reference point and directional spatial imaging line of view corresponding to each matched viewpoint, and construct the object space intersection geometry relationship for each matched viewpoint.

[0198] For each matched viewpoint, obtain the corresponding UAV spatial location reference point (longitude, latitude, elevation) and the directional spatial imaging line of sight, construct the ray equation from the UAV position to the navigation mark, and form the object space intersection geometry relationship of that viewpoint.

[0199] Step S621: Solve the top object space coordinates of each skeleton node in the set of skeleton nodes of the same beacon instance under each matching viewpoint into the top object space coordinates of the beacon under that viewpoint through the object space intersection geometry relationship of their respective viewpoints, and obtain the set of top object space coordinates of the beacon from multiple viewpoints.

[0200] For each matched viewpoint, the image pixel coordinates of each node in the set of skeleton nodes at the top of the beacon extracted in step S619 are used to calculate the corresponding three-dimensional coordinates in the object space using the object space intersection geometry constructed in step S620. The beacon top coordinates calculated from all views constitute a multi-view beacon top object space coordinate set.

[0201] Step S622: Solve the bottom skeleton nodes of each support skeleton chain in the set of support skeleton chains of the same navigation beacon instance under each matching viewpoint into the support bottom object space coordinates under that viewpoint through the object space intersection geometry relationship of their respective viewpoints, and obtain the set of support bottom object space coordinates of multiple viewpoints.

[0202] For each matching viewpoint, the image pixel coordinates of the bottom skeleton node of each support skeleton chain in the beacon support skeleton chain set are used to calculate the three-dimensional coordinates of the bottom node in the object space using object space intersection geometry. The bottom support coordinates calculated from all views constitute a multi-view support bottom object space coordinate set.

[0203] Step S623: Perform cross-validation processing between viewpoints on the set of object space coordinates on the top of the multi-view navigation beacon, calculate the spatial Euclidean distance between the object space coordinates on the top of the navigation beacon calculated from any two different viewpoints, and mark the object space coordinates on the top of the navigation beacon whose spatial Euclidean distance exceeds the preset viewpoint consistency deviation threshold as viewpoint abnormal coordinates.

[0204] For any two viewpoints in the multi-view object space coordinate set of the beacon top, calculate the Euclidean distance between them. If this distance is greater than a preset viewpoint consistency deviation threshold, these two coordinates may be incorrect, and they are marked as viewpoint abnormal coordinates.

[0205] Step S624: After removing the abnormal coordinates of the viewpoint from the set of top object space coordinates of the multi-view navigation beacon, the remaining top object space coordinates of the navigation beacon are subjected to weighted average aggregation processing, wherein the weight of each viewpoint is determined by the cosine value of the angle between the imaging line of that viewpoint and the vertical direction, so as to obtain the multi-view aggregated navigation beacon geospatial coordinates.

[0206] Coordinates marked as anomaly points are removed. The remaining coordinates are then weighted and averaged, with each coordinate's weight equal to the cosine of the angle between the image line of sight and the vertical direction (the smaller the angle, the more perpendicular the line of sight, and the greater the weight). The weighted average coordinates are used as the geospatial coordinates of the multi-view aggregated navigation beacon.

[0207] Step S625: Perform cross-validation processing between viewpoints on the set of multi-view support bottom object space coordinates, remove abnormal viewpoint coordinates, and then perform weighted average aggregation processing to obtain multi-view aggregated support bottom space coordinates.

[0208] Perform the same cross-validation and weighted average processing as in steps S623 and S624 on the set of support bottom coordinates to obtain the multi-view aggregated support bottom spatial coordinates.

[0209] Step S626: Calculate the spatial tilt vector of the navigation beacon under multi-view constraints based on the geospatial coordinates of the multi-view aggregated navigation beacon and the spatial coordinates of the bottom of the multi-view aggregated support, and determine the tilt angle parameter of the multi-view navigation beacon based on the angle between the spatial tilt vector and the gravity direction.

[0210] Calculate the vector pointing from the spatial coordinates of the bottom of the multi-view aggregation support to the geospatial coordinates of the multi-view aggregation beacon, and use it as the spatial tilt vector. Calculate the angle between this vector and the unit vector of the gravity direction, and use it as the tilt angle parameter of the multi-view beacon.

[0211] Step S627: Perform tilt error correction processing on the geospatial coordinates of the multi-view aggregated navigation beacon according to the tilt angle parameters of the multi-view navigation beacon to generate multi-view fused navigation beacon precise positioning information, which includes multi-view fused navigation beacon geospatial coordinates and multi-view fused navigation beacon elevation information.

[0212] Using the tilt angle parameters of the multi-view navigation beacon obtained in step S626, tilt error correction is performed on the geospatial coordinates of the multi-view aggregated navigation beacon to eliminate the influence of tilt on horizontal position and elevation measurements. The corrected coordinates are used as the geospatial coordinates of the multi-view fused navigation beacon, and the corrected elevation is used as the elevation information of the multi-view fused navigation beacon. The two are combined to form the precise positioning information of the multi-view fused navigation beacon.

[0213] Based on the same inventive concept, please refer to Figure 2 The diagram shows a schematic block diagram of a machine vision-based navigation mark precision positioning system 100 provided in an embodiment of this application. The machine vision-based navigation mark precision positioning system 100 may include a communication unit 110, a machine-readable storage medium 120, and a processor 130.

[0214] In this embodiment, alternatively, the machine-readable storage medium 120 can also be integrated into the processor 130 and can communicate and interact with external systems through the communication unit 110. The machine-readable storage medium 120 is used to store machine-executable instructions for executing the scheme of this application, and the processor 130 is used to execute the machine-executable instructions stored in the machine-readable storage medium 120 to implement the machine vision-based precise beacon positioning method provided in the aforementioned method embodiments.

[0215] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A machine vision-based method for precise positioning of navigational aids, characterized in that, The method includes: A set of aerial images of navigation marks in water areas taken by drones is acquired, wherein the set of aerial images of navigation marks in water areas taken by drones includes multiple continuously acquired water surface image frame units with acquisition time stamps from the perspective of drones; The image dataset of the navigation mark in the water area taken by the UAV is subjected to water surface specular reflection component stripping processing to obtain a diffuse reflection substrate image and a specular reflection intensity distribution map of the water area. Then, the diffuse reflection substrate image and the specular reflection intensity distribution map of the water area are subjected to light field consistency reconstruction processing to obtain an illumination-normalized water area image. The illumination-normalized water area image is processed to construct a multi-scale structural tensor field for navigation marks, resulting in a navigation mark structure tensor field description map. The navigation mark structure tensor field description map represents the main structural direction and structural saliency intensity of the navigation mark at each pixel position. Based on the tensor field description graph of the navigation beacon structure, the topology of the navigation beacon structure skeleton is extracted to generate a topological skeleton graph of the navigation beacon structure. The topological skeleton graph of the navigation beacon structure includes a set of skeleton nodes at the top of the navigation beacon and a set of skeleton chains supporting the navigation beacon. A preset UAV-to-ground spatial rendezvous and positioning model is invoked to perform spatial rendezvous calculation on the set of top skeleton nodes of the navigation beacon and the set of supporting skeleton chains of the navigation beacon, thereby generating precise positioning information of the navigation beacon, which includes the geographic spatial coordinates and elevation information of the navigation beacon.

2. The machine vision-based precise positioning method for navigation marks according to claim 1, characterized in that, The process involves stripping the specular reflection component from the drone-captured navigation mark water area image data set to obtain a diffuse reflection substrate image and a specular reflection intensity distribution map. Then, the diffuse reflection substrate image and the specular reflection intensity distribution map are subjected to light field consistency reconstruction processing to obtain an illumination-normalized water area image, including: Multi-color channel separation processing is performed on the water surface image frame units from the UAV perspective in the UAV aerial image data set of navigation mark water area to obtain multiple single-color channel images. The single-color channel image with the highest water absorption rate is selected as the strong water absorption channel image, and the single-color channel image with the lowest water absorption rate is selected as the weak water absorption channel image. Dark channel prior statistical processing is performed on the weak water absorption channel image. The minimum pixel intensity value is extracted in each local neighborhood window of the weak water absorption channel image as the dark channel pixel value to generate the dark channel prior image. The atmospheric light component estimation process is performed on the dark channel prior image, and the top preset proportion of pixels with the largest pixel intensity value in the dark channel prior image are extracted. The average pixel intensity value of the top preset proportion of pixels at the corresponding position in the weak absorption channel image of the water body is calculated as the global atmospheric light intensity value. Based on the global atmospheric light intensity value, the transmittance distribution map of the strong absorption channel image and the weak absorption channel image of the water body is estimated to obtain the water body transmittance distribution map. The transmittance value of each pixel in the water body transmittance distribution map represents the proportion of light transmitted by the water body at that pixel location. Based on the water transmittance distribution map and the global atmospheric light intensity value, specular reflection component stripping is performed on each single-color channel image of the UAV-view water surface image frame unit. The additive light component caused by specular reflection is removed from each single-color channel image to obtain the diffuse reflection image of each single-color channel. All single-color channel diffuse reflection images are merged to generate a water diffuse reflection base image. The UAV-view water surface image frame unit and the water diffuse reflection base image are subjected to pixel-by-pixel difference operation to obtain the specular reflection intensity distribution map. The spatial variation illumination field estimation process is performed on the specular reflection intensity distribution map to extract the illumination intensity variation trend surface of the specular reflection intensity distribution map in the spatial dimension, and generate the spatial variation illumination field of the specular reflection. The specular reflection spatial variation illumination field and the diffuse reflection base image of the water area are processed by pixel-by-pixel illumination compensation to eliminate local brightness differences caused by the unevenness of the specular reflection space in the diffuse reflection base image of the water area, resulting in an illumination-normalized water area image. The illumination-normalized water area image is then subjected to water surface texture detail preservation filtering to suppress residual noise while maintaining the clarity of the navigation mark edge structure, generating an illumination-normalized water area image with preserved texture details. The illumination-normalized water area image with preserved texture details is then subjected to local contrast stretching to enhance the visual difference between the navigation mark main area and the water background area, resulting in an illumination-normalized water area image with enhanced contrast.

3. The machine vision-based precise positioning method for navigation marks according to claim 1, characterized in that, The process of constructing a multi-scale structure tensor field for navigation marks on the illumination-normalized water area image yields a navigation mark structure tensor field description map, including: The illumination-normalized water image is subjected to multi-scale smoothing processing. Multiple smoothing kernels with different smoothing scales are used to perform convolution smoothing operations on the illumination-normalized water image to generate a multi-scale smoothed image set. Each multi-scale smoothed image in the multi-scale smoothed image set corresponds to a smoothing scale. For each multi-scale smoothed image in the multi-scale smoothed image set, horizontal and vertical gradient calculations are performed to obtain the horizontal and vertical gradient distribution maps corresponding to each multi-scale smoothed image. Based on the horizontal and vertical gradient distribution maps corresponding to each multi-scale smoothed image, a local gradient structure tensor is constructed for each pixel location. The local gradient structure tensor includes a horizontal gradient squared term, a vertical gradient squared term, and a horizontal-vertical gradient product term. Local neighborhood Gaussian weighted aggregation is performed on the local gradient structure tensors of all pixel locations corresponding to each multi-scale smoothed image to obtain the aggregated structure tensor field corresponding to each multi-scale smoothed image. The aggregated structure tensor fields corresponding to each smoothing scale are then fused at multiple scales. The aggregated structure tensor with the strongest structural saliency at each pixel location in each smoothing scale is selected as the fused structure tensor at that pixel location to generate a description map of the beacon structure tensor field. The fused structure tensor at each pixel location in the navigation mark structure tensor field description diagram is subjected to feature decomposition processing to obtain feature value pairs and feature vector pairs at each pixel location. The larger feature value in the feature value pair represents the structural saliency intensity at that pixel location, and the feature vector corresponding to the larger feature value in the feature vector pair represents the structural principal direction at that pixel location. The feature value difference analysis is performed on the feature value pairs of each pixel position in the tensor field description diagram of the navigation mark structure. The difference between the larger and smaller feature values ​​in the feature value pairs is calculated. The pixel positions where the difference value exceeds the preset difference threshold are marked as anisotropic structure pixels. Based on the structural principal direction of the anisotropic structural pixels, a direction-consistent region growth process is performed. Spatially adjacent anisotropic structural pixels with an angular deviation of less than a preset direction deviation threshold are aggregated into direction-consistent structural regions, generating a set of direction-consistent structural regions. For each direction-consistent structural region in the set, a region geometric attribute description process is performed. The region area parameter and the region principal direction parameter of each direction-consistent structural region are calculated, and the region area parameter and the region principal direction parameter are associated with each pixel position within the direction-consistent structural region. The navigation beacon structure tensor field description map associated with the region geometric attribute description information is used as the updated navigation beacon structure tensor field description map.

4. The machine vision-based precise positioning method for navigation marks according to claim 1, characterized in that, The process of extracting the topology of the navigation beacon structure skeleton based on the tensor field description graph of the navigation beacon structure to generate a navigation beacon structure topology skeleton graph includes: Non-maximum suppression processing is performed on the structural saliency intensity of each pixel position in the tensor field description diagram of the navigation beacon structure. The structural saliency intensity of each pixel position is compared with the structural saliency intensity of its neighboring pixel positions along the main structural direction. Only pixels with local maxima are retained as candidate pixels for the structural skeleton, generating a set of candidate pixels for the structural skeleton. The set of candidate pixels for the structural skeleton is then processed by hysteresis double-threshold connection processing. Candidate pixels for the structural skeleton whose structural saliency intensity exceeds a preset high connection threshold are selected as skeleton seed pixels. Starting from each skeleton seed pixel, candidate pixels for the structural skeleton whose structural saliency intensity exceeds a preset low connection threshold are tracked along the main structural direction. The tracked pixel chains are used as skeleton chain segments to generate a set of skeleton chain segments. For each skeleton segment in the skeleton segment set, perform segment endpoint detection processing, extract the two terminating pixels of each skeleton segment as the segment endpoints of the skeleton segment, and record the pixel coordinates of each segment endpoint. Endpoint connection processing is performed on the endpoints of different skeleton chain segments in the skeleton chain segment set whose spatial distance is less than a preset endpoint connection distance threshold and whose directional deviation is less than a preset directional deviation threshold. The endpoints of the chain segments that meet the connection conditions are merged into a continuous skeleton chain to generate a continuous skeleton chain set. Topological node identification processing is performed on each continuous skeleton chain in the continuous skeleton chain set. The intersection pixels of at least three skeleton chain segments in the continuous skeleton chain are extracted as skeleton branch nodes, and the endpoint pixels of the continuous skeleton chain connected by only one skeleton chain segment are extracted as skeleton terminal nodes to generate a skeleton node set. Spatial clustering analysis is performed on each skeleton branch node and each skeleton terminal node in the skeleton node set. Skeleton nodes of the same type whose spatial distance is less than the preset node merging distance threshold are merged into one skeleton node to obtain a merged skeleton node set. The skeleton chain segments between skeleton nodes in the merged skeleton node set are simplified by curve processing, retaining the start node and end node of the skeleton chain segment, and simplifying the middle pixel sequence of the skeleton chain segment into a straight line connecting the start node and the end node to generate a navigation beacon structure topology skeleton diagram. The skeleton nodes in the topological skeleton diagram of the navigation mark are sorted by spatial position. The skeleton nodes are sorted in ascending order of their vertical pixel coordinates in the normalized water image. The first preset number of skeleton nodes with the smallest vertical pixel coordinates are extracted as the set of skeleton nodes at the top of the navigation mark. Extract the straight connecting edges from the topological skeleton diagram of the navigation beacon structure, connecting each skeleton node in the set of skeleton nodes at the top of the navigation beacon with the skeleton node with the larger vertical pixel coordinate, and organize the extracted straight connecting edges into a set of navigation beacon support skeleton chains according to the vertical extension direction.

