Multi-source Fusion Point Cloud Data-driven 3D Street View Generation and Optimization Method

Abstract

The present invention discloses a multi-source fusion point cloud data-driven three-dimensional street view generation and optimization method, which relates to the technical field of three-dimensional street view generation, and includes the following steps: According to the spatial range of the input large-scale point cloud data, combined with the distribution characteristics of its three-dimensional coordinate system, a structured space division algorithm is used to divide the overall registration area into multiple sub-unit areas with equal area and uniform distribution; for each sub-unit area, the geometric structure features of the internal point cloud are obtained, and a geometric feature descriptor including multiple dimensions is constructed. The present invention realizes the recognition of the sub-unit structure complexity through a heterogeneity evaluation mechanism based on geometric features and a support vector machine model, dynamically adjusts the upper limit of the number of matching points, improves the registration accuracy of high-heterogeneity areas, reduces the risk of misregistration in low-heterogeneity areas, thereby optimizing the structure restoration and continuity of the street view model, and enhancing the stability and intelligence of the three-dimensional modeling system.

Description

Technical Field

[0001] The present invention relates to the technical field of 3D street view generation, and particularly to a method for generating and optimizing 3D street views driven by multi-source fusion point cloud data. Background Art

[0002] The multi-source fusion point cloud data-driven 3D street view refers to generating point cloud information by integrating data sources from multiple sensors (such as lidar, stereo cameras, UAV aerial survey systems, mobile measurement platforms, etc.), and using data fusion technology to perform unified registration, accuracy improvement, and structure completion processing on spatial information from different sources, with different accuracies, and different perspectives, so as to construct a high-precision and highