Extraction optimization method and system of ground points for water conservancy survey

Abstract

The present invention discloses a method and system for optimizing the extraction of ground points for water conservancy surveys, and relates to the field of data processing technology. The method comprises: using a camera to perform side-view photography of a target water conservancy survey area to obtain side-view photography data; marking the heights of buildings in the water conservancy survey area according to the side-view photography data, clustering the marked height data, and determining multiple clustering heights; planning multiple photography paths according to multiple flight altitude parameters, and performing overhead image acquisition of the target water conservancy survey area based on the multiple photography paths, outputting multiple layers of overhead photography data; converting the multiple layers of overhead photography data into point clouds, and then inputting them into a ground point extraction model for multi-scale optimization processing to obtain ground point extraction results. The method solves the technical problem of insufficient ground point extraction accuracy in water conservancy surveys in complex terrain and landforms in the prior art, and achieves the technical effect of improving the accuracy of ground point extraction and enhancing adaptability to different terrain and landforms.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a method and system for extracting and optimizing ground points for water conservancy survey. Background Art

[0002] As the foundational work for the development and utilization of water resources, water conservancy survey is crucial for ensuring the quality of planning, design, and construction of water conservancy projects. However, traditional water conservancy survey methods often face many challenges when faced with complex and ever-changing terrain and landforms. Accurate extraction of ground points is a key link in water conservancy surveys, which is directly related to the accuracy and feasibility of subsequent engineering designs. However, traditional ground point extraction methods, such as manual measurement and remote sensing image interpretation, are not only time-consuming and labor-intensive, but are also easily affected by various factors such as terrain and climate, resulting in low extraction accuracy. Therefore, how to improve the accuracy of ground point extraction in complex terrain and landforms and ensure the adaptability of the algorithm to various terrain and landforms has become a technical problem that needs to be solved urgently in the current field of water conservancy surveys. Summary of the Invention

[0003] The present application provides a method and system for optimizing the extraction of ground points for water conservancy surveys, which solves the technical problem of insufficient accuracy in the extraction of ground points for water conservancy surveys in complex terrain in the prior art.

[0004] A first aspect of the present application provides a method for optimizing extraction of ground points for water conservancy survey, the method comprising:

[0005] A camera is used to take a side-view camera of a target water conservancy survey area to obtain side-view camera data; the building heights in the water conservancy survey area are marked according to the side-view camera data, and clustered according to the marked height data to determine a plurality of cluster heights; a plurality of flight height parameters are set based on the plurality of cluster heights, a plurality of camera paths are planned according to the plurality of flight height parameters, and a top-view image of the target water conservancy survey area is collected based on the plurality of camera paths to output multi-layer top-view camera data; the multi-layer top-view camera data is converted into a point cloud and then input into a ground point extraction model for multi-scale optimization processing to obtain a ground point extraction result.

[0006] A second aspect of the present application provides a system for optimizing extraction of ground points for water conservancy surveys, the system comprising:

[0007] The first data acquisition module is used to use a camera to take a side-view camera of the target water conservancy survey area to obtain side-view camera data; the clustering module is used to mark the building heights in the water conservancy survey area according to the side-view camera data, cluster the marked height data according to the marked height data, and determine multiple cluster heights; the second data acquisition module is used to set multiple flight height parameters based on the multiple cluster heights, plan multiple camera paths according to the multiple flight height parameters, collect overhead images of the target water conservancy survey area based on the multiple camera paths, and output multi-layer overhead camera data; the optimization processing module is used to convert the multi-layer overhead camera data into a point cloud and input the converted data into a ground point extraction model for multi-scale optimization processing to obtain ground point extraction results.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0009] First, a camera is used to take a side-view camera of the target water conservancy survey area to obtain side-view camera data. Next, the building heights in the water conservancy survey area are marked based on the side-view camera data, and clustered according to the marked height data to determine multiple cluster heights. Then, multiple flight altitude parameters are set based on the multiple cluster heights, and multiple camera paths are planned according to the multiple flight altitude parameters. Based on the multiple camera paths, overhead images of the target water conservancy survey area are collected, and multi-layer overhead camera data is output. Finally, the multi-layer overhead camera data is converted into a point cloud and input into the ground point extraction model for multi-scale optimization processing to obtain the ground point extraction results. This solves the technical problem of insufficient ground point extraction accuracy in water conservancy surveys in complex terrain and landforms in the existing technology, and achieves the technical effect of improving the accuracy of ground point extraction and enhancing adaptability to different terrain and landforms. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0011] Figure 1 A flow chart of a method for optimizing the extraction of ground points for water conservancy surveys provided in an embodiment of the present application;

[0012] Figure 2 Schematic diagram of the structure of the ground point extraction and optimization system for water conservancy survey provided in an embodiment of the present application.