5. The machine vision-based precise positioning method for navigation marks according to claim 1, characterized in that, The process involves calling a pre-defined UAV ground-to-ground spatial rendezvous and positioning model to perform spatial rendezvous calculations on the set of top skeleton nodes of the beacon and the set of support skeleton chains of the beacon, generating precise beacon positioning information, including: The airborne positioning data of the UAV when it collects the water surface image frame unit from the perspective of the UAV is obtained. The airborne positioning data includes the longitude coordinates of the UAV, the latitude coordinates of the UAV, the altitude of the UAV, and the heading angle parameters of the UAV. Acquire airborne imaging attitude data of the UAV when it is collecting the UAV's viewpoint water surface image frame unit, the airborne imaging attitude data including gimbal pitch angle parameters and gimbal roll angle parameters; Based on the airborne positioning data, a spatial position reference point for the UAV is constructed. Based on the airborne imaging attitude data and the UAV heading angle parameters, an imaging line of sight direction vector of the airborne imaging device in space is constructed, generating a directional spatial imaging line of sight. The image pixel coordinates of each buoy top skeleton node in the set of buoy top skeleton nodes in the normalized water image are converted into normalized coordinates of the imaging plane through the intrinsic parameter mapping relationship of the airborne imaging device, thus obtaining the set of normalized coordinates of buoy top skeleton nodes. Based on the UAV spatial location reference point, the directional spatial imaging line of sight, and the normalized coordinate set of the top skeleton node of the navigation beacon, an imaging ray is constructed. The average elevation plane of the water area where the navigation beacon is located is introduced as the intersection constraint plane. The intersection point of each imaging ray with the constraint plane is calculated, and the three-dimensional spatial coordinates of the top skeleton node of the navigation beacon in geographic space are obtained. A set of top spatial coordinates of the navigation beacon is generated. The image pixel coordinates of the bottom skeleton node of each navigation beacon support skeleton chain in the set of navigation beacon support skeleton chains are extracted. The image pixel coordinates of the bottom skeleton node are converted into normalized coordinates of the imaging plane through the intrinsic parameter mapping relationship to obtain the set of normalized coordinates of the bottom support. Based on the UAV spatial location reference point, the directional spatial imaging line of sight, and the normalized coordinate set of the support bottom, an imaging ray is constructed, and the average elevation plane of the known water area where the navigation beacon is located is introduced as the intersection constraint plane. The intersection point of each imaging ray with the constraint plane is calculated, and the three-dimensional spatial coordinates of the bottom skeleton node of each navigation beacon support skeleton chain in geographic space are obtained, generating the support bottom spatial coordinate set. The spatial vector of the buoy support skeleton chain in the object space is calculated based on the set of spatial coordinates of the top of the buoy and the set of spatial coordinates of the bottom of the support. The tilt state parameters of the buoy are determined based on the angle between the spatial vector and the direction of gravity. The three-dimensional spatial coordinates of each top skeleton node of the navigation beacon in the set of top spatial coordinates of the navigation beacon are fused to obtain the navigation beacon geospatial coordinates. The tilt error compensation processing of the navigation beacon geospatial coordinates is then performed according to the navigation beacon tilt state parameters to generate the navigation beacon precise positioning information, which includes the navigation beacon geospatial coordinates and the navigation beacon elevation information.

