Drone tilt photography-based intelligent monitoring method and system for a river basin

Abstract

This invention relates to the field of watershed topographic monitoring technology, providing a method and system for intelligent watershed monitoring based on UAV oblique photography. The method includes the following steps: Based on a watershed digital elevation model, the flight altitude is dynamically adjusted according to topographic undulations to acquire oblique photographic images; potential feature points are initially screened on each image, and the quality index of each potential feature point is determined based on neighborhood conditions and elevation information; the final retained feature points are determined based on a quadtree algorithm, topographic conditions, and the quality index; the texture information of the surrounding images of the final retained feature points is encoded, and this encoded information is used for comparison and matching in different images; feature point matching is performed, and the matching confidence of each feature point is determined. When the matching confidence is lower than a confidence threshold, dynamic accuracy compensation is performed based on structure tensor and texture gradient to determine the 3D model. This invention can promptly detect and correct errors that occur during the matching process, improving the accuracy and quality of the 3D model.

Description

Technical Field

[0001] This invention relates to the field of watershed topographic monitoring technology, specifically to a watershed intelligent monitoring method and system based on UAV oblique photography. Background Technology

[0002] Accurate 3D reconstruction of watershed topography is a key technological foundation for fields such as hydrological analysis, flood simulation, soil and water conservation, and geological disaster early warning. In recent years, the combination of UAV technology and oblique photogrammetry has made it possible to rapidly acquire multi-view, high-resolution image data of watershed surfaces, gradually replacing traditional manual surveys and satellite remote sensing, demonstrating significant advantages in efficiency and cost. However, when applying UAV oblique photogrammetry to large-scale, topographically complex watershed scenarios, its technological bottlenecks become increasingly apparent. Specifically, traditional UAV flight path planning often employs fixed flight altitudes and uniform gridded flight paths, a model that cannot adapt to the dramatic topographic undulations of watersheds. Furthermore, watershed surfaces often contain numerous areas with repetitive textures (such as dense vegetation and homogeneous water surfaces) or textureless areas (such as smooth rocks), making traditional 3D reconstruction algorithms based on motion reconstruction structures and multi-view stereos prone to mismatches and matching drift during feature point extraction and matching stages. These errors are continuously propagated and accumulated in the data processing chain, ultimately leading to significant systematic geometric errors in the generated 3D model. Therefore, a watershed intelligent monitoring method and system based on UAV oblique photogrammetry is needed to address these problems. Summary of the Invention

[0003] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a method and system for intelligent watershed monitoring based on UAV oblique photography, so as to solve the problems existing in the above-mentioned background technology.

[0004] This invention is implemented as follows: a method for intelligent watershed monitoring based on UAV oblique photography, the method comprising the following steps:

[0005] Based on the watershed digital elevation model, the flight path is generated using the Hilbert spiral curve, and the flight altitude is dynamically adjusted according to the terrain undulations to collect images by oblique photography.

[0006] Potential feature points are initially screened out on each image, and the quality index of each potential feature point is determined based on neighborhood conditions and elevation information;

[0007] All potential feature points are projected onto a two-dimensional horizontal plane, and the final retained feature points are determined based on the quadtree algorithm, terrain conditions, and quality index.

[0008] The surrounding image texture information of the final retained feature points is encoded, and the encoded information is used for comparison and matching in different images;

[0009] Feature point matching is performed to determine the matching confidence of each feature point. When the matching confidence is lower than the confidence threshold, dynamic accuracy compensation is performed based on the structural tensor and texture gradient to determine the 3D model.

[0010] As a further aspect of the present invention, the step of initially screening potential feature points on each image and determining the quality index of each potential feature point based on neighborhood conditions and elevation information specifically includes:

[0011] An improved FAST corner detection algorithm is used to initially screen out several potential feature points on each image;

[0012] Determine the local neighborhood of each potential feature point. The local neighborhood is a spherical neighborhood with a fixed radius defined in three-dimensional space, centered on the potential feature point.

[0013] Statistically analyze the elevation values ​​of all other potential feature points in the local neighborhood, calculate the standard deviation σ of these elevation values, and determine the maximum elevation Hmax and the minimum elevation Hmin among them;

[0014] Calculate the quality index Q for each potential feature point, Q = λ × (Hmax - Hmin) / (σ × N represents the number of potential feature points in the local neighborhood, and λ is the weighting coefficient.