[0013] Description of reference numerals: first data acquisition module 11 , clustering module 12 , second data acquisition module 13 , optimization processing module 14 . DETAILED DESCRIPTION

[0014] This application solves the technical problem of insufficient accuracy in extracting ground points for water conservancy surveys in complex terrain in the prior art by providing an optimized method and system for extracting ground points for water conservancy surveys.

[0015] The following will be combined with the accompanying drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only some of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0016] It should be noted that the terms "including" and "having" are intended to cover non-exclusive inclusions. For example, a process, method, system, product or server that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or modules that are not clearly listed or are inherent to these processes, methods, products or devices.

[0017] Example 1, as Figure 1 As shown, the present application provides a method for optimizing the extraction of ground points for water conservancy survey, wherein the method includes:

[0018] The camera equipment is used to take side-view photos of the target water conservancy survey area to obtain side-view photo data.

[0019] Furthermore, a camera device is used to take side-view photos of the target water conservancy survey area. The camera device includes a first camera element and a second camera element. The first camera element is for horizontal side-view shooting, and the second camera element is for pitch and tilt angle side-view shooting.

[0020] The target water conservancy survey area is photographed from the side using a camera to obtain side view photographic data, wherein side view photographing means that the camera shoots in a manner parallel to the target water conservancy survey area, and the obtained image shows the depth structure of the area.

[0021] The camera device includes a first camera and a second camera. The first camera is configured to perform horizontal side view shooting, and its main function is to provide a flat, wide-angle view, which can clearly capture the overall layout of the water conservancy survey area, the appearance of buildings, ground features and other information; the horizontal viewing angle of the first camera can maximize the coverage of the target area, ensuring that a complete view of important facilities such as buildings, rivers, and dams can be captured. The second camera is configured to perform angled side view shooting at a pitch angle, that is, the camera can shoot the target area at a certain tilt angle to obtain a view from the side or obliquely above, which can better show the depth structure and details of the area, such as the height of the building, slope changes, etc. The two cameras with different shooting angles work together, and the horizontal view captured by the first camera and the pitch angle view captured by the second camera complement each other to provide more comprehensive regional data.

[0022] The building heights in the water conservancy survey area are marked according to the side-view camera data, and clustering is performed according to the marked height data to determine a plurality of cluster heights.

[0023] By analyzing the image information in the side-view camera data and using image processing technology or deep learning models, the buildings in the image are detected and the top height value of each building is marked. The labeling results can accurately reflect the relative height information of each building.

[0024] After the building heights are labeled, cluster analysis is performed based on this labeled height data. Specifically, all the labeled building heights are grouped according to a clustering algorithm, such as K-means or DBSCAN, to group buildings of similar heights into the same cluster. Through cluster analysis, multiple different cluster heights can be obtained, each of which represents the common characteristics of a group of buildings within a certain height range.

[0025] Furthermore, the heights of buildings in the water conservancy survey area are marked according to the side view camera data, and the method includes:

[0026] A building height recognition model is trained according to the YOLOv7 architecture; the building top heights of the side-view camera data are annotated according to the building height recognition model to obtain top height annotation results, and the top height annotation results are density clustered to output multiple cluster heights, where each cluster height is the highest value within the cluster group to which it belongs.

[0027] Preferably, the building height recognition model is trained using the YOLOv7 architecture. YOLOv7 is an efficient and highly accurate object detection algorithm suitable for real-time detection of buildings in images. YOLOv7 is trained using a large dataset of labeled building images, enabling it to automatically identify the location and features of buildings in images. During training, the model learns the morphological characteristics and positional relationships of buildings in images, enabling accurate identification of buildings in images. Next, the trained building height recognition model is used to detect and annotate buildings in the side-view camera data. The building height recognition model detects buildings and annotates the top height of each building in the image, generating top height annotation results. These top height annotation results are then subjected to density clustering. Density clustering algorithms (such as DBSCAN) can group buildings based on their top height data. The algorithm analyzes the height similarities between buildings and groups similar buildings together. Finally, based on the density clustering results, multiple cluster heights are output. Each cluster height corresponds to the top height of the highest building in the cluster group to which it belongs. Each cluster group represents a specific height range, and the buildings within this height range have similar height characteristics. These clustering results provide a basis for subsequent flight path planning and camera path setting.

[0028] A plurality of flight height parameters are set based on the plurality of cluster heights, a plurality of camera paths are planned according to the plurality of flight height parameters, a bird's-eye view image is collected for the target water conservancy survey area based on the plurality of camera paths, and multi-layer bird's-eye view camera data is output.

