Complex terrain three-dimensional modeling and earthwork volume calculation method based on multi-source fusion point cloud

Abstract

The invention discloses a complex terrain three-dimensional modeling and earthwork volume calculation method based on a multi-source fusion point cloud, belongs to the technical field of surveying and mapping and engineering surveying, and mainly solves the problems of low terrain modeling precision and insufficient earthwork calculation efficiency in a complex scene. The method comprises the following steps: constructing a ground-air integrated multi-source sensing network to synchronously acquire laser point cloud, multi-view images and positioning data; adopting PointNet + +-based initial registration and multi-scale ICP fine registration fusion to generate a unified point cloud; combining semantic segmentation and penetration probability filtering to accurately extract a digital elevation model; constructing a similar triangular prism voxel model by using a constrained Delaunay triangulation network; and finally, the earth volume is rapidly calculated by adopting a GPU parallel voxel cutting and filling algorithm, so that high-precision terrain modeling and rapid engineering quantity calculation are realized, and the method can be efficiently and reliably applied to large-scale projects such as roads and mines.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of surveying and mapping technology. More specifically, the present application relates to a complex terrain three-dimensional modeling and earthwork volume calculation method based on multi-source fused point clouds. BACKGROUND

[0002] With the large-scale promotion of infrastructure construction, mineral resource development and national space planning, high-precision three-dimensional modeling and earthwork volume calculation under complex terrain conditions are in urgent need in engineering practice. Traditional topographic survey methods mainly rely on discrete point collection methods such as total station and GPS-RTK, which not only have low efficiency, but also are difficult to fully reflect the spatial continuity of surface details and complex topography, especially in vegetation-covered areas, steep slopes and areas with frequent human disturbance, the accuracy and reliability are significantly limited.

[0003] In recent years, advanced technologies such as airborne LiDAR and unmanned aerial vehicle multi-view photogrammetry have been gradually introduced in the field of remote sensing and surveying and mapping, providing new data sources for three-dimensional modeling of terrain by obtaining high-resolution point clouds and image data. However, single data source often has inherent limitations: LiDAR data has strong penetration capability and can partially obtain terrain information under vegetation cover, but the point cloud density is unevenly distributed and the cost is high; multi-view images can generate dense point clouds, with lower cost and rich texture information, but are easily affected by light, occlusion and vegetation cover, making it difficult to directly represent the real terrain surface. Therefore, how to effectively fuse multi-source heterogeneous data to form a comprehensive and consistent accuracy surface model has become one of the focuses of current research.

[0004] In the process of multi-source data fusion, there are still several technical problems to be solved. First, there are significant differences in scale, resolution and coordinate system between different sensor data, which need to be unified and registered with high-precision space-time reference. Especially in air-ground collaborative observation, the inconsistency of point cloud density, incident angle and geometric features can easily lead to accumulation of fusion errors, affecting the accuracy of subsequent modeling. Second, point cloud data in complex scenes often contain a large amount of noise and non-ground points (such as vegetation, buildings, moving objects), and traditional filtering algorithms are difficult to effectively remove these interference elements while preserving terrain features, especially in areas with dramatic terrain undulations or complex surface cover, existing methods can easily lead to terrain distortion or loss of details.

[0005] In addition, existing registration methods such as the Iterative Closest Point (ICP) algorithm are sensitive to the initial pose when processing large-scale, multi-source point clouds, and are prone to local optimization, lacking the use of point cloud semantic structure features and local geometric constraints, resulting in insufficient registration accuracy. In terms of earthwork calculation, traditional methods are mostly based on regular grids or simple triangular networks, which are difficult to adapt to complex design surfaces and excavation boundary conditions, have low calculation efficiency and limited accuracy, and especially cannot achieve efficient and accurate excavation and filling volume estimation in large-scale engineering applications.

[0006] To solve the above problems, it is urgent to develop a new method of three-dimensional modeling and earthwork calculation that can integrate multi-source data, adapt to complex terrain conditions, and balance accuracy and efficiency. By constructing a ground-air integrated multi-source perception network, combining semantic segmentation and intelligent filtering technology, the accurate extraction of terrain features is realized. On this basis, the development of high-fidelity three-dimensional modeling and efficient earthwork calculation will provide reliable technical support for engineering design, construction and management. SUMMARY

[0007] An object of the present application is to solve at least the above problems and to provide at least the advantages described later.

[0008] Another object of the present application is to provide a complex terrain three-dimensional modeling and earthwork calculation method based on multi-source fusion point cloud, which significantly improves the accuracy and integrity of complex terrain three-dimensional modeling through ground-air integrated perception and multi-source data fusion, effectively overcomes the limitations of single data source, accurately extracts the real terrain by combining semantic segmentation and penetration probability filtering, and greatly improves the efficiency and accuracy of earthwork calculation by using GPU parallel computing, providing reliable technical support for engineering applications.

[0009] To achieve these objects and other advantages in accordance with the present application, a complex terrain three-dimensional modeling and earthwork calculation method based on multi-source fusion point cloud is provided, comprising: S1, constructing a ground-air integrated multi-source perception network, synchronously collecting laser point cloud data, multi-view image data and high-precision positioning data of the target area, and generating a multi-source heterogeneous data set based on a unified space-time reference protocol; S2, registering and fusing the multi-source heterogeneous data set generated in step S1 to generate a global unified point cloud model; S3, performing semantic segmentation on the unified point cloud model fused in step S2, and combining the semantic segmentation result with the penetration probability filtering algorithm based on the multiple echo characteristics of the LiDAR point element to denoise and classify, and accurately extracting the digital elevation model representing the real terrain; S4, generating a triangular mesh surface model by using a constrained Delaunay triangulation algorithm on the digital elevation model extracted in step S3, and constructing a quasi-triprism element model based on the triangular mesh surface model and the engineering design surface model; S5, calculating the excavation and filling volume of each element by using a GPU parallel accelerated voxel cutting and filling algorithm on the quasi-triprism element model constructed in step S4; In step S2, the multi-source heterogeneous data is first subjected to initial transformation matrix estimation by using a pre-trained PointNet++ grid to extract deep semantic features of the point cloud, and then subjected to precise registration and fusion to generate a global unified point cloud model by using a multi-scale ICP algorithm with a normal vector consistency constraint and dynamically adjusting a weight factor according to the point cloud density and the incident angle.

[0010] Preferably, the ground-air integrated multi-source sensing network in step S1 comprises a ground measurement system arranged on the ground and an air measurement system carried on the unmanned aerial vehicle, wherein the ground measurement system comprises a ground three-dimensional laser scanner and a close-range photogrammetry system, the ground three-dimensional laser scanner obtains high-density three-dimensional point cloud data and reflectivity information of the target area from multiple station angles, and the close-range photogrammetry system collects high-resolution fine texture images of the ground features; the air measurement system comprises a multi-modal acquisition system carried on the unmanned aerial vehicle, the multi-modal acquisition system integrates a multi-lens oblique photography system, an airborne laser radar scanner, and a high-precision GNSS / IMU integrated navigation system, the multi-lens oblique photography system uses multi-view spectral digital images of the target area, the airborne laser radar scanner obtains multi-echo laser point cloud data with intensity information, and the GNSS / IMU integrated navigation system records the high-precision three-dimensional coordinates and attitude angles of the platform at the acquisition time in real time.

[0011] Preferably, the unified space-time reference protocol in step S1 is a space-time unified protocol based on the integration of Beidou third-generation satellite signals and UWB ultra-wideband, and the process of generating a multi-source heterogeneous data set comprises: S11, obtaining reference UTC time and latitude and longitude coordinates through Beidou third-generation satellite signals, and distributing them to each measurement point as a master clock source; S12, covering the target area by a UWB anchor base station network, each measurement device is provided with a UWB tag, and the relative positions between each measurement device and the base station are solved in real time by using a TOA positioning method, and the absolute coordinates provided by the Beidou third-generation satellite are fused to form space positioning data with centimeter-level accuracy; S13, using a high-speed time synchronization protocol to uniformly calibrate the time stamps of the data streams of each measurement device to the master clock source; S14, based on the unified space-time reference, all of the laser point cloud data of different sources, multi-view spectral digital image data, high-precision three-dimensional coordinates and attitude angle data are associated and integrated to generate a multi-source heterogeneous data set unified in the space-time reference.

[0012] Preferably, the process of generating the global unified point cloud model in step S2 comprises: S21, inputting the multi-source heterogeneous point cloud data obtained in step S1 into a pre-trained PointNet++ network, and after the PointNet++ network is pre-trained using a multi-source point cloud training set containing semantic labels of ground, vegetation, building and temporary soil, calculating matching according to the similarity of semantic features between different source point clouds to obtain an initial transformation matrix, and completing coarse registration of the multi-source heterogeneous point cloud data according to the initial transformation matrix; S22, on the basis of coarse registration, performing fine registration by using a multi-scale ICP algorithm, in the data association stage of the ICP algorithm, first introducing a point pair normal vector angle threshold to eliminate error point pairs with inconsistent normal vector directions when searching for the nearest point corresponding relationship; in the iteration process, each point pair is assigned a dynamic weight factor ω to correct the initial transformation matrix to obtain an optimal transformation matrix, and the multi-source heterogeneous point cloud data is further fine registered according to the optimal transformation matrix; S23, fusing all the point cloud data converted to the unified coordinates after fine registration, and performing down-sampling by using a voxelization grid method, eliminating overlapping points generated in the fusion process, and generating a high-precision, non-redundant global unified point cloud model; wherein the dynamic weight factor ω = ω ρi × ω θi ; ω ρi is a weight distribution based on local point cloud density, , ρ max is the density of the densest region in the point cloud, ρ i is the local point cloud density of the i point, σ ρ is a bandwidth parameter of a Gaussian function, σ ρ = 0.3 × ρ max ; ω θi is a weight distribution based on laser incidence angle, , θ i is the laser beam incidence angle, and k is an attenuation adjustment coefficient, which is 2-4.

[0013] Preferably, the specific extraction process of the digital elevation model representing the real terrain in step S3 comprises: S31, input the global unified point cloud model generated in step S2 into a pre-trained deep semantic segmentation neural network, which adopts a PointNet++ architecture and assigns each point cloud a semantic label {l i} containing ground, vegetation, buildings, temporary dumps and other ground objects, i.e.: ; S32, for the point cloud classified as vegetation, based on the multi-echo characteristics of the laser radar points, calculate the penetration probability p i of the point cloud: ; The point cloud classified as ground is set to have a penetration probability of 1; The point cloud classified as buildings, temporary dumps and other ground objects is set to have a penetration probability of 0; S33, set the penetration probability threshold to 0.5-0.8, and determine the point cloud with a penetration probability lower than the penetration probability threshold as a non-ground point cloud and filter it out; S34, perform PMF and CSF filtering algorithms based on slope changes on the point cloud retained after filtering to remove outliers, and then use a moving least squares method to perform smoothing interpolation to fill in missing areas, generating a digital elevation model representing the real terrain; Wherein, {l i} is the output semantic label set, ; {P i} is the input point cloud set; A and B are both laser echo influence coefficients, A is 0.8-1.0, and B is 0.5-2; is the number of laser echoes; S i is a scattering characteristic factor, with a value of 0.7.

[0014] Preferably, the specific process of step S4 to construct a pseudo-triangular prism model includes: S41, input the point set of the digital elevation model extracted in step S3 to generate an initial Delaunay triangular network, embed terrain feature lines as constraint edges into the triangular network to generate a triangular mesh surface model M1 representing the real terrain, and the terrain feature lines include terrain fracture lines, water boundary lines, cliff edges and boundary lines of the calculation area; S42, according to the engineering design drawing or parameters, use a non-uniform rational B-spline surface to perform parametric modeling to generate a smooth and continuous engineering design surface model M2; S43. Perform a two-dimensional Delaunay subdivision on the calculation area where earthwork excavation and filling volumes need to be calculated, generate an irregular triangular mesh, and assign each triangular facet T to the irregular triangular mesh. i Projecting these surfaces onto the triangular mesh surface model M1, representing the real terrain, and the engineering design surface model M2, respectively, creates two surfaces, one above the other. Connecting the corresponding vertices of the two surfaces generates a prism-like element V. i The set of all triangular prism-like elements V i That is, a triangular prism-like meta-model was constructed.

[0015] Preferably, the specific process of calculating the cut and fill volume of each unit in step S5 includes: S51. For each prism-like element V generated in step S4... i Based on the required computational accuracy, the voxel resolution is set, and each prism-like element V is... i The internal space is divided into a uniform voxel grid, and each voxel in the voxel grid is a cubic unit. S52. For each voxel divided in step S51, calculate the elevation values ​​of its center point coordinates relative to the triangular mesh surface model M1 representing the real terrain and the engineering design surface model M2, and calculate the difference Δh between the two elevation values. S53. Determine the cut-and-fill type for each voxel based on the Δh obtained in step S52: If △h>0, then the voxel belongs to the cut area; If △h < 0, then the voxel belongs to the fill region; If △h=0, then the voxel does not need to be excavated or filled. S54, employs a GPU parallel computing architecture to process all triangular prism-like elements V. i The voxel processing tasks are distributed across multiple GPU thread blocks, each thread block being responsible for one or more prism-like V elements. i The voxel calculation is performed in parallel, with steps S52 and S53 executed. S55. The total excavation volume is obtained by summing the volumes of all excavation voxels, and the total fill volume is obtained by summing the volumes of all fill voxels.

[0016] Preferably, the method for setting the voxel resolution in step S51 according to its calculation accuracy requirements includes: For each prism-like element V i Calculate its triangular face T i The equivalent radius R i : ; Determine the prism-like element V i Local voxel resolution r i : ; Among them, Ai Area of the triangular facet Ti m 2 ; R i Equivalent radius of the triangular facet T i m r i Voxel resolution of the pseudo-prism volume V i m C is the design calculation precision, m E is the resolution constant, and is 3-5.

[0017] The present application at least includes the following beneficial effects: Firstly, the present application constructs a ground-air integrated multi-source perception network, comprehensively utilizes multi-source data such as ground laser scanning, close-range photography and unmanned aerial vehicle airborne LiDAR, realizes high-density and multi-angle data acquisition of complex terrain, effectively overcomes the perception limitations of single data source in vegetation covered area and steep slope terrain, significantly improves the integrity and reliability of three-dimensional spatial data, and lays a solid foundation for high-precision modeling; Secondly, the present application adopts a space-time unified protocol based on the fusion of Beidou III and UWB, realizes the space-time reference unification of multi-source heterogeneous data with centimeter-level precision, completely solves the registration error problem caused by the space-time reference difference of different sensors, provides a key guarantee for the high-precision fusion of subsequent multi-source point clouds, and greatly improves the overall data consistency of large-scale terrain measurement; Thirdly, the present application introduces an initial registration based on deep semantic features and a multi-scale ICP fine registration strategy, effectively overcomes the problems of initial pose sensitivity and easy falling into local optimum in large-scale point cloud registration, significantly improves the precision and robustness of multi-source point cloud fusion, and generates a global unified point cloud model with rich details and small error, providing a high-quality data basis for terrain extraction; Fourthly, the present application combines semantic segmentation and LiDAR multiple echo penetration probability filtering, realizes accurate extraction of real terrain under vegetation cover, effectively solves the pain points that traditional filtering algorithms are easy to misremove or retain noise in mixed ground areas, and generates a digital elevation model that can better reflect the real ground relief, providing high-reliability input data for earthwork calculation; Fifthly, the present application adopts a constrained Delaunay triangular network and a pseudo-prism volume model, and performs voxelization excavation and filling analysis based on a GPU parallel computing architecture, breaks through the low efficiency bottleneck of traditional methods in processing complex design surfaces and large-scale calculations, realizes the synchronous improvement of earthwork calculation precision and speed, and provides efficient and reliable technical support for engineering design and construction.

[0018] Other advantages, objects, and features of the application will be apparent to those skilled in the art from the following specification. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 A flowchart of a complex terrain three-dimensional modeling and earthwork calculation method based on multi-source fused point clouds according to one of the technical solutions of the present application is shown in the figure. Figure 2 A specific flowchart of a complex terrain three-dimensional modeling and earthwork calculation method based on multi-source fused point clouds according to another of the technical solutions of the present application is shown in the figure. DETAILED DESCRIPTION

[0020] The present application will be further described in detail below with reference to the accompanying drawings, so that those skilled in the art can implement it according to the description.

[0021] It should be understood that the terms such as "have", "contain" and "include" used herein do not exclude the presence or addition of one or more other elements or combinations thereof.

[0022] As shown in Figure 1 , 2 , the present application provides a complex terrain three-dimensional modeling and earthwork calculation method based on multi-source fused point clouds, comprising: S1, constructing a ground-air integrated multi-source perception network, synchronously collecting laser point cloud data, multi-view image data and high-precision positioning data of a target area, and generating a multi-source heterogeneous data set based on a unified space-time reference protocol; S2, registering and fusing the multi-source heterogeneous data set generated in step S1 to generate a global unified point cloud model; S3, performing semantic segmentation on the unified point cloud model fused in step S2, and combining the semantic segmentation result with a penetration probability filtering algorithm of the LiDAR point element multiple echo characteristics to perform denoising and classification, and accurately extracting a digital elevation model representing the real terrain; S4, according to the digital elevation model extracted in step S3, generating a triangular mesh surface model using a constrained Delaunay triangular mesh division algorithm, and based on the triangular mesh surface model and an engineering design surface model, constructing a pseudo-triangular prism element model; S5, according to the pseudo-triangular prism element model constructed in step S4, using a GPU parallel accelerated voxel cutting and filling algorithm to calculate the excavation and filling amount of each unit; In the step S2, the multi-source heterogeneous data is firstly subjected to initial transformation matrix estimation by pre-trained PointNet++ network to extract deep semantic features of the point cloud, and then subjected to multi-scale ICP algorithm registration and fusion to generate a global unified point cloud model.

[0023] The technical scheme above realizes high-precision acquisition and fusion of complex terrain data by designing the ground-air integrated multi-source perception network. The complex terrain three-dimensional modeling and earthwork calculation method based on multi-source fusion point cloud provided by the application firstly constructs a ground-air integrated multi-source perception network, synchronously acquires laser point cloud data, multi-view image data and high-precision positioning data of a target region, and generates a multi-source heterogeneous data set based on a unified space-time reference protocol, thereby ensuring the consistency of the data in time and space and providing a basis for subsequent processing. Further, the multi-source heterogeneous data set is registered and fused to generate a global unified point cloud model. Specifically, the deep semantic features of the point cloud are firstly extracted by a pre-trained PointNet++ network to estimate the initial transformation matrix, and then a multi-scale ICP algorithm is used for accurate registration. The algorithm introduces a normal vector consistency constraint to eliminate false points, and dynamically adjusts a weight factor according to the point cloud density and the incident angle (wherein the attenuation adjustment coefficient k of the weight factor is 2-4) to realize high-precision fusion. Then, the unified point cloud model after fusion is subjected to semantic segmentation, and combined with the semantic segmentation result and a penetration probability filtering algorithm of the LiDAR point meta multiple echo characteristics to remove noise and classify, so as to accurately extract a digital elevation model representing the real terrain. The penetration probability filtering algorithm uses a penetration probability threshold (the value range is 0.5-0.8) to filter out non-ground point clouds. Then, according to the extracted digital elevation model, a constrained Delaunay triangulation algorithm is used to generate a triangular mesh surface model, and based on the model and an engineering design surface model, a pseudo-triangular prism element model is constructed. Finally, according to the pseudo-triangular prism element model, a GPU parallel accelerated voxel cutting and filling algorithm is used to calculate the excavation and filling amount of each element, and parallel processing is used to improve the calculation efficiency.

[0024] According to the method for complex terrain three-dimensional modeling and earthwork calculation based on multi-source fused point cloud provided by the application, one specific work flow is as follows: first, the laser point cloud data, multi-view image data and high-precision positioning data of a target region are synchronously adopted by a ground-air integrated multi-source perception network, and a multi-source heterogeneous data set is generated based on a unified space-time reference protocol integration to ensure that all data have consistent time and space references. Then, the multi-source heterogeneous data is subjected to initial transformation matrix estimation by a pre-trained PointNet++ network, coarse registration is completed, and then multi-scale ICP algorithm is adopted for fine registration, the algorithm introduces normal vector consistency constraint and a dynamic weight factor (k is 2-4) to optimize registration accuracy, and finally a global unified point cloud model is generated. The fused point cloud model is subjected to semantic segmentation and penetration probability filtering algorithm processing, wherein the penetration probability threshold is set to 0.5-0.8, and is used to filter out non-ground point clouds such as vegetation and buildings, and extract a pure digital elevation model. By using the digital elevation model, a triangular mesh surface model is generated by constraint Delaunay triangulation, and a quasi-triprism element model is constructed in combination with an engineering design surface model. Finally, a voxel cutting and filling algorithm accelerated by GPU parallelism is adopted to divide each quasi-triprism element into a voxel grid, calculate the excavation and filling amount of each voxel, and obtain the total earthwork amount in parallel.

[0025] The technical solution described above fundamentally improves the accuracy and efficiency of complex terrain three-dimensional modeling and earthwork calculation, ensures the synchronicity and consistency of data acquisition through the ground-air integrated multi-source perception network and the unified space-time reference protocol, reduces the error caused by data source heterogeneity, uses PointNet++ and multi-scale ICP algorithm for registration and fusion, combines the normal vector constraint and the dynamic weight factor, improves the accuracy and robustness of point cloud fusion, and adapts to different point cloud densities and incident angles. The combination of semantic segmentation and penetration probability filtering algorithm effectively filters out non-ground point clouds, accurately extracts a digital elevation model, and avoids the deviation of terrain representation in traditional methods. The construction of constraint Delaunay triangulation and quasi-triprism element model provides a more realistic terrain representation and engineering design integration, and the voxel calculation accelerated by GPU parallelism greatly improves the earthwork calculation speed, is suitable for large-scale engineering applications, and realizes efficient and accurate excavation and filling amount evaluation.

[0026] In one of the technical solutions, the ground-air integrated multi-source perception network in step S1 includes a ground measurement system arranged on the ground and an air measurement system carried on the unmanned aerial vehicle, wherein the ground measurement system includes a ground three-dimensional laser scanner and a close-range photogrammetry system, the ground three-dimensional laser scanner obtains high-density three-dimensional point cloud data and reflectivity information of the target area from multiple station angles, and the close-range photogrammetry system collects high-resolution fine texture images of ground features; the air measurement system includes a multi-modal acquisition system carried on the unmanned aerial vehicle, the multi-modal acquisition system integrates a multi-lens oblique photography system, an airborne laser radar scanner, and a high-precision GNSS / IMU combined navigation system, the multi-lens oblique photography system adopts multi-view spectral digital images of the target area, the airborne laser radar scanner obtains multi-echo laser point cloud data with intensity information, and the GNSS / IMU combined navigation system records high-precision three-dimensional coordinates and attitude angles of the platform at the acquisition time in real time.

[0027] The layout of the ground-air integrated multi-source perception network is further optimized in the application, and the preferred ground-air integrated multi-source perception network includes a ground measurement system arranged on the ground and an air measurement system carried on the unmanned aerial vehicle. The ground measurement system includes a ground three-dimensional laser scanner and a close-range photogrammetry system; wherein the ground three-dimensional laser scanner (one preferred type is RIEGL VZ-400i) is arranged around the target area from multiple station angles, and is used to obtain high-density three-dimensional point cloud data and reflectivity information, and the close-range photogrammetry system (which can be selected as a high-precision measurement lens matched with a Canon EOS5DS R) is installed on a tripod, and is used to collect high-resolution fine texture images of ground features. The air measurement system includes a multi-modal acquisition system carried on the unmanned aerial vehicle; the system is integrated on an unmanned aerial vehicle platform (which can be selected as a DJI Matrice300 RTK), and specifically includes a multi-lens oblique photography system (such as a five-lens oblique camera Smart3D), an airborne laser radar scanner (such as a RIEGL VUX-1LR), and a high-precision GNSS / IMU combined navigation system (such as a NovAtel SPAN-IGM-A1). The multi-lens oblique photography system is installed on the unmanned aerial vehicle gimbal, and is used to collect multi-view spectral digital images of the target area; the airborne laser radar scanner and the GNSS / IMU system are tightly fixed and connected, and are installed on the unmanned aerial vehicle body, and are used to obtain multi-echo laser point cloud data with intensity information, and to record high-precision three-dimensional coordinates and attitude angles of the platform at the acquisition time in real time. It should be noted that the device models listed in the application are not mandatory, and other models that can achieve the same function can be used instead.

[0028] The optimized ground-air integrated multi-source perception network layout is extremely beneficial to realize comprehensive and high-precision data acquisition, and provides rich data sources for three-dimensional modeling. Through the combination of ground and air systems, omnidirectional coverage of the terrain is realized: the ground system provides high-density side and detail data, effectively making up for the possible occlusion and bottom point cloud sparseness problems of the air system; and the air system efficiently acquires large-range top information of the ground. The optimized ground-air integrated multi-source perception network significantly improves the integrity and reliability of the data set, and the collaborative work of the multi-modal sensors brings about the complementarity of the data types, the laser radar provides accurate three-dimensional geometric information, the multi-view image provides rich texture and spectral information, and the high-precision positioning and attitude data ensure that all data have a unified spatial reference. The collected multi-source heterogeneous data set lays a solid foundation for subsequent generation of high-precision unified point cloud models and real three-dimensional models, and finally improves the accuracy and reliability of complex terrain earthwork calculation.

[0029] In one of the technical solutions, the unified space-time reference protocol in step S1 is a space-time unified protocol based on the fusion of Beidou third-generation satellite signals and UWB ultra-wideband, and the process of generating a multi-source heterogeneous data set includes: S11, obtaining reference UTC time and latitude and longitude coordinates through Beidou third-generation satellite signals, and distributing them to each measurement point as a master clock source; S12, covering the target area through the UWB anchor base station network, each measurement device is built-in with a UWB tag, and the relative positions between each measurement device and the base station are solved in real time through the TOA positioning method, and are fused with the absolute coordinates provided by the Beidou third-generation satellite to form centimeter-level precision spatial positioning data; S13, using a high-speed time synchronization protocol to uniformly calibrate the time stamps of the data streams of each measurement device to the master clock source; S14, based on the unified space-time reference, integrating all the laser point cloud data, multi-view spectral digital image data, high-precision three-dimensional coordinates and attitude angle data of different sources to generate a multi-source heterogeneous data set unified in space-time reference.

[0030] The application further optimizes the space-time reference protocol, and the space-time protocol based on the Beidou third-generation satellite signal and UWB ultra-wideband fusion is first proposed, and the reference unification of multi-source data is realized by using the space-time protocol based on the Beidou third-generation satellite signal and UWB ultra-wideband fusion. The space-time protocol based on the Beidou third-generation satellite signal and UWB ultra-wideband fusion first uses the Beidou third-generation satellite signal as a master clock source to provide the UTC time and the latitude and longitude coordinates of the reference for the whole measurement system, so that all devices have a unified absolute time reference and a geographical coordinate frame. Then, the UWB anchor base station network is deployed in the target area, and the UWB tag is built-in in each measurement device, and the relative positions between each measurement device and the base station are solved in real time by using the TOA (time of arrival) positioning method; the solved relative position data and the absolute coordinates provided by the Beidou are fused, and finally the space positioning data with centimeter-level precision is formed. In terms of time synchronization, the space-time protocol based on the Beidou third-generation satellite signal and UWB ultra-wideband fusion uses a high-speed time synchronization protocol to unify and calibrate the time stamps of the data streams of each measurement device to the master clock source provided by the Beidou, so as to eliminate the time deviation caused by the different sampling time of different devices. Based on the unified space-time reference, the laser point cloud data, multi-view image data, high-precision three-dimensional coordinates and attitude angle data from different sources such as the ground three-dimensional laser scanner, the close-range photogrammetry system, the airborne laser radar and the oblique photography system are associated and integrated to generate a multi-source heterogeneous data set with completely unified space-time reference.

[0031] According to the space-time protocol based on the Beidou third-generation satellite signal and UWB ultra-wideband fusion provided by the application, the specific working process is as follows: the Beidou third-generation satellite receiver first obtains the reference UTC time and latitude and longitude coordinates, and distributes this information as a master clock source to each measurement device node on the ground and in the air; at the same time, the multiple UWB anchor base stations previously arranged in the target area start to work, the UWB tag built-in in each measurement device and the base station perform distance measurement and communication through the TOA algorithm, and the accurate relative position of the device relative to the base station network is solved in real time. The solved UWB relative position data and the absolute coordinates provided by the Beidou satellite are combined in the fusion algorithm to generate space positioning data with centimeter-level precision. The high-speed time synchronization protocol then runs to align the time stamps in the data streams of all measurement devices (such as laser scanners, cameras, GNSS / IMU) with the master clock source, and realizes microsecond-level time synchronization. Finally, the laser point cloud, multi-view image, platform attitude and other data collected at the unified time and with the unified space coordinates are automatically associated and packaged to generate the final multi-source heterogeneous data set.