[0015] As a further aspect of the present invention, the step of determining the final retained feature points based on the quadtree algorithm, terrain conditions, and quality index specifically includes:

[0016] The terrain is determined based on the digital elevation model. A first point density and a second point density are set for each type of terrain, with the first point density being greater than the second point density.

[0017] The entire watershed is recursively divided into multiple rectangular cells of different sizes based on the quadtree algorithm, so that the number of feature points in each rectangular cell meets the first point density requirement.

[0018] Sort the potential feature points in each rectangular cell in descending order of quality index, determine the feature point with the highest quality index, remove feature points in the neighborhood of the feature point whose quality index is lower than the quality threshold, and delete the feature point with the highest quality index from the sort. Repeat this step until the number of feature points in the rectangular cell meets the second point density requirement.

[0019] As a further aspect of the present invention, the step of performing feature point matching and determining the matching confidence of each feature point specifically includes:

[0020] Based on the encoded information, bidirectional nearest neighbor matching is performed among all images to initially establish a set of correspondences between feature points in different images.

[0021] The matching results are used as observations and substituted into the bundle adjustment model of motion reconstruction structure for overall optimization calculation to solve for the exterior orientation elements of all images and the three-dimensional coordinates of all feature points.

[0022] Calculate the reprojection error, matching consistency, and geometric consistency for each feature point to determine the matching confidence.

[0023] As a further aspect of the present invention, the step of performing dynamic accuracy compensation based on structural tensor and texture gradient specifically includes:

[0024] Principal component analysis is performed on the 3D coordinates of all feature points in the neighborhood of each feature point to calculate the covariance matrix. Eigenvalue decomposition is then performed on the covariance matrix to determine the largest eigenvalue and its corresponding eigenvector, thus obtaining the structure tensor. ;

[0025] Among all images that match the feature points, the one with the highest resolution and the most upright angle is selected as the main reference image. On the main reference image, a pixel block is extracted with the location of the feature point as the center, and the gradient values ​​of the pixel block in the X and Y directions are calculated to determine the texture gradient. ;

[0026] Calculate the three-dimensional coordinate compensation amount ΔP = α × (1 - e) for the feature point. {-βt} )×( • The coordinates of the feature points are corrected, where α is the compensation coefficient and t is the number of iterations.

[0027] As a further aspect of the present invention, when dynamically adjusting the flight altitude according to the terrain undulations, h = k × (DEM / sinθ) + h0, where h is the real-time flight altitude, k is the terrain adaptation coefficient, DEM is the standard deviation of the digital elevation model within the flight area, θ is the camera pitch angle, and h0 is the reference flight altitude.

[0028] Another objective of this invention is to provide a watershed intelligent monitoring system based on unmanned aerial vehicle (UAV) oblique photography, the system comprising:

[0029] The flight path determination module is used to generate flight paths based on the watershed digital elevation model, using Hilbert spiral curves, dynamically adjusting the flight altitude according to the terrain undulations, and acquiring images through oblique photography.

[0030] The quality index calculation module is used to initially screen potential feature points on each image and determine the quality index of each potential feature point based on neighborhood conditions and elevation information.

[0031] The feature point filtering module is used to project all potential feature points onto a two-dimensional horizontal plane and determine the final retained feature points based on the quadtree algorithm, terrain conditions, and quality index.

[0032] The texture information encoding module is used to encode the texture information of the surrounding image of the final retained feature points. The encoded information is used for comparison and matching in different images.

[0033] The dynamic compensation module is used to perform feature point matching, determine the matching confidence of each feature point, and when the matching confidence is lower than the confidence threshold, perform dynamic accuracy compensation based on the structural tensor and texture gradient to determine the 3D model.

[0034] As a further aspect of the present invention, the quality index calculation module includes:

[0035] The latent feature point unit is used to initially screen out a number of latent feature points on each image using the improved FAST corner detection algorithm.

[0036] The local neighborhood determination unit is used to determine the local neighborhood of each potential feature point. The local neighborhood is a spherical neighborhood with a fixed radius defined in three-dimensional space, centered on the potential feature point.

[0037] The elevation determination unit is used to statistically analyze the elevation values ​​of all other potential feature points in the local neighborhood, calculate the standard deviation σ of these elevation values, and determine the maximum elevation Hmax and the minimum elevation Hmin among them.