[0029] Based on the multiple cluster heights obtained from clustering, multiple flight altitude parameters are set to ensure that complete information of the water conservancy survey area is captured. For clusters of lower buildings, a lower flight altitude can be set to ensure clear capture of ground and building details; for clusters of taller buildings, a higher flight altitude can be set to avoid obstruction of the camera equipment by the buildings and ensure coverage of a wider area.

[0030] After setting the flight altitude parameters, multiple camera paths are planned according to these flight altitude parameters. Specifically, first, the target water conservancy survey area is divided into multiple sub-areas, each of which corresponds to a specific flight altitude, so that it can be optimized based on the clustered altitude data; for lower flight altitudes, path planning can adopt shorter linear or rectangular paths to obtain higher image resolution; for higher flight altitudes, longer paths can be used to ensure coverage of larger areas while avoiding path overlap. Path planning can be done in a grid, spiral, or rectangular manner; for larger water conservancy survey areas, grid path planning is used to ensure that each area is covered by dividing the area into regular grids; if higher image resolution is required in the center of the area, spiral path planning can be used to ensure that the area is scanned from different angles; and for areas where buildings are more evenly distributed, rectangular path planning can be used to cover the entire area and minimize path overlap.

[0031] For example, the collected building height data is clustered and analyzed. Assume the following clustering results: low-rise buildings, such as low houses and facilities, are clustered in a height range of 5 to 10 meters; medium-rise buildings, such as medium-sized water conservancy facilities, are clustered in a height range of 10 to 20 meters; and high-rise buildings, such as tall dams and bridges, are clustered in a height range of 20 to 40 meters. For low-rise buildings (5 to 10 meters), the lower flight altitude is set to 20 meters; for medium-rise buildings (10 to 20 meters), the flight altitude is set to 40 meters; and for high-rise buildings (20 to 40 meters), the higher flight altitude is set to 60 meters. Based on these flight altitude parameters, multiple camera paths are planned, and the survey area can be divided into multiple sub-areas, for example, the area can be divided into a 5x5 km grid. At low flight altitudes (20 meters), a shorter rectangular path with a 50-meter spacing is selected to ensure coverage of each sub-area and capture detailed images of buildings. At medium flight altitudes (40 meters), a longer rectangular path with a 100-meter spacing is used to ensure image capture over a wide area. At high flight altitudes (60 meters), a longer grid path with a 150-meter spacing is selected to reduce path overlap and ensure image capture coverage of the entire area. After completing path planning, the aircraft flies along the planned path and performs real-time image capture. The aircraft first flies along the low-altitude path to capture detailed overhead images of low-altitude buildings. Then, the aircraft flies at the set medium altitude to capture overhead images of the area with medium-altitude buildings. Finally, the aircraft flies at the high altitude to obtain overhead images of high-altitude buildings and the surrounding environment.

[0032] After planning the camera paths, actual flight and image acquisition are performed according to these paths. The camera equipment flies along the planned paths and captures overhead images in real time during flight. The overhead images capture detailed data such as ground information, building outlines, and facility layouts in the water conservancy survey area through the camera's bird's-eye view. Based on the overhead image acquisition of these multiple camera paths, multiple layers of overhead camera data are output. Each layer of overhead camera data reflects the regional image information at different flight altitudes. This data provides rich raw material for subsequent point cloud conversion, ground point extraction, and further data processing.

[0033] Furthermore, the method for collecting overhead images of the target water conservancy survey area based on the multiple camera paths includes:

[0034] Acquire multiple camera devices, where the number of the multiple camera devices corresponds to the multiple camera paths; connect the multiple camera devices to a data center as edge devices, and when the multiple camera devices respectively perform camera shooting according to the multiple camera paths, transmit the acquired real-time camera data to the data center through the edge devices for storage.

[0035] Preferably, multiple cameras are acquired, and the number of cameras corresponds to the planned multiple camera paths. Each camera path represents a specific flight trajectory, ensuring coverage of all parts of the target water conservancy survey area. Each camera is responsible for capturing overhead images along a specific path to obtain high-definition image data of the area along that path.

[0036] To ensure real-time data transmission and efficient data transmission, cameras are usually configured as edge devices. Each camera is connected to the data center as an edge device. That is, while capturing images, the camera will transmit the captured video or image data to the edge device in real time via a wireless network or other communication methods. Specifically, when the camera begins to fly and capture images according to multiple predetermined camera paths, each device will simultaneously transmit data in real time; the edge device collects data streams from each camera and stores the data in categories according to the path. In this way, all image data can be transmitted to the data center in real time during the flight, without having to wait until the flight ends to upload the data, ensuring the timeliness and integrity of the data.

[0037] After receiving all the real-time data, the data center stores it uniformly, typically compressing the data, converting storage formats, or performing preliminary data analysis to facilitate further ground point extraction, 3D modeling, or other survey and analysis tasks. Through the collaborative work of edge devices and the data center, image data collection, transmission, and storage across the entire water conservancy survey area can be completed efficiently and stably.

[0038] Furthermore, after the multi-layer overhead camera data is converted into a point cloud and before being input into the ground point extraction model, the multi-layer overhead camera data is registered based on the ICP algorithm, and the method includes:

[0039] Acquire a multi-layer point cloud dataset corresponding to the multi-layer overhead camera data for point cloud conversion; extract two adjacent layers of point cloud data groups based on the multi-layer point cloud dataset; obtain a feature vector representing point cloud characteristics, the feature vector including three gradient components, curvature and normal vector; perform a KD tree accelerated nearest neighbor search on the two adjacent layers of point cloud data groups based on the feature vector to obtain a corresponding registration transformation matrix between the two adjacent layers of point cloud data groups; register the multi-layer overhead camera data based on the registration transformation matrix to obtain a registered multi-layer point cloud dataset.

[0040] Before inputting the ground point extraction model into the multi-layered overhead camera data through point cloud conversion, it is necessary to register the data using the Iterative Closest Point (ICP) algorithm. First, the multi-layered overhead camera data, acquired from multiple flight altitudes and different viewing angles, is converted to point clouds, resulting in multiple corresponding point cloud datasets. Each layer of the overhead camera data, after conversion, represents the 3D spatial information of the target area at that specific flight altitude and viewing angle. Next, two adjacent layers of point cloud data are selected from the multiple point cloud datasets for registration. These two adjacent layers of point cloud data are derived from different flight altitudes and viewing angles, and therefore may exhibit spatial deviations or misalignments. For accurate registration, feature vectors are extracted from each point cloud dataset. These feature vectors consist of three gradient components, curvature, and normal vectors. These feature vectors describe the shape characteristics, curvature variations, and surface normal information of a local region in the point cloud data. The three gradient components capture the direction of surface variation, curvature reflects the degree of curvature, and normal vectors provide key information about the surface orientation. Afterwards, a KD tree (K-dimensional tree) is used to accelerate the nearest neighbor search. The KD tree is an efficient spatial data structure that can quickly search for nearest neighbor points in a multidimensional data space. Through the KD tree acceleration algorithm, the corresponding point pairs in two adjacent layers of point cloud data sets can be quickly found, and the optimal registration transformation matrix is ​​calculated based on these point pairs. This matrix includes parameters for geometric transformations such as translation and rotation, and describes the relative transformation relationship between the two sets of point cloud data. Finally, the calculated registration transformation matrix is ​​used to perform registration operations on the point cloud data corresponding to all multi-layer overhead camera data. Through this transformation matrix, all point cloud data are aligned to a unified coordinate system, so that point cloud data from different flight altitudes and perspectives can be accurately matched, eliminating deviations in spatial position. The registered multi-layer point cloud dataset can more accurately reflect the three-dimensional structure of the entire water conservancy survey area, providing accurate data support for subsequent ground point extraction and other analysis and processing.

[0041] Furthermore, when the multi-layer overhead camera data is aligned based on the ICP algorithm, if the multi-layer overhead camera data includes low-layer overhead camera data, middle-layer overhead camera data and high-layer overhead camera data; the priority of the feature vector of the low-layer overhead camera data is that the three gradient components are greater than the normal vector and greater than the curvature, the priority of the feature vector of the middle-layer overhead camera data is that the three gradient components are greater than the curvature and greater than the normal vector, and the priority of the feature vector of the high-layer overhead camera data is that the normal vector is greater than the curvature and greater than the three gradient components.

[0042] When aligning multi-layer overhead camera data based on the ICP algorithm, if the multi-layer overhead camera data includes low-layer, mid-layer, and high-layer overhead camera data, the priority of the feature vector will vary depending on the data characteristics of the different layers. For low-layer overhead camera data, the priority of its feature vector is three gradient components greater than normal vector greater than curvature; for mid-layer overhead camera data, the priority of its feature vector is three gradient components greater than curvature greater than normal vector; for high-layer overhead camera data, the priority of its feature vector is normal vector greater than curvature greater than three gradient components. By setting the feature vector priority in this way, the accuracy and efficiency of the ICP algorithm in the multi-layer data registration process can be improved according to the characteristics of the data at different layers, ensuring the optimal registration effect of each layer of data, thereby obtaining more accurate registration results.