[0032] The space-time protocol based on the third-generation Beidou satellite signal and UWB ultra-wideband fusion provided by the application fundamentally guarantees the accuracy and reliability of multi-source data fusion, overcomes the problems of signal shielding, drift or insufficient accuracy of a single navigation system in a complex environment by fusing the advantages of absolute positioning of Beidou and relative positioning of UWB, provides a stable and centimeter-level-precision spatial reference for the entire system, and ensures the consistency of the spatial positions of all collected data. Through the high-speed time synchronization protocol, the timestamps of various heterogeneous sensors are unified to the Beidou master clock, eliminating data time delay and misplacement caused by different internal clocks of devices, and laying a solid foundation for precise synchronization collection and subsequent inter-frame matching of data between ground and air equipment. The strict space-time unification protocol provides crucial data quality guarantee for generating a high-precision global unified point cloud model, and is the core prerequisite for the accuracy of subsequent three-dimensional modeling and earthwork calculation results.

[0033] In one of the technical solutions, the process of step S2 of generating a global unified point cloud model includes: S21, inputting the multi-source heterogeneous point cloud data obtained in step S1 into a pre-trained PointNet++ network, and after the PointNet++ network is pre-trained using a multi-source point cloud training set containing semantic labels of the ground, vegetation, buildings and temporary earth, calculating matching according to the similarity of semantic features between different source point clouds to obtain an initial transformation matrix, and completing coarse registration of the multi-source heterogeneous point cloud data according to the initial transformation matrix; S22, on the basis of coarse registration, performing fine registration by using a multi-scale ICP algorithm, in the data association stage of the ICP algorithm, first introducing a point pair normal vector angle threshold to eliminate incorrect point pairs with inconsistent normal vector directions; in the iteration process, each point pair is assigned a dynamic weight factor ω to correct the initial transformation matrix to obtain an optimal transformation matrix, and the multi-source heterogeneous point cloud data is further fine registered according to the optimal transformation matrix; S23, fusing all point cloud data converted to a unified coordinate after fine registration, and performing down-sampling by using a voxelization grid method, eliminating overlapping points generated in the fusion process, and generating a high-precision, non-redundant global unified point cloud model; wherein the dynamic weight factor ω = ω ρi × ω θi ; ω ρi is a weight distribution based on local point cloud density, , ρ max is the density of the densest region in the point cloud, ρ i is the local point cloud density of the i point, σ ρ is a bandwidth parameter of a Gaussian function, σ ρ = 0.3 × ρ max ; ω θi is the weight based on laser incidence angle, ,θ i is the laser beam incidence angle, and k is the attenuation adjustment coefficient, taking 2-4.

[0034] The application further optimizes the method of generating a global unified point cloud model through a fine three-step process. First, the multi-source heterogeneous point cloud data with unified space-time reference is input into the pre-trained PointNet++ network. The network is pre-trained using a dataset containing semantic labels of ground, vegetation, buildings and temporary earthwork. The initial matching is calculated by comparing the similarity of semantic features between different source point clouds, and the initial transformation matrix is obtained accordingly, completing the coarse registration of multi-source point cloud data and laying a foundation for subsequent processing. On the basis of coarse registration, multi-scale ICP algorithm is used for fine registration. In the data association stage, the algorithm introduces a point pair normal vector angle threshold, effectively eliminating incorrect point pairs with inconsistent normal vectors, and improving the accuracy of matching. In the iterative optimization process, a dynamic weight factor ω is assigned to each point pair to correct the transformation matrix; the weight factor ω is composed of the weight ω ρi based on local point cloud density and the weight ω θi based on laser incidence angle, wherein the attenuation adjustment coefficient k in the incidence angle weight takes a value range of 2-4. By dynamically adjusting the contribution degree of different point pairs, the optimal transformation matrix is finally obtained, realizing high-precision registration of point clouds. Finally, all the fine-registered point cloud data is converted to a unified coordinate system, and the voxelization grid method is used for downsampling to eliminate overlapping points generated in the fusion process, generating a high-precision, non-redundant global unified point cloud model, providing a high-quality data basis for subsequent terrain extraction and modeling.

[0035] According to the above technical scheme, in the coarse registration stage, the pre-trained PointNet++ network receives multi-source point cloud data from the ground and the air. The network uses the semantic features (such as ground, vegetation, buildings, etc.) learned by it to find similar structures in different source point clouds, calculates and outputs an initial transformation matrix, and preliminarily aligns point clouds of different angles and sources to a unified coordinate system, completing coarse registration. Then enter the fine registration stage, and the multi-scale ICP algorithm starts to run. First, when finding the nearest point correspondence, the normal vector angle between the points is calculated and a threshold constraint is applied to filter out incorrect point pairs with large direction differences. Then, a dynamic weight factor ω is calculated for each point pair, which considers the local point cloud density of the point and the incidence angle of the laser beam of the point. The weight is used to iteratively optimize the transformation matrix until convergence, realizing high-precision registration of point clouds at sub-pixel level.

[0036] The above optimization process greatly benefits the high-precision and high-robustness registration and fusion of multi-source point clouds. By adopting coarse registration based on deep semantic features, the accuracy and reliability of initial registration are significantly improved, overcoming the problem that traditional methods based only on geometric features are prone to matching failure in complex scenes, providing an excellent starting point for subsequent fine registration. The introduction of the multi-scale ICP algorithm with normal vector constraint and dynamic weight factor greatly improves the accuracy and robustness of fine registration. The normal vector constraint effectively suppresses false matches, while the dynamic weight factor adaptively balances the contributions of different point cloud density regions and different incident angle point cloud data, reducing the adverse effects of uneven density and incident angle distortion on the registration result, ensuring the high quality of the final registration model. Through voxelization and downsampling for final fusion, while preserving key terrain features, redundant points in the massive point cloud data are effectively eliminated, generating a lightweight, non-redundant global unified point cloud model, significantly improving the efficiency of subsequent data processing and providing a reliable guarantee for accurate extraction of digital elevation models.

[0037] In one of the technical solutions, the specific extraction process of the digital elevation model representing the real terrain includes: S31, input the global unified point cloud model generated in step S2 into a pre-trained deep semantic segmentation neural network, which adopts a PointNet++ architecture and assigns each point cloud a semantic label {l i} containing ground, vegetation, buildings, temporary dumps and other ground objects. ; S32, for the point cloud classified as vegetation, based on the multi-echo characteristics of laser radar points, calculate the penetration probability p i of the point cloud: ; The penetration probability of the point cloud classified as ground is set to 1; The penetration probability of the point cloud classified as buildings, temporary dumps and other ground objects is set to 0; S33, set the penetration probability threshold to 0.5-0.8, and determine the point cloud with a penetration probability lower than the penetration probability threshold as a non-ground point cloud and filter it out; S34, perform PMF and CSF filtering algorithms based on slope changes to remove outliers from the remaining point cloud after filtering, and then use the moving least squares method to smooth interpolation and fill in the missing areas, generating a digital elevation model representing the real terrain; where {l i} is the output semantic label set, ; {P iis the input point cloud set; A and B are laser echo influence coefficients, A is 0.8-1.0, and B is 0.5-2; is the laser echo times; S i is a scattering characteristic factor, and the value is 0.7.