[0038] The quality index calculation unit is used to calculate the quality index Q for each potential feature point, Q = λ × (Hmax - Hmin) / (σ × ... N represents the number of potential feature points in the local neighborhood, and λ is the weighting coefficient.

[0039] As a further embodiment of the present invention, the feature point filtering module includes:

[0040] The point density preset unit is used to determine the terrain conditions based on the digital elevation model and set a first point density and a second point density for each type of terrain, with the first point density being greater than the second point density.

[0041] The rectangular grid segmentation unit is used to recursively divide the entire watershed area into multiple rectangular cells of different sizes based on the quadtree algorithm, so that the number of feature points in each rectangular cell meets the first point density requirement.

[0042] The feature point filtering unit is used to sort the potential feature points in each rectangular cell in descending order of quality index, determine the feature point with the highest quality index, remove feature points in the neighborhood of the feature point whose quality index is lower than the quality threshold, delete the feature point with the highest quality index from the sort, and repeat this step until the number of feature points in the rectangular cell meets the second point density requirement.

[0043] As a further aspect of the present invention, the dynamic compensation module includes:

[0044] The preliminary matching unit is used to perform bidirectional nearest neighbor matching among all images based on the encoded information, and to initially establish a set of correspondences of feature points among different images.

[0045] The matching optimization unit is used to take the matching results as observations and substitute them into the bundle adjustment model of the motion restoration structure for overall optimization calculation, so as to solve the exterior orientation elements of all images and the three-dimensional coordinates of all feature points.

[0046] The matching confidence unit is used to calculate the reprojection error, matching consistency, and geometric consistency of each feature point to determine the matching confidence.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] This invention dynamically adjusts the flight altitude based on terrain undulations to acquire oblique photographic images, better adapting to complex terrain and enabling UAVs to obtain more comprehensive and accurate image data under various terrain conditions. After initially screening potential feature points on each image, the quality index of each potential feature point is determined based on neighborhood conditions and elevation information. Finally, based on a quadtree algorithm, terrain conditions, and the quality index, the final retained feature points are determined, accurately selecting higher-quality feature points and reducing mismatches caused by repetitive or textureless areas, thus improving the accuracy of feature point extraction. Furthermore, dynamic accuracy compensation is performed based on structural tensors and texture gradients, allowing for timely detection and correction of errors during the matching process, preventing errors from propagating and accumulating in the data processing chain, and effectively reducing systematic geometric errors in the generated 3D model. Attached Figure Description

[0049] Figure 1 This is a flowchart of a watershed intelligent monitoring method based on UAV oblique photography.

[0050] Figure 2 This is a flowchart for determining the quality index in a watershed intelligent monitoring method based on UAV oblique photography.

[0051] Figure 3 This is a flowchart illustrating the process of determining the final retained feature points in a watershed intelligent monitoring method based on UAV oblique photography.

[0052] Figure 4 This is a flowchart for determining the matching confidence level in a watershed intelligent monitoring method based on UAV oblique photography.

[0053] Figure 5 This is a flowchart illustrating the dynamic accuracy compensation process in a watershed intelligent monitoring method based on UAV oblique photography.

[0054] Figure 6 This is a schematic diagram of a watershed intelligent monitoring system based on UAV oblique photography. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0056] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0057] like Figure 1 As shown in the figure, this invention provides a method for intelligent watershed monitoring based on UAV oblique photography, the method comprising the following steps:

[0058] S100, based on the watershed digital elevation model, uses Hilbert spiral curves to generate flight paths, dynamically adjusts flight altitude according to terrain undulations, and performs oblique photography to acquire images.

[0059] S200 initially screens out potential feature points on each image and determines the quality index of each potential feature point based on neighborhood conditions and elevation information.

[0060] S300 projects all potential feature points onto a two-dimensional horizontal plane and determines the final retained feature points based on the quadtree algorithm, terrain conditions, and quality index.

[0061] S400 encodes the surrounding image texture information of the final retained feature points. The encoded information is used for comparison and matching in different images.

[0062] S500 performs feature point matching, determines the matching confidence of each feature point, and when the matching confidence is lower than the confidence threshold, performs dynamic accuracy compensation based on the structural tensor and texture gradient to determine the 3D model.