[0043] Furthermore, a multi-layer point cloud data set corresponding to the multi-layer overhead camera data is obtained for point cloud conversion; wherein, if the multi-layer overhead camera data includes low-layer overhead camera data, middle-layer overhead camera data and high-layer overhead camera data, the point cloud density of the low-layer overhead camera data is greater than the point cloud density of the middle-layer overhead camera data, which is greater than the point cloud density of the high-layer overhead camera data.

[0044] In these point cloud datasets, if the multi-layer overhead camera data includes low-layer, mid-layer, and high-layer overhead camera data, the point cloud density of each layer will vary. Specifically, the point cloud density of the low-layer overhead camera data is generally greater than that of the mid-layer overhead camera data, which in turn is greater than that of the high-layer overhead camera data. Low-layer aircraft capture images closer to the ground, and their cameras capture more image detail. Therefore, when converted into point clouds, the point cloud data has a higher density, providing richer ground detail and higher spatial resolution. Mid-layer aircraft, which are farther from the ground, cover a wider area, but due to their relative distance, have a lower point cloud density and resolution. High-layer aircraft fly at a higher altitude, and their cameras capture a larger ground area but with less detail, resulting in a lower point cloud density. Based on the differences in point cloud density between different layers of data, different optimization strategies can be employed during subsequent processing to ensure the registration accuracy of high-density and low-density point cloud data, thereby guaranteeing the overall quality and accuracy of the final point cloud dataset.

[0045] The multi-layer overhead camera data is converted into a point cloud and then input into a ground point extraction model for multi-scale optimization processing to obtain a ground point extraction result.

[0046] After converting the multi-layer overhead camera data into point clouds, multi-layer point cloud datasets are generated. These point cloud datasets include point clouds corresponding to the low-level, mid-level, and high-level overhead camera data. Next, these point cloud data are input into the ground point extraction model for multi-scale optimization processing to obtain the final ground point extraction results.

[0047] Furthermore, the multi-layer overhead camera data is converted into a point cloud and then input into a ground point extraction model for multi-scale optimization processing to obtain a ground point extraction result. The method includes:

[0048] The ground point extraction model includes an encoder, a decoder and a convolution layer; the multi-layer point cloud dataset obtained by converting the multi-layer overhead camera data into a point cloud is input into the ground point extraction model, and the convolution layer is used to perform waveform convolution on each layer of the point cloud dataset to obtain a first convolution dataset; the encoder is used to perform 4-level down-sampling on the first convolution dataset, and the decoder is used to perform 4-level up-sampling on the first convolution dataset to obtain a first sampling feature set; ground points are classified according to the first sampling feature set, and the ground point extraction result is output.

[0049] First, a multi-layer point cloud dataset, generated by converting multiple layers of overhead camera data, is input into the ground point extraction model. The ground point extraction model primarily consists of an encoder, a decoder, and convolutional layers. Each layer of the point cloud dataset is fed into the model separately. The convolutional layers perform waveform convolution on each layer of the point cloud data. This waveform convolution effectively extracts local features from the point cloud data, resulting in a first convolutional dataset containing the feature information of the convolved point cloud. The encoder then performs a four-level downsampling process on the first convolutional dataset to compress detailed information and improve computational efficiency. The decoder then performs a four-level upsampling process on the first convolutional dataset, gradually restoring details in the point cloud data and enhancing its resolution. This process also reintroduces details lost during downsampling, allowing the model to recover finer ground point information from coarse features, resulting in more accurate ground point extraction. After obtaining the first sampled feature set, the model uses this feature set to classify ground points in the point cloud. This classification process distinguishes ground points from non-ground points based on each point's spatial features and contextual information. Ultimately, through this process, high-precision ground point extraction results are output, which provide accurate data support for subsequent 3D modeling, terrain analysis and other survey tasks.

[0050] Furthermore, the ground point extraction model further includes a transposed convolutional layer, where the transposed convolutional layer is the transpose of the convolutional layer, and the method includes:

[0051] After using the encoder to perform 4-level downsampling on the first convolution data set, a downsampled feature set is obtained; waveform convolution is performed on the downsampled feature set according to the transposed convolution layer to obtain a second convolution data set; the decoder performs 4-level upsampling on the second convolution data set to obtain an upsampled feature set, ground point classification is performed according to the upsampled feature set, and the ground point extraction result is output.