[0038] The application further optimizes the accurate extraction method of the digital elevation model, firstly inputs the global unified point cloud model into the pre-training deep semantic segmentation neural network based on the PointNet++ architecture, the network allocates a semantic label to each point cloud, and the label set includes the ground, vegetation, building, temporary dump and other ground objects, so that the preliminary semantic understanding and classification of the point cloud data are completed. Subsequently, according to the semantic segmentation result, different point clouds of different categories are processed by using a differentiated penetration probability processing strategy. For the point cloud classified as vegetation, based on the multi-echo characteristics of the laser radar point, a penetration probability model containing laser echo influence coefficients (A is 0.8-1.0, B is 0.5-2) and a scattering characteristic factor (S i is 0.7) is used for calculation; for the ground point cloud, the penetration probability is directly set to 1; and for the building, temporary dump and other non-ground point clouds, the penetration probability is set to 0. Then, a penetration probability threshold (the value range is 0.5-0.8) is set, the point cloud with a penetration probability lower than the threshold is determined as a non-ground point and is filtered out. Finally, the remaining point cloud is subjected to PMF and CSF filtering based on the slope change to remove outliers, and the moving least square method is used for smoothing interpolation to repair the missing area, and finally a high-precision digital elevation model is generated.

[0039] According to the technical solution, the pre-trained deep semantic segmentation network receives a global point cloud model, performs inference on each point in the point cloud model, and outputs a semantic label of each point, so that the whole point cloud scene is decomposed into categories such as ground, vegetation, building, temporary dump and other ground objects, and semantic-level understanding of the scene is completed. Then, the penetration probability calculation and filtering are performed. All the points are traversed, and the corresponding penetration probability calculation or assignment rule is selected according to the semantic label of each point: the probability of the vegetation point is calculated according to the echo times (and coefficients A and B (A is 0.8-1.0, and B is 0.5-2)); the probability of the ground point is set to 1; and the probability of the non-ground point is set to 0. After the calculation, the penetration probability of each point is compared with a preset threshold (0.5-0.8, which can be slightly revised according to different field environments), and all points below the threshold are removed to preliminarily separate the ground point cloud. Finally, the DEM is generated through fine processing, and the preliminarily filtered ground point cloud is processed through PMF, CSF and other filtering algorithms based on slope change to further remove residual outlier points. Then, the mobile least square method is used to smooth and interpolate the point cloud missing area caused by filtering or occlusion to fill the holes, and finally a complete, smooth and accurate digital elevation model representing the real terrain is output.

[0040] The accurate extraction method of the optimized digital elevation model is beneficial to the accurate extraction of the real terrain in a complex scene. Through deep semantic segmentation and deep fusion of the physical characteristics (multiple echoes) of LiDAR, intelligent identification and differentiation of ground objects such as vegetation coverage are realized, the accuracy of terrain point and non-terrain point classification is significantly improved, and the limitations of traditional single geometric filtering methods are overcome. The penetration probability processing strategy based on semantic category differentiation is adopted, and the adjustable penetration probability threshold (0.5-0.8) is set, so that the filtering process is more targeted and flexible, and the real terrain points can be effectively retained while the non-ground points are filtered out to the maximum extent. In combination with the advanced slope filtering and mobile least square interpolation algorithm, the finally generated digital elevation model not only has high precision and strong fidelity, but also has good integrity and smoothness, thereby providing an extremely reliable terrain data basis for subsequent triangular net modeling and earthwork calculation, and ensuring the accuracy of the final engineering calculation results.

[0041] In one of the technical solutions, the specific process of step S4 for constructing the pseudo-triangular prism model includes: S41, taking the point set of the digital elevation model extracted in step S3 as input, generating an initial Delaunay triangular net, embedding terrain feature lines as constraint edges into the triangular net, and generating a triangular net grid surface model M1 representing the real terrain, wherein the terrain feature lines include terrain fracture lines, water boundary lines, cliff edges and boundary lines of the calculation area; S42, according to the engineering design drawing or parameter, parameterized modeling is carried out by using non-uniform rational B-spline surface, and a smooth and continuous engineering design surface model M2 is generated; S43, the two-dimensional Delaunay is divided for the calculation area needing to calculate the earthwork volume, and the irregular triangular net is generated, and each triangular facet T i is projected onto the triangular mesh surface model M1 and the engineering design surface model M2 respectively, forming the upper and lower two surfaces, connecting the corresponding vertices of the two surfaces, and generating a pseudo-triangular prism element V i , all the pseudo-triangular prism elements V i are collected, and a pseudo-triangular prism element model is constructed.

[0042] The application further structurally designs the construction method of the pseudo-triangular prism element model, takes the accurately extracted digital elevation model point set as input, generates an initial triangular net by using a constraint Delaunay triangular net division algorithm, and embeds key terrain feature lines as constraint edges into the triangular net, wherein the terrain feature lines include but are not limited to terrain fracture lines, water area boundary lines, scarp edges and calculation area boundary lines, so as to generate a high-fidelity real terrain triangular mesh surface model M1, and ensure the accurate expression of terrain details. Subsequently, according to the engineering design drawing or parameterized design data, parameterized modeling is carried out by using a non-uniform rational B-spline surface (NURBS), and a smooth and continuous engineering design surface model M2 is generated. The engineering design surface model can accurately reflect the geometric shape and requirements of the engineering design, and provide an ideal terrain reference surface for subsequent earthwork volume calculation. Finally, the two-dimensional Delaunay is divided for the area needing to calculate the earthwork volume, and an irregular triangular net is generated. Each triangular facet T i is vertically projected onto the real terrain surface M1 and the engineering design surface M2 respectively, and the corresponding vertices of the two surfaces are connected, and a pseudo-triangular prism element V i is constructed. The collection of all these element bodies constitutes the final pseudo-triangular prism element model, and provides an ideal element body basis for accurately calculating the excavation and filling volume.

[0043] The method for constructing the quasi-triangular prism element model provided by the application constructs a real terrain surface by introducing a terrain characteristic line as a constraint condition, maximally retains the detail features and fracture lines and other key landforms of the terrain, and ensures the accuracy of the real terrain model, thereby providing reliable actual terrain data basis for calculation. The parameterized modeling of the engineering design surface by using the NURBS surface can efficiently and flexibly represent various complex engineering design surfaces, ensures the continuity and smoothness of the design surface model, and meets the actual needs of engineering design. The calculation region is divided and projected and connected into quasi-triangular prism elements, the complex earthwork calculation problem is decomposed into the volume calculation problem of a large number of regular elements, the concept is clear, and subsequent parallel calculation processing is very convenient, thereby laying a solid foundation for efficient and accurate earthwork calculation.

[0044] In one of the technical solutions, the specific process of step S5 for calculating the earthwork volume of each unit includes: S51, for each quasi-triangular prism element V i generated in step S4, the internal space of each quasi-triangular prism element V i is divided into a uniform voxel grid according to the voxel resolution set according to the calculation accuracy requirement, and each voxel in the voxel grid is a cubic element; S52, for each voxel divided in step S51, the coordinates of the center point are calculated, and the elevation values of the triangular mesh surface model M1 representing the real terrain and the engineering design surface model M2 are calculated, and the difference Δh between the two elevation values is calculated; S53, according to the Δh obtained in step S52, the digging type of each voxel is judged: Δh>0, the voxel belongs to the excavation area; Δh<0, the voxel belongs to the filling area; Δh=0, the voxel does not need to be excavated and filled; S54, using the GPU parallel computing architecture, the voxel processing tasks of all quasi-triangular prism elements V i are allocated to multiple GPU thread blocks, each thread block is responsible for the voxel calculation of one or more quasi-triangular prism elements V i , and steps S52 and S53 are executed in parallel; S55, the volumes of all excavation area voxels are accumulated to obtain the total excavation volume, and the volumes of all filling area voxels are accumulated to obtain the total filling volume.

[0045] This invention further designs an efficient calculation process for the excavation and fill volumes of each unit. For each constructed quasi-triangular prism voxel, a voxel resolution is set according to the required calculation accuracy, dividing its internal space into a uniform cubic voxel mesh, providing the basic unit for refined calculations. Subsequently, for each voxel, the elevation values ​​of its center point coordinates on the real terrain surface M1 and the engineering design surface M2 are calculated, and the difference Δh is obtained. The sign of Δh determines whether the voxel belongs to an excavation area, a fill area, or an area requiring no excavation or fill. A GPU parallel computing architecture is used to accelerate the calculation process, distributing the voxel processing tasks of all voxels to multiple GPU thread blocks for parallel execution. Each thread block is responsible for calculating the voxel elevation difference and determining the excavation / fill type for one or more voxels. Finally, after the parallel calculation is completed, the volumes of all excavation area voxels are summed to obtain the total excavation volume, and the volumes of all fill area voxels are summed to obtain the total fill volume, achieving efficient and accurate earthwork volume statistics. By dividing each prism-like voxel element into a finer voxel mesh, the computational unit is refined, enabling a more accurate description of the spatial relationship between complex terrain and design surfaces. This significantly improves the accuracy of earthwork volume calculation, particularly for complex boundaries and abrupt terrain changes. A GPU parallel computing architecture is employed to accelerate massive voxel calculation tasks, distributing the tasks across multiple thread blocks for simultaneous execution. This greatly improves computational efficiency, allowing earthwork volume calculations for large-scale projects to be completed within an acceptable timeframe, meeting the speed requirements of practical engineering applications. Through voxel-level cut-and-fill judgment and volume accumulation, the final cut-and-fill volume results are accurate and reliable, providing precise data support for engineering decisions and construction organization, effectively avoiding resource waste or rework caused by calculation errors.

[0046] In one of the technical solutions, the method for setting the voxel resolution according to its calculation accuracy requirements in step S51 includes: For each prism-like element V i Calculate its triangular face T i The equivalent radius R i : ; Determine the prism-like element V i Local voxel resolution ri: ; Among them, A i Triangular facet Ti area, m 2 ; R i For triangular facet T i The equivalent radius, m; r i The element V is a prism-like shape. i voxel resolution, m; C represents the design calculation accuracy, in meters (m). E is a resolution constant, and is 3-5.

[0047] The application further limits the preferred method of setting the voxel resolution, which adaptively determines the voxel resolution according to the geometric characteristics of each voxel, avoids the problem of waste of computing resources or insufficient accuracy caused by uniform resolution, and makes small voxels use higher resolution to ensure accuracy and large voxels use relatively lower resolution to improve efficiency. The dual-factor decision mechanism of equivalent radius and design accuracy not only considers the geometric characteristics of the voxel itself, but also ensures that the final resolution meets the accuracy requirements of engineering design, so that the calculation result is accurate and reliable. The introduction of the resolution constant (value 3-5) provides a flexible adjustment means for different application scenarios, and users can balance between calculation accuracy and efficiency according to actual needs, enhancing the practicality and adaptability of the method.

[0048] The number of devices and the scale of processing described herein are used to simplify the description of the application. It is obvious to those skilled in the art that the application of the application based on multi-source fusion point cloud complex terrain three-dimensional modeling and earthwork calculation method, modification and change is obvious.

[0049] Although the embodiments of the application have been disclosed as above, it is not limited to the application listed in the specification and the embodiments, and can be fully applied to various fields suitable for the application, and additional modifications can be easily realized by those skilled in the art, therefore the application is not limited to specific details and the figures shown and described herein, without departing from the general concept defined by the claims and the equivalent scope.

Claims

1. A method for complex terrain three-dimensional modeling and earthwork volume calculation based on multi-source fused point clouds, characterized in that, The method comprises the following steps: S1, constructing a ground-air integrated multi-source perception network, synchronously collecting laser point cloud data, multi-view image data and high-precision positioning data of a target area, and generating a multi-source heterogeneous data set based on a unified space-time reference protocol; S2, registering and fusing the multi-source heterogeneous data set generated in step S1 to generate a global unified point cloud model; S3, performing semantic segmentation on the unified point cloud model fused in step S2, and combining the semantic segmentation result with a penetration probability filtering algorithm based on the multi-return characteristics of the LiDAR point element to perform denoising and classification, and accurately extracting a digital elevation model representing the real terrain; S4, according to the digital elevation model extracted in step S3, generating a triangular mesh surface model using a constrained Delaunay triangulation algorithm, and based on the triangular mesh surface model and the engineering design surface model, constructing a pseudo-triangular prism element model; S5, according to the pseudo-triangular prism element model constructed in step S4, using a GPU parallel accelerated voxel cutting and filling algorithm to calculate the excavation and filling amount of each element; In step S2, the multi-source heterogeneous data is first transformed by a pre-trained PointNet++ grid to extract the deep semantic features of the point cloud for initial transformation matrix estimation, and then a multi-scale ICP algorithm is used to accurately register and fuse the global unified point cloud model by introducing a normal vector consistency constraint and dynamically adjusting a weight factor according to the point cloud density and the incident angle.

2. The method for complex terrain 3D modeling and earthwork calculation based on multi-source fused point cloud according to claim 1, characterized in that, In step S1, the ground-air integrated multi-source perception network comprises a ground measurement system arranged on the ground and an air measurement system carried on the unmanned aerial vehicle, wherein the ground measurement system comprises a ground three-dimensional laser scanner and a close-range photogrammetry system, the ground three-dimensional laser scanner obtains high-density three-dimensional point cloud data and reflectivity information of the target area from multiple station angles, and the close-range photogrammetry system collects high-resolution fine texture images of the ground features; the air measurement system comprises a multi-modal acquisition system carried on the unmanned aerial vehicle, the multi-modal acquisition system integrates a multi-lens oblique photography system, an airborne laser radar scanner and a high-precision GNSS / IMU integrated navigation system, the multi-lens oblique photography system uses multi-view spectral digital images of the target area, the airborne laser radar scanner obtains multi-return laser point cloud data with intensity information, and the GNSS / IMU integrated navigation system records the high-precision three-dimensional coordinates and attitude angles of the platform at the time of acquisition in real time.