[0063] In this embodiment of the invention, a digital elevation model (DEM) of the watershed is first pre-set, and a Hilbert spiral curve is used to plan the flight path. Compared with the traditional zigzag flight path, this spiral path has smoother turns, effectively reducing the rotation and distortion of the images and laying a better foundation for subsequent feature matching. The flight altitude is dynamically adjusted according to the terrain undulations, with the real-time flight altitude h = k × (DEM / sinθ) + h0. This adapts to watersheds with drastic terrain undulations, where k is a terrain adaptation coefficient ranging from 1.2 to 1.5, DEM is the standard deviation of the DEM within the flight area, θ is the camera pitch angle, and h0 is the reference flight altitude. A 70% forward overlap rate and an 80% side overlap rate are also set, meaning that 70%-80% of the content in two adjacent images overlaps, ensuring that any ground point can be seen in multiple images, providing sufficient data for high-precision 3D positioning by the computer. Then, the acquired images are pre-processed, including noise reduction, distortion correction, and color balancing. Next, potential feature points are initially screened on each image, and the quality index of each potential feature point is determined. All potential feature points are projected onto a two-dimensional horizontal plane, and the final retained feature points are determined based on the quadtree algorithm, terrain conditions, and quality index. This achieves accurate and efficient screening of the most stable and unique feature points from massive image data, reducing mismatches caused by repetitive or textureless regions and improving the accuracy of feature point extraction. Then, for the final retained feature points, the surrounding image texture information is encoded using FAST-HE (high-speed, high-efficiency) feature descriptors for rapid and accurate comparison and matching in different images. Finally, feature point matching is performed, determining the matching confidence of each feature point. When the matching confidence falls below a confidence threshold, dynamic accuracy compensation is performed. This dynamic compensation mechanism can promptly detect and correct errors that occur during the matching process, preventing errors from continuously propagating and accumulating in the data processing chain, effectively reducing systematic geometric errors in the generated 3D model, thereby improving the accuracy and quality of the 3D model.

[0064] like Figure 2 As shown, in a preferred embodiment of the present invention, the steps of initially screening potential feature points on each image and determining the quality index of each potential feature point based on neighborhood conditions and elevation information specifically include:

[0065] S201 uses an improved FAST corner detection algorithm to initially screen out several potential feature points on each image.

[0066] S202, determine the local neighborhood of each potential feature point;

[0067] S203, statistically analyze the elevation values ​​of all other potential feature points in the local neighborhood, calculate the standard deviation σ of these elevation values, and determine the maximum elevation Hmax and the minimum elevation Hmin among them;

[0068] S204, calculate the quality index Q for each potential feature point, Q=λ×(Hmax-Hmin) / (σ× N represents the number of potential feature points in the local neighborhood, and λ is the weighting coefficient.

[0069] In this embodiment of the invention, to ensure that no potentially useful features are missed, an improved FAST (Features from Accelerated Segment Test) corner detection algorithm is used to quickly screen out a large number of potential feature points on each image. Then, a local neighborhood is determined for each potential feature point. The local neighborhood is a spherical neighborhood with a fixed radius (e.g., 10 meters) defined in three-dimensional space, centered on the potential feature point. The elevation values ​​of all other potential feature points within the local neighborhood are then statistically analyzed, and the standard deviation σ of these elevation values ​​is calculated. The maximum elevation Hmax and minimum elevation Hmin are determined, thus yielding the quality index Q: Q = λ × (Hmax - Hmin) / (σ × ... (Hmax-Hmin) reflects the topographic relief. The larger the value, the more unique the spatial location of the point, indicating that it is located in an area with drastic topographic changes and should receive a higher score. σ reflects topographic consistency. The smaller the value, the more consistent and smooth the elevation changes within the neighborhood, indicating that the point is not an isolated noise point and is more reliable. This represents the point density factor, which is used to balance the effects of uneven point density. In areas where points are too densely packed, the weight of individual points is appropriately reduced to avoid duplicate point selection. The weighting coefficient is used to fine-tune the sensitivity under different terrain conditions.

[0070] like Figure 3 As shown, in a preferred embodiment of the present invention, the step of determining the final retained feature points based on the quadtree algorithm, terrain conditions, and quality index specifically includes:

[0071] S301, determine the terrain conditions based on the digital elevation model, and set the first point density and second point density for each type of terrain;

[0072] S302, based on the quadtree algorithm, recursively divides the entire watershed into multiple rectangular cells of different sizes, so that the number of feature points in each rectangular cell meets the first point density requirement;

[0073] S303, sort the potential feature points in each rectangular cell in descending order of quality index, determine the feature point with the highest quality index, remove feature points in the neighborhood of the feature point whose quality index is lower than the quality threshold, delete the feature point with the highest quality index from the sort, and repeat this step until the number of feature points in the rectangular cell meets the second point density requirement.