[0052] First, the encoder performs a four-level downsampling operation on the input first convolutional dataset to obtain a downsampled feature set. During this stage, the encoder continuously downsamples the input point cloud data, gradually reducing its spatial resolution while extracting key features from the data, particularly the abstract features of ground points. Next, a transposed convolutional layer performs waveform convolution on the downsampled feature set. The transposed convolutional layer, the inverse of the convolutional layer, gradually recovers low-resolution feature information through upsampling, expanding its spatial dimensions and restoring details lost during the downsampling process. In this way, the transposed convolutional layer effectively converts low-resolution feature maps into high-resolution ones, enabling more detailed spatial information to be recovered at each point. After processing by the transposed convolutional layer, a second convolutional dataset is generated, containing richer details and feature information, which facilitates accurate ground point identification. The decoder then performs a four-level upsampling operation on the second convolutional dataset to obtain an upsampled feature set. The upsampling process restores spatial resolution by progressively interpolating and expanding the feature maps, thereby restoring detailed information to a higher level of accuracy. Through this process, the decoder is able to gradually restore the spatial features in the point cloud data, ensuring the accurate extraction of ground points in the final result. Finally, ground point classification is performed based on the feature set obtained by upsampling. The classification operation distinguishes ground points from non-ground points in the point cloud data based on the spatial features and local information of each point. Through this ground point classification step, the model can effectively extract ground points and output the final ground point extraction results. This ground point data can be used for subsequent 3D modeling, terrain analysis, and other tasks, providing accurate ground information for water conservancy surveys and other engineering applications.

[0053] In summary, the embodiments of the present application have at least the following technical effects:

[0054] First, a camera is used to take a side-view camera of the target water conservancy survey area to obtain side-view camera data. Next, the building heights in the water conservancy survey area are marked based on the side-view camera data, and clustered according to the marked height data to determine multiple cluster heights. Then, multiple flight altitude parameters are set based on the multiple cluster heights, and multiple camera paths are planned according to the multiple flight altitude parameters. Based on the multiple camera paths, overhead images of the target water conservancy survey area are collected, and multi-layer overhead camera data is output. Finally, the multi-layer overhead camera data is converted into a point cloud and input into the ground point extraction model for multi-scale optimization processing to obtain the ground point extraction results. This solves the technical problem of insufficient ground point extraction accuracy in water conservancy surveys in complex terrain and landforms in the existing technology, and achieves the technical effect of improving the accuracy of ground point extraction and enhancing adaptability to different terrain and landforms.

[0055] Example 2, based on the same inventive concept as the method for extracting and optimizing ground points for water conservancy survey in the previous embodiment, Figure 2 As shown, the present application provides a system for extracting and optimizing ground points for water conservancy survey, wherein the system includes:

[0056] The first data acquisition module 11 is used to use a camera to take a side-view camera of the target water conservancy survey area to obtain side-view camera data; the clustering module 12 is used to mark the building heights in the water conservancy survey area according to the side-view camera data, cluster the marked height data, and determine multiple cluster heights; the second data acquisition module 13 is used to set multiple flight height parameters based on the multiple cluster heights, plan multiple camera paths according to the multiple flight height parameters, collect overhead images of the target water conservancy survey area based on the multiple camera paths, and output multi-layer overhead camera data; the optimization processing module 14 is used to convert the multi-layer overhead camera data into point clouds and input them into a ground point extraction model for multi-scale optimization processing to obtain ground point extraction results.

[0057] Furthermore, the optimization processing module 14 is used to perform the following method:

[0058] Acquire a multi-layer point cloud dataset corresponding to the multi-layer overhead camera data for point cloud conversion; extract two adjacent layers of point cloud data groups based on the multi-layer point cloud dataset; obtain a feature vector representing point cloud characteristics, the feature vector including three gradient components, curvature and normal vector; perform a KD tree accelerated nearest neighbor search on the two adjacent layers of point cloud data groups based on the feature vector to obtain a corresponding registration transformation matrix between the two adjacent layers of point cloud data groups; register the multi-layer overhead camera data based on the registration transformation matrix to obtain a registered multi-layer point cloud dataset.

[0059] Furthermore, the optimization processing module 14 is used to perform the following method:

[0060] When the multi-layer overhead camera data is aligned based on the ICP algorithm, if the multi-layer overhead camera data includes low-layer overhead camera data, middle-layer overhead camera data and high-layer overhead camera data; the priority of the feature vector of the low-layer overhead camera data is that the three gradient components are greater than the normal vector and greater than the curvature, the priority of the feature vector of the middle-layer overhead camera data is that the three gradient components are greater than the curvature and greater than the normal vector, and the priority of the feature vector of the high-layer overhead camera data is that the normal vector is greater than the curvature and greater than the three gradient components.

[0061] Furthermore, the optimization processing module 14 is used to perform the following method:

[0062] Obtain a multi-layer point cloud dataset corresponding to the multi-layer overhead camera data for point cloud conversion; wherein, if the multi-layer overhead camera data includes low-layer overhead camera data, middle-layer overhead camera data and high-layer overhead camera data, the point cloud density of the low-layer overhead camera data is greater than the point cloud density of the middle-layer overhead camera data, which is greater than the point cloud density of the high-layer overhead camera data.

[0063] Furthermore, the optimization processing module 14 is used to perform the following method:

[0064] The ground point extraction model includes an encoder, a decoder and a convolution layer; the multi-layer point cloud dataset obtained by converting the multi-layer overhead camera data into a point cloud is input into the ground point extraction model, and the convolution layer is used to perform waveform convolution on each layer of the point cloud dataset to obtain a first convolution dataset; the encoder is used to perform 4-level down-sampling on the first convolution dataset, and the decoder is used to perform 4-level up-sampling on the first convolution dataset to obtain a first sampling feature set; ground points are classified according to the first sampling feature set, and the ground point extraction result is output.

[0065] Furthermore, the optimization processing module 14 is used to perform the following method:

[0066] After using the encoder to perform 4-level downsampling on the first convolution data set, a downsampled feature set is obtained; waveform convolution is performed on the downsampled feature set according to the transposed convolution layer to obtain a second convolution data set; the decoder performs 4-level upsampling on the second convolution data set to obtain an upsampled feature set, ground point classification is performed according to the upsampled feature set, and the ground point extraction result is output.

[0067] Furthermore, the first data acquisition module 11 is used to perform the following method:

[0068] The target water conservancy survey area is photographed from the side using a camera device, wherein the camera device includes a first camera element and a second camera element, wherein the first camera element is for horizontal side view photography, and the second camera element is for pitch angle side view photography.

[0069] Furthermore, the clustering module 12 is configured to perform the following method:

[0070] A building height recognition model is trained according to the YOLOv7 architecture; the building top heights of the side-view camera data are annotated according to the building height recognition model to obtain top height annotation results, and the top height annotation results are density clustered to output multiple cluster heights, where each cluster height is the highest value within the cluster group to which it belongs.

[0071] Furthermore, the second data acquisition module 13 is configured to execute the following method:

[0072] Acquire multiple camera devices, where the number of the multiple camera devices corresponds to the multiple camera paths; connect the multiple camera devices to a data center as edge devices, and when the multiple camera devices respectively perform camera shooting according to the multiple camera paths, transmit the acquired real-time camera data to the data center through the edge devices for storage.

[0073] It should be noted that the order in which the embodiments of the present application are presented is for illustrative purposes only and does not necessarily represent the superiority or inferiority of the embodiments. Furthermore, the foregoing descriptions of specific embodiments of this specification are provided. The processes depicted in the accompanying drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0074] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

[0075] This specification and drawings are merely illustrative of the present application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Obviously, those skilled in the art may make various modifications and variations to this application without departing from the scope of this application. Thus, this application is intended to include such modifications and variations as fall within the scope of this application and its equivalents.

Claims

1. A method for optimizing the extraction of ground points for water conservancy survey, characterized in that: The method comprises: Using a camera device to take a side-view video of the target water conservancy survey area to obtain side-view video data; Marking the building heights in the water conservancy survey area according to the side-view camera data, clustering the marked height data, and determining a plurality of cluster heights; Setting a plurality of flight height parameters based on the plurality of cluster heights, planning a plurality of camera paths according to the plurality of flight height parameters, collecting overhead images of the target water conservancy survey area based on the plurality of camera paths, and outputting multi-layer overhead camera data; The multi-layer overhead camera data is converted into a point cloud and then input into a ground point extraction model for multi-scale optimization processing to obtain a ground point extraction result; After the multi-layer overhead camera data is converted into a point cloud and before being input into a ground point extraction model, the multi-layer overhead camera data is registered based on an ICP algorithm, the method comprising: Obtaining a multi-layer point cloud data set corresponding to the point cloud conversion of the multi-layer overhead camera data; Extracting point cloud data groups of two adjacent layers according to the multi-layer point cloud data set; Obtaining a feature vector representing a characteristic of the point cloud, the feature vector including three gradient components, a curvature, and a normal vector; Performing a KD tree accelerated nearest neighbor search on two adjacent layers of point cloud data groups based on the feature vector to obtain a corresponding registration transformation matrix between the two adjacent layers of point cloud data groups, and registering the multi-layer overhead camera data according to the registration transformation matrix to obtain a registered multi-layer point cloud data set; When the multi-layer overhead camera data is registered based on the ICP algorithm, if the multi-layer overhead camera data includes low-layer overhead camera data, middle-layer overhead camera data and high-layer overhead camera data; The priority of the feature vector of the low-level bird's-eye view camera data is that the three gradient components are greater than the normal vector and greater than the curvature; the priority of the feature vector of the middle-level bird's-eye view camera data is that the three gradient components are greater than the curvature and greater than the normal vector; and the priority of the feature vector of the high-level bird's-eye view camera data is that the normal vector is greater than the curvature and greater than the three gradient components.