6. The machine vision-based precise positioning method for navigation marks according to claim 1, characterized in that, The method further includes: The time-series displacement tracking processing of the navigation mark time-series displacement vector field is performed on the multi-frame water surface image frame unit from the UAV aerial image data set of navigation mark water area, and the navigation mark precise positioning information is subjected to time-series filtering processing based on the navigation mark time-series displacement vector field to generate time-stable navigation mark precise positioning information. The step of performing navigation mark time-series displacement tracking processing on multiple frames of UAV-view water surface image units in the UAV aerial image data set to obtain the navigation mark time-series displacement vector field includes: Extract the first and second UAV-view water surface image frame units that are temporally adjacent from the UAV aerial image data set of navigation mark water areas. The water surface image frame unit from the first UAV perspective and the water surface image frame unit from the second UAV perspective are respectively subjected to the water surface specular reflection component stripping process and the light field consistency reconstruction process to obtain the first illumination normalized water surface image and the second illumination normalized water surface image. The first illumination-normalized water area image and the second illumination-normalized water area image are respectively subjected to the navigation mark multi-scale structure tensor field construction processing to obtain the first navigation mark structure tensor field description map and the second navigation mark structure tensor field description map; Based on the first and second navigation beacon structure tensor field description maps, the temporal variation analysis of the structure tensor field is performed to calculate the angular change of the main direction of the structure and the intensity change of the structural salience intensity at the corresponding pixel positions, and to generate a temporal variation distribution map of the structure tensor field. The temporal variation distribution map of the structural tensor field is processed by beacon motion region segmentation, and the pixel regions in the temporal variation distribution map of the structural tensor field where the structural significance intensity change exceeds the preset intensity change threshold are extracted as beacon motion regions. Pixel-level displacement vector estimation is performed within the movement area of ​​the navigation mark. Local image patch matching is performed between the first illumination-normalized water area image and the second illumination-normalized water area image for each pixel within the movement area of ​​the navigation mark to obtain the pixel displacement vector of each pixel and generate a set of pixel-level displacement vectors. The pixel-level displacement vector set is subjected to motion consistency region clustering processing. Pixels that are spatially adjacent and whose pixel displacement vector direction deviation and magnitude deviation are both within the preset motion consistency tolerance range are aggregated into motion consistency region blocks to generate a set of motion consistency region blocks. The average displacement vector is calculated for each motion-consistent region block in the set of motion-consistent region blocks to obtain a representative displacement vector for each motion-consistent region block. The representative displacement vector is then assigned to all pixels within the motion-consistent region block to generate a beacon temporal displacement vector field in the pixel domain of the generated image.

7. The machine vision-based precise positioning method for navigation marks according to claim 6, characterized in that, The step of performing time-series filtering on the precise positioning information of the navigation beacon based on the time-series displacement vector field of the navigation beacon to generate time-stable precise positioning information of the navigation beacon includes: Extract the displacement vector of each pixel position of the top skeleton node in the top skeleton node set of the navigation beacon from the navigation beacon time-series displacement vector field, and use it as the navigation beacon top node displacement vector set; Based on the set of displacement vectors of the top nodes of the beacon and the changes in the spatial position reference point of the UAV, the temporal state prediction processing is performed on each top spatial coordinate of the top nodes of the beacon in the current frame. According to the top spatial coordinates of the top nodes of the previous frame, the image motion trend reflected by the displacement vectors of the top nodes of the beacon and the changes in the position of the UAV, the top spatial coordinates of the top nodes of the beacon in the current frame are predicted by the spatial forward intersection geometric relationship, and a set of predicted coordinates of the top nodes of the beacon is generated. The set of spatial coordinates of the top of the navigation beacon obtained by the spatial intersection solution processing of the current frame is used as the set of spatial observation coordinates of the top of the navigation beacon. The predicted coordinate set and the observed coordinate set of the top space of the navigation beacon are weighted and fused. The predicted coordinates and observed coordinates are weighted and averaged according to the preset prediction confidence weight and observation confidence weight to obtain the fused coordinate set of the top space of the navigation beacon. The displacement vectors of the pixel positions of each bottom skeleton node in the navigation beacon support skeleton chain set are extracted from the navigation beacon temporal displacement vector field and used as the support bottom node displacement vector set. Based on the set of displacement vectors of the support bottom node and the changes of the UAV spatial position reference point, the temporal state prediction processing is performed on each support bottom spatial coordinate in the support bottom spatial coordinate set of the current frame. The support bottom spatial prediction coordinate of the current frame is predicted according to the support bottom spatial coordinate of the previous frame and the displacement vector of the support bottom node, and a set of support bottom spatial prediction coordinates is generated. The set of spatial coordinates of the support bottom end obtained by the spatial intersection solution of the current frame is used as the set of spatial observation coordinates of the support bottom end. The predicted coordinate set and the observed coordinate set of the support bottom space are weighted and fused. The predicted coordinates and observed coordinates are weighted and averaged according to the preset prediction confidence weight and observation confidence weight to obtain the fused coordinate set of the support bottom space. Based on the fused coordinate set of the top space of the buoy and the fused coordinate set of the bottom space of the support, the fused spatial vector of the buoy support skeleton chain in the object space is recalculated, and the fused buoy tilt state parameters are determined based on the angle between the fused spatial vector and the gravity direction. The three-dimensional spatial coordinates of each top skeleton node of the navigation beacon in the fused coordinate set are fused to obtain the fused navigation beacon geospatial coordinates. The fused navigation beacon geospatial coordinates are then compensated for tilt error based on the tilt state parameters of the fused navigation beacon to generate time-stable and accurate navigation beacon positioning information.

8. The machine vision-based precise positioning method for navigation marks according to claim 1, characterized in that, The method further includes: The navigation mark structure topology skeleton diagram is processed for navigation mark type identification to obtain navigation mark type identification results. Based on the navigation mark type identification results, the navigation mark precise positioning information is processed for type specificity correction to generate type-corrected navigation mark precise positioning information. The process of performing navigation beacon type identification processing on the navigation beacon structure topology skeleton diagram to obtain navigation beacon type identification results includes: The top geometric shape description processing is performed on the set of top skeleton nodes of the navigation beacon in the top skeleton node diagram of the navigation beacon structure. The spatial distribution convex hull of each skeleton node in the set of top skeleton nodes of the navigation beacon is calculated, and the shape description parameters of the spatial distribution convex hull are extracted. The shape description parameters include the convex hull area parameter and the convex hull roundness parameter. The set of support skeleton chains in the topological skeleton diagram of the navigation beacon structure is subjected to support structure topology description processing. The number of support skeleton chains in the set of support skeleton chains is counted, the angle distribution characteristics between each support skeleton chain are calculated, and the length distribution characteristics of the support skeleton chains are extracted to obtain the support structure topology descriptor. The structural saliency intensity of each skeleton node in the set of skeleton nodes at the top of the navigation beacon is statistically processed in the tensor field description diagram of the navigation beacon structure to generate the distribution characteristics of the top structural saliency intensity. The shape description parameters, the topological descriptor of the supporting structure, and the salience intensity distribution features of the top structure are combined into a navigation mark structure feature vector; The navigation mark structure feature vector is input into a preset navigation mark type classifier for type discrimination processing. The navigation mark type classifier outputs the navigation mark type label corresponding to the navigation mark structure topology skeleton diagram based on the mapping relationship between the structure feature vector and the navigation mark type label, which is used as the navigation mark type identification result. Based on the navigation mark type identification result, the standard navigation mark geometric parameters of this navigation mark type are determined. The standard navigation mark geometric parameters include the standard top diameter parameter and the standard support height parameter. Obtain the top circle radius parameter from the top shape fitting parameters of the navigation mark and the vertical distance parameter between the end point coordinates and the beginning point coordinates of the support structure from the vertical support shape fitting parameters of the navigation mark, and use them as the measured top radius and measured support height of the navigation mark; The dimensional fit of the beacon top is calculated based on the ratio of the standard top diameter parameter to the measured top radius of the beacon, and the confidence weight for beacon top identification is calculated based on the fit. The dimensional fit of the support structure is calculated based on the ratio of the standard support height parameter to the measured support height of the beacon, and the confidence weight for support structure identification is calculated based on the fit. The confidence weights of the top of the navigation mark and the support structure are used as weighting factors for each positioning coordinate in the precise positioning information of the navigation mark. The geographic spatial coordinates and elevation information of the navigation mark are then adjusted to generate the precise positioning information of the navigation mark after type correction.