[0074] In this embodiment of the invention, the terrain conditions are determined based on the digital elevation model. These terrain conditions include complex areas and flat areas. A first point density and a second point density are set for each terrain type, with the first point density being greater than the second point density. It is easy to understand that the point density in complex areas is greater than that in flat areas. Then, the entire watershed area is recursively divided into multiple rectangular cells of different sizes using a quadtree algorithm, ensuring that the number of feature points in each rectangular cell meets the first point density requirement for the corresponding terrain. Next, the potential feature points in each rectangular cell are sorted in descending order of quality index. The feature point with the highest quality index is determined, and feature points in their neighborhood with a quality index lower than a fixed quality threshold are removed. The feature point with the highest quality index is then deleted from the sorting. This deletion only affects the retention of the feature point with the highest quality index, thus updating the sorting and facilitating the repeated execution of step S303 until the number of feature points in the rectangular cell meets the second point density requirement.

[0075] like Figure 4 As shown, in a preferred embodiment of the present invention, the step of performing feature point matching and determining the matching confidence of each feature point specifically includes:

[0076] S501, based on the coding information, perform bidirectional nearest neighbor matching among all images to initially establish a set of correspondences of feature points among different images;

[0077] S502, the matching results are used as observations and substituted into the bundle adjustment model of motion reconstruction structure for overall optimization calculation, and the exterior orientation elements of all images and the three-dimensional coordinates of all feature points are solved;

[0078] S503, calculate the reprojection error, matching consistency, and geometric consistency for each feature point, and determine the matching confidence.

[0079] In this embodiment of the invention, the previously generated FAST-HE feature descriptor is used to perform bidirectional nearest neighbor matching among all images. That is, for a feature point in image A, the most similar feature point in image B is searched, and vice versa. A match is considered successful only when the two are optimally matched, thus initially establishing a set of correspondences between feature points in different images. Then, bundle adjustment is used for preliminary optimization. Specifically, the matching results are used as observations and substituted into the bundle adjustment model of motion-reconstructed structures for overall optimization calculation, solving for the high-precision exterior orientation elements (position and attitude) of all images and the preliminary three-dimensional coordinates of all feature points. Next, for each successfully matched feature point i, the reprojection error J1, matching consistency J2, and geometric consistency J3 are determined. When determining the reprojection error, the three-dimensional coordinates of the feature point are reprojected back onto all images that observed it according to the solved image exterior orientation elements. The average pixel distance between the projected point and the original image point is calculated. The smaller the error, the higher the confidence level. When determining matching consistency, the number of different images that successfully matched the feature point is identified. The more images that are independently observed and matched, the higher the reliability and confidence. When determining geometric consistency, the standard deviation of the reprojection error of other neighboring feature points within the local neighborhood of the feature point is calculated. The smaller this value, the better the overall geometric consistency of the region, and the higher the confidence of the point. Matching confidence = (J2 / Jmax) × (1 / (1+J1)) × (1 / (1+J3)), where Jmax represents the maximum number of observations among all feature points.

[0080] like Figure 5 As shown in the preferred embodiment of the present invention, the step of performing dynamic accuracy compensation based on structural tensor and texture gradient specifically includes:

[0081] S504: Perform principal component analysis on the 3D coordinates of all feature points in the neighborhood of each feature point, calculate the covariance matrix, perform eigenvalue decomposition on the covariance matrix, determine the largest eigenvalue and the corresponding eigenvector, and obtain the structure tensor. ;

[0082] S505, among all images that match the feature points, select the one with the highest resolution and the most upright angle as the main reference image; on the main reference image, extract a pixel block centered on the location of the feature point, calculate the gradient values ​​of the pixel block in the X and Y directions, and determine the texture gradient. ;

[0083] S506, calculate the three-dimensional coordinate compensation amount ΔP=α×(1-e) of the feature point. {-βt} )×( • The coordinates of the feature points are corrected.

[0084] In this embodiment of the invention, when the matching confidence of a certain feature point is lower than the confidence threshold, principal component analysis is performed on the coordinates of all three-dimensional points in the neighborhood centered on the feature point i. The covariance matrix of the point set is calculated, and eigenvalue decomposition is performed on the covariance matrix to obtain three eigenvalues ​​(λ1, λ2, λ3) and corresponding eigenvectors. The largest eigenvalue λ1 characterizes the intensity of change in the neighborhood along the main spatial extension direction, and its corresponding eigenvector indicates that direction. The structure tensor ∇S is defined by these eigenvectors and eigenvalues, quantifying the three-dimensional geometric structure features and directionality of the point's location. Next, among all images matching the feature point, the one with the highest resolution and the most positive angle is selected as the master reference image. On the master reference image, a pixel block is extracted centered on the feature point's location, and the gradient values ​​Gx and Gy of the pixel block (e.g., 5x5 or 7x7) in the X and Y directions are calculated to determine the texture gradient ∇T. ∇T is determined by the gradient magnitude... Defined together with the gradient direction arctan(Gy / Gx), it quantifies the texture edge intensity and direction of the point in the 2D image. Finally, the 3D coordinate compensation amount ΔP = α × (1 - e^(-Gx)) of the feature point is calculated. {-βt} )×( • ), α is the compensation coefficient, controlling the overall compensation intensity, t is the iteration number, β is a constant coefficient, (1-e {-βt} () is a time decay factor to ensure that the compensation amount is smoothly introduced during the optimization process, avoiding oscillations. • The dot product operation between the structure tensor and the texture gradient essentially evaluates the consistency between the 3D geometry and the 2D image texture. This value is maximized when the 3D edge direction aligns with the 2D image edge direction, resulting in the most precise and reliable compensation direction. The calculated 3D vector ΔP is then added to the current coordinates of feature point i to complete the position correction.

[0085] In this embodiment of the invention, the compensated and corrected feature point coordinates are used as new initial values ​​and re-input into the bundle adjustment for the next iteration of optimization. Confidence is continuously monitored, compensation is triggered, and smoothing is performed until the reprojection error of the entire model converges globally, and the average confidence of all feature points is higher than a set threshold. At this point, the final high-precision 3D model is output.

[0086] like Figure 6 As shown, this embodiment of the invention also provides a watershed intelligent monitoring system based on UAV oblique photography, the system comprising:

[0087] The flight path determination module 100 is used to generate a flight path based on the watershed digital elevation model, using a Hilbert spiral curve, dynamically adjusting the flight altitude according to the terrain undulations, and acquiring images through oblique photography.

[0088] The quality index calculation module 200 is used to initially screen potential feature points on each image and determine the quality index of each potential feature point based on neighborhood conditions and elevation information.

[0089] The feature point filtering module 300 is used to project all potential feature points onto a two-dimensional horizontal plane and determine the final retained feature points based on the quadtree algorithm, terrain conditions, and quality index.

[0090] The texture information encoding module 400 is used to encode the surrounding image texture information of the final retained feature points. The encoded information is used for comparison and matching in different images.

[0091] The dynamic compensation module 500 is used to perform feature point matching, determine the matching confidence of each feature point, and when the matching confidence is lower than the confidence threshold, perform dynamic accuracy compensation based on the structural tensor and texture gradient to determine the 3D model.

[0092] In a preferred embodiment of the present invention, the quality index calculation module 200 includes:

[0093] The latent feature point unit is used to initially screen out a number of latent feature points on each image using the improved FAST corner detection algorithm.

[0094] The local neighborhood determination unit is used to determine the local neighborhood of each potential feature point. The local neighborhood is a spherical neighborhood with a fixed radius defined in three-dimensional space, centered on the potential feature point.

[0095] The elevation determination unit is used to statistically analyze the elevation values ​​of all other potential feature points in the local neighborhood, calculate the standard deviation σ of these elevation values, and determine the maximum elevation Hmax and the minimum elevation Hmin among them.

[0096] The quality index calculation unit is used to calculate the quality index Q for each potential feature point, Q = λ × (Hmax - Hmin) / (σ × ... N represents the number of potential feature points in the local neighborhood, and λ is the weighting coefficient.

[0097] In a preferred embodiment of the present invention, the feature point filtering module 300 includes:

[0098] The point density preset unit is used to determine the terrain conditions based on the digital elevation model and set a first point density and a second point density for each type of terrain, with the first point density being greater than the second point density.

[0099] The rectangular grid segmentation unit is used to recursively divide the entire watershed area into multiple rectangular cells of different sizes based on the quadtree algorithm, so that the number of feature points in each rectangular cell meets the first point density requirement.