2. The method for extracting and optimizing ground points for water conservancy survey according to claim 1, characterized in that: Obtaining a multi-layer point cloud data set corresponding to the point cloud conversion of the multi-layer overhead camera data; Among them, if the multi-layer overhead camera data includes low-layer overhead camera data, middle-layer overhead camera data and high-layer overhead camera data, the point cloud density of the low-layer overhead camera data is greater than the point cloud density of the middle-layer overhead camera data, which is greater than the point cloud density of the high-layer overhead camera data.

3. The method for extracting and optimizing ground points for water conservancy survey according to claim 1, wherein: The multi-layer overhead camera data is converted into a point cloud and then input into a ground point extraction model for multi-scale optimization processing to obtain a ground point extraction result. The method includes: The ground point extraction model includes an encoder, a decoder and a convolutional layer; Inputting the multi-layer point cloud dataset obtained by performing point cloud conversion on the multi-layer overhead camera data into the ground point extraction model, and performing waveform convolution on each layer of the point cloud dataset using the convolution layer to obtain a first convolution dataset; Performing 4-level downsampling on the first convolution dataset by the encoder, and performing 4-level upsampling on the first convolution dataset by the decoder to obtain a first sampling feature set; Ground point classification is performed according to the first sampling feature set, and a ground point extraction result is output.

4. The method for extracting and optimizing ground points for water conservancy survey according to claim 3, characterized in that: The ground point extraction model further includes a transposed convolutional layer, where the transposed convolutional layer is the transpose of the convolutional layer. The method includes: After downsampling the first convolutional dataset by 4 levels using the encoder, a downsampled feature set is obtained; Performing waveform convolution on the downsampled feature set according to the transposed convolution layer to obtain a second convolution data set; The decoder performs 4-level upsampling on the second convolution data set to obtain an upsampled feature set, performs ground point classification according to the upsampled feature set, and outputs a ground point extraction result.

5. The method for extracting and optimizing ground points for water conservancy survey according to claim 1, wherein: The target water conservancy survey area is photographed from the side using a camera device, wherein the camera device includes a first camera element and a second camera element, wherein the first camera element is for horizontal side view photography, and the second camera element is for pitch angle side view photography.

6. The method for extracting and optimizing ground points for water conservancy survey according to claim 1, wherein: Marking the building heights in the water conservancy survey area according to the side view camera data, the method comprising: Train a building height recognition model based on the YOLOv7 architecture; The building top heights of the side-view camera data are marked according to the building height recognition model to obtain top height marking results, and the top height marking results are density clustered to output multiple cluster heights, each cluster height being the highest value in the cluster group to which it belongs.

7. The method for extracting and optimizing ground points for water conservancy survey according to claim 1, wherein: The method for collecting overhead images of the target water conservancy survey area based on the multiple camera paths includes: Acquire a plurality of imaging devices, where the number of the plurality of imaging devices corresponds to the plurality of imaging paths; The multiple camera devices are connected to the data center as edge devices. When the multiple camera devices respectively perform camera shooting according to the multiple camera paths, the acquired real-time camera data is transmitted to the data center through the edge devices for storage.

8. A system for extracting and optimizing ground points for water conservancy surveys, characterized in that: A system for implementing the method for extracting and optimizing ground points for water conservancy survey according to any one of claims 1 to 7, comprising: A first data acquisition module is used to use a camera to take a side view camera image of the target water conservancy survey area to obtain side view camera data; A clustering module, configured to mark the building heights in the water conservancy survey area according to the side-view camera data, cluster the marked building height data, and determine a plurality of cluster heights; a second data acquisition module, configured to set a plurality of flight height parameters based on the plurality of cluster heights, plan a plurality of camera paths according to the plurality of flight height parameters, collect overhead images of the target water conservancy survey area based on the plurality of camera paths, and output multi-layer overhead camera data; The optimization processing module is used to convert the multi-layer overhead camera data into point clouds and input them into the ground point extraction model for multi-scale optimization processing to obtain ground point extraction results.

Images