9. The machine vision-based precise positioning method for navigation marks according to claim 1, characterized in that, The method further includes: Perform multi-navigational beacon instance detection processing on the same frame of water surface image unit from the UAV perspective in the UAV aerial image data set of navigational beacon water area, extract all navigational beacon instances existing in the UAV perspective water surface image frame unit, and obtain the navigational beacon structure topology skeleton diagram corresponding to each navigational beacon instance. Extract the set of navigational beacon top skeleton nodes from the navigational beacon structure topology skeleton diagram corresponding to each navigational beacon instance, calculate the navigational beacon top center pixel coordinates of each navigational beacon instance, and obtain the set of multi-navigational beacon top center coordinates. The spatial distance between the top center coordinates of the multiple navigation marks is calculated. The pixel Euclidean distance between the top center pixel coordinates of any two navigation mark instances is calculated to generate a navigation mark spacing matrix. The relative orientation between navigation marks is calculated for the set of top center coordinates of the multiple navigation marks. The direction angle of the line connecting the top center pixel coordinates of any two navigation mark instances is calculated to generate a relative orientation matrix between navigation marks. Based on the navigation mark spacing matrix, the relative orientation matrix between navigation marks, and the initial spatial distance estimate calculated from the navigation mark geospatial coordinate set, a spatial association description graph is constructed between navigation marks. The weight of the graph edges is jointly determined by the spatial association strength in the image domain and the initial distance estimate in the object space. The spatial association description graph between navigation marks uses each navigation mark instance as a graph node and navigation mark pairs whose spatial distance between navigation marks is less than a preset navigation mark association distance threshold as graph edges. The weight of the graph edges is jointly determined by the spatial distance between navigation marks and the relative orientation between navigation marks. The spatial relationship description graph between navigation marks corresponding to the frame units of multi-frame UAV-view water surface images is subjected to temporal fusion processing. Graph nodes and graph edges that appear stably in consecutive frames are retained, while isolated graph nodes that appear only in a single frame are removed to generate a spatial relationship topology graph between navigation marks. Navigation mark instances corresponding to all graph nodes are extracted from the spatial relationship topology graph between navigation marks. The geospatial coordinates of each navigation mark instance are obtained through the spatial intersection solution processing to generate a set of geospatial coordinates of navigation marks. Extract the two navigation mark instances connected by each graph edge from the spatial association topology graph between navigation marks, and obtain the spatial distance between the two navigation mark instances and the relative orientation between the navigation marks as graph edge constraints; Based on the set of geospatial coordinates of the navigation marks, the map edge constraints, and the imaging geometric relationships of each navigation mark instance, a spatial consistency optimization objective function is constructed. The spatial consistency optimization objective function represents the cumulative deviation between the pixel spacing and relative orientation of the images formed after back-projecting the currently optimized geospatial coordinates of each navigation mark to each original image coordinate system and the image domain observations recorded in the map edge constraints, while also incorporating the smoothness constraint of the spatial position of the navigation marks. The spatial consistency optimization objective function is solved by iterative minimization. In each iteration, the geospatial coordinates of each navigation mark instance are adjusted to gradually reduce the cumulative deviation until the cumulative deviation is less than the preset convergence threshold, thus obtaining the optimized set of navigation mark geospatial coordinates. Obtain the navigation beacon tilt state parameters for each navigation beacon instance, associate the navigation beacon tilt state parameters with the corresponding optimized navigation beacon geospatial coordinates, and perform elevation direction correction processing on the optimized navigation beacon geospatial coordinates based on the navigation beacon tilt state parameters to obtain optimized navigation beacon elevation information for each navigation beacon instance. Combine the optimized navigation beacon geospatial coordinate set and the optimized navigation beacon elevation information into globally optimized precise navigation beacon positioning information.

10. A machine vision-based precise positioning system for navigational aids, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the machine vision-based beacon positioning method according to any one of claims 1 to 9 by executing the machine executable instructions.

Images