[0100] The feature point filtering unit is used to sort the potential feature points in each rectangular cell in descending order of quality index, determine the feature point with the highest quality index, remove feature points in the neighborhood of the feature point whose quality index is lower than the quality threshold, delete the feature point with the highest quality index from the sort, and repeat this step until the number of feature points in the rectangular cell meets the second point density requirement.

[0101] In a preferred embodiment of the present invention, the dynamic compensation module 500 includes:

[0102] The preliminary matching unit is used to perform bidirectional nearest neighbor matching among all images based on the encoded information, and to initially establish a set of correspondences of feature points among different images.

[0103] The matching optimization unit is used to take the matching results as observations and substitute them into the bundle adjustment model of the motion restoration structure for overall optimization calculation, so as to solve the exterior orientation elements of all images and the three-dimensional coordinates of all feature points.

[0104] The matching confidence unit is used to calculate the reprojection error, matching consistency, and geometric consistency of each feature point to determine the matching confidence.

[0105] The above description only details the preferred embodiments of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0106] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0107] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0108] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the disclosure in the specification and embodiments. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the claims.

Claims

1. A watershed intelligent monitoring method based on unmanned aerial vehicle oblique photography, characterized in that, The method includes the following steps: Based on the watershed digital elevation model, the flight path is generated using the Hilbert spiral curve, and the flight altitude is dynamically adjusted according to the terrain undulations to collect images by oblique photography. Potential feature points are initially screened out on each image, and the quality index of each potential feature point is determined based on neighborhood conditions and elevation information; All potential feature points are projected onto a two-dimensional horizontal plane, and the final retained feature points are determined based on the quadtree algorithm, terrain conditions, and quality index. The surrounding image texture information of the final retained feature points is encoded, and the encoded information is used for comparison and matching in different images; Feature point matching is performed, and the matching confidence of each feature point is determined. When the matching confidence is lower than the confidence threshold, dynamic accuracy compensation is performed based on the structural tensor and texture gradient to determine the 3D model. In the step of screening out potential feature points on each image and determining the quality index of each potential feature point based on neighborhood condition and elevation information, the step specifically comprises: screening out a plurality of potential feature points on each image; determining a local neighborhood of each potential feature point, wherein the local neighborhood is a spherical neighborhood with a fixed radius centered on the potential feature point in a three-dimensional space; counting elevation values of all other potential feature points in the local neighborhood, calculating a standard deviation σ of the elevation values, and determining a maximum elevation Hmax and a minimum elevation Hmin among the elevation values; and calculating a quality index Q of each potential feature point, Q = λ × (Hmax - Hmin) / (σ × ), wherein N represents the number of potential feature points in the local neighborhood, and λ is a weight coefficient. ) The step of dynamic accuracy compensation based on structural tensor and texture gradient specifically includes: performing principal component analysis on the three-dimensional coordinates of all feature points in the neighborhood of each feature point, calculating the covariance matrix, performing eigenvalue decomposition on the covariance matrix, determining the maximum eigenvalue and the corresponding eigenvector, and obtaining the structural tensor ∇S; selecting the image with the highest resolution and the most positive angle from all images that match the feature point as the main reference image; extracting a pixel block centered on the feature point on the main reference image, calculating the gradient values ​​of the pixel block in the X and Y directions, and determining the texture gradient ∇T; and calculating the three-dimensional coordinate compensation amount of the feature point. The coordinates of the feature points are corrected, where α is the compensation coefficient, t is the number of iterations, and β is a constant coefficient.

2. The intelligent watershed monitoring method based on UAV oblique photography according to claim 1, characterized in that, The steps for determining the final retained feature points based on the quadtree algorithm, terrain conditions, and quality index specifically include: The terrain is determined based on the digital elevation model. A first point density and a second point density are set for each type of terrain, with the first point density being greater than the second point density. The entire watershed is recursively divided into multiple rectangular cells of different sizes based on the quadtree algorithm, so that the number of feature points in each rectangular cell meets the first point density requirement. Sort the potential feature points in each rectangular cell in descending order of quality index, determine the feature point with the highest quality index, remove feature points in the neighborhood of the feature point whose quality index is lower than the quality threshold, and delete the feature point with the highest quality index from the sort. Repeat this step until the number of feature points in the rectangular cell meets the second point density requirement.

3. The intelligent watershed monitoring method based on UAV oblique photography according to claim 1, characterized in that, The step of performing feature point matching and determining the matching confidence of each feature point specifically includes: Based on the encoded information, bidirectional nearest neighbor matching is performed among all images to initially establish a set of correspondences between feature points in different images. The matching results are used as observations and substituted into the bundle adjustment model of motion reconstruction structure for overall optimization calculation to solve for the exterior orientation elements of all images and the three-dimensional coordinates of all feature points. Calculate the reprojection error, matching consistency, and geometric consistency for each feature point to determine the matching confidence.

4. The intelligent watershed monitoring method based on UAV oblique photography according to claim 1, characterized in that, When dynamically adjusting the flight altitude based on terrain undulations, h = k × (DEM / sinθ) + h0, where h is the real-time flight altitude, k is the terrain adaptation coefficient, DEM is the standard deviation of the digital elevation model within the flight area, θ is the camera pitch angle, and h0 is the reference flight altitude.

5. A watershed intelligent monitoring system based on UAV oblique photography, characterized in that, For implementing the watershed intelligent monitoring method based on UAV oblique photography as described in any one of claims 1 to 4, the system comprises: The flight path determination module is used to generate flight paths based on the watershed digital elevation model, using Hilbert spiral curves, dynamically adjusting the flight altitude according to the terrain undulations, and acquiring images through oblique photography. The quality index calculation module is used to initially screen potential feature points on each image and determine the quality index of each potential feature point based on neighborhood conditions and elevation information. The feature point filtering module is used to project all potential feature points onto a two-dimensional horizontal plane and determine the final retained feature points based on the quadtree algorithm, terrain conditions, and quality index. The texture information encoding module is used to encode the texture information of the surrounding image of the final retained feature points. The encoded information is used for comparison and matching in different images. The dynamic compensation module is used to perform feature point matching, determine the matching confidence of each feature point, and when the matching confidence is lower than the confidence threshold, perform dynamic accuracy compensation based on the structural tensor and texture gradient to determine the 3D model.

6. The intelligent watershed monitoring system based on UAV oblique photography according to claim 5, characterized in that, The quality index calculation module includes: The latent feature point unit is used to initially screen out a number of latent feature points on each image using the improved FAST corner detection algorithm. The local neighborhood determination unit is used to determine the local neighborhood of each potential feature point. The local neighborhood is a spherical neighborhood with a fixed radius defined in three-dimensional space, centered on the potential feature point. The elevation determination unit is used to statistically analyze the elevation values ​​of all other potential feature points in the local neighborhood, calculate the standard deviation σ of these elevation values, and determine the maximum elevation Hmax and the minimum elevation Hmin among them. The quality index calculation unit is used to calculate the quality index Q for each potential feature point, Q = λ × (Hmax - Hmin) / (σ × ... N represents the number of potential feature points in the local neighborhood, and λ is the weighting coefficient.

7. The intelligent watershed monitoring system based on UAV oblique photography according to claim 5, characterized in that, The feature point filtering module includes: The point density preset unit is used to determine the terrain conditions based on the digital elevation model and set a first point density and a second point density for each type of terrain, with the first point density being greater than the second point density. The rectangular grid segmentation unit is used to recursively divide the entire watershed area into multiple rectangular cells of different sizes based on the quadtree algorithm, so that the number of feature points in each rectangular cell meets the first point density requirement. The feature point filtering unit is used to sort the potential feature points in each rectangular cell in descending order of quality index, determine the feature point with the highest quality index, remove feature points in the neighborhood of the feature point whose quality index is lower than the quality threshold, delete the feature point with the highest quality index from the sort, and repeat this step until the number of feature points in the rectangular cell meets the second point density requirement.

8. The intelligent watershed monitoring system based on UAV oblique photography according to claim 5, characterized in that, The dynamic compensation module includes: The preliminary matching unit is used to perform bidirectional nearest neighbor matching among all images based on the encoded information, and to initially establish a set of correspondences of feature points among different images. The matching optimization unit is used to take the matching results as observations and substitute them into the bundle adjustment model of the motion restoration structure for overall optimization calculation, so as to solve the exterior orientation elements of all images and the three-dimensional coordinates of all feature points. The matching confidence unit is used to calculate the reprojection error, matching consistency, and geometric consistency of each feature point to determine the matching confidence.

Images