3. The method for complex terrain 3D modeling and earth volume calculation based on multi-source fused point cloud according to claim 2, characterized in that, In step S1, the unified space-time reference protocol is a space-time unified protocol based on the fusion of Beidou third-generation satellite signals and UWB ultra-wideband, and the process of generating a multi-source heterogeneous data set comprises: S11, obtaining reference UTC time and latitude and longitude coordinates through Beidou third-generation satellite signals, and distributing them to each measurement point as a master clock source; S12, covering the target area with a UWB anchor base station network, each measurement device is provided with a UWB tag, and the relative positions between each measurement device and the base station are solved in real time by a TOA positioning method, and are fused with the absolute coordinates provided by the Beidou third-generation satellite to form space positioning data with centimeter-level accuracy; S13, using a high-speed time synchronization protocol to uniformly calibrate the time stamps of the data streams of each measurement device to the master clock source; S14, based on the unified space-time reference, the laser point cloud data of different sources, multi-view spectral digital image data, high-precision three-dimensional coordinates and attitude angle data are all correlated and integrated to generate a multi-source heterogeneous data set unified in the space-time reference.

4. The method for complex terrain 3D modeling and earth volume calculation based on multi-source fused point cloud according to claim 3, characterized in that, The process of step S2 of generating a global unified point cloud model includes: S21, input the multi-source heterogeneous point cloud data obtained in step S1 into a pre-trained PointNet++ network, the PointNet++ network is pre-trained using a multi-source point cloud training set containing semantic labels of ground, vegetation, building and temporary soil, and then the similarity calculation and matching between different source point clouds are performed according to the similarity of the semantic features of the different source point clouds to obtain an initial transformation matrix, and the coarse registration of the multi-source heterogeneous point cloud data is completed according to the initial transformation matrix; S22, on the basis of coarse registration, fine registration is performed by using a multi-scale ICP algorithm, in the data correlation stage of the ICP algorithm, a point pair normal vector angle threshold is first introduced when searching for the nearest point corresponding relationship to eliminate error point pairs with inconsistent normal vector directions; in the iteration process, a dynamic weight factor ω is allocated to each point pair to correct the initial transformation matrix to obtain an optimal transformation matrix, and the multi-source heterogeneous point cloud data is further fine registered according to the optimal transformation matrix; S23, fuse all the point cloud data converted to a unified coordinate after fine registration, and perform down-sampling by using a voxelization grid method, eliminate the overlapping points generated in the fusion process, and generate a high-precision, non-redundant global unified point cloud model; wherein the dynamic weight factor ω = ω ρi x ω θi ; ω ρi For the weight assignment based on local point cloud density, exp , p max is the density of the densest region in the point cloud, p i is the local point cloud density of the i-th point, s ρ is the bandwidth parameter of the Gaussian function, s ρ = 0.3 x p max ; ω θi For the weight distribution based on the laser incidence angle, ,θ i For the laser beam incidence angle, k is the attenuation adjustment coefficient, taking 2~4.

5. The method for complex terrain 3D modeling and earth volume calculation based on multi-source fused point cloud according to claim 4, characterized in that, The specific extraction process of step S3 of representing a digital elevation model of a real terrain includes: S31, input the global unified point cloud model generated in step S2 into a pre-trained deep semantic segmentation neural network, wherein the deep semantic segmentation neural network adopts a PointNet++ architecture, and assigns a semantic label {l i} containing ground, vegetation, buildings, temporary earthwork and other ground objects to each point cloud, i.e.: ； S32、for the point cloud classified as vegetation, based on the multi-echo characteristics of the lidar points, calculate the penetration probability p of the point cloud i : ； The point cloud classified as ground is set to have a penetration probability of 1; The point cloud classified as building, temporary soil and other ground objects is set to have a penetration probability of 0; S33, set the penetration probability threshold to 0.5-0.8, and determine the point cloud with a penetration probability lower than the penetration probability threshold as non-ground point cloud and filter it out; S34, perform PMF and CSF filtering algorithms based on slope change on the remaining point cloud after filtering to eliminate outliers, and then perform smoothing interpolation to fill in the missing area by using a moving least square method to generate a digital elevation model representing a real terrain; wherein, {l i} is the set of output semantic labels, {l i} {P i} is a set of input point clouds; A and B are both laser echo influence coefficients, A is 0.8-1.0, and B is 0.5-2; is the number of laser returns; S i is the scattering characteristic factor, which takes the value 0.

7.

6. The method for complex terrain 3D modeling and earth volume calculation based on multi-source fused point cloud according to claim 5, characterized in that, The specific process of step S4 of constructing a pseudo-triangular prism element model includes: S41, input the point set of the digital elevation model extracted in step S3 to generate an initial Delaunay triangular network, embed the terrain feature lines including terrain fracture lines, water boundary lines, cliff edges and the boundary lines of the calculation area into the triangular network to generate a triangular grid surface model M1 representing a real terrain; S42, according to the engineering design drawing or parameter, perform parametric modeling by using a non-uniform rational B-spline surface to generate a smooth and continuous engineering design surface model M2; S43, the calculation area needing earthwork calculation is two-dimensionally Delaunay divided to generate irregular triangle net, and each triangle T on the irregular triangle net is recorded i is projected onto the triangular mesh surface model M1 representing the real terrain and the engineering design surface model M2 respectively to form upper and lower surfaces, and the corresponding vertices of the two surfaces are connected to generate pseudo-prism elements V i All pseudo-prism elements V are collected i That is, the pseudo-prism element model is constructed.

7. The method for complex terrain 3D modeling and earth volume calculation based on multi-source fused point cloud according to claim 6, characterized in that, The specific process of step S5 of calculating the excavation and filling amount of each unit includes: S51. For each prism-like element V generated in step S4... i Based on the required computational accuracy, the voxel resolution is set, and each prism-like element V is... i The internal space is divided into a uniform voxel grid, and each voxel in the voxel grid is a cubic unit. S52, for each voxel divided in step S51, calculate the elevation values of the center point coordinates of the voxel with respect to the triangular grid surface model M1 representing a real terrain and the engineering design surface model M2, and calculate the difference Δh between the two elevation values. S53, judging the type of each voxel according to the △h obtained in step S52: △h>0, the voxel belongs to the excavation area; △h<0, the voxel belongs to the filling area; △h=0, the voxel does not need to be excavated or filled; S54, employs a GPU parallel computing architecture to process all triangular prism-like elements V. i The voxel processing tasks are distributed across multiple GPU thread blocks, each thread block being responsible for one or more prism-like V elements. i The voxel calculation is performed in parallel, with steps S52 and S53 executed. S55, accumulating the volume of all the voxels in the excavation area to obtain the total excavation amount, and accumulating the volume of all the voxels in the filling area to obtain the total filling amount.

8. The method for complex terrain 3D modeling and earth volume calculation based on multi-source fused point cloud according to claim 7, characterized in that, The method for setting the voxel resolution according to the calculation accuracy requirement in step S51 includes: For each pseudo-triangular prism volume V i , the equivalent radius R i of its triangular facets T i is calculated: ; Determine the prism-like element V i Local voxel resolution r i : ; where A i is the area of the triangular facet, m Ti 2 ;​ R i is the equivalent radius of the triangular facet T i , m; r i For a hexagonal prism-like volume element V i the voxel resolution, m; C is the design calculation accuracy, m; E is the resolution constant, and the value is 3-5.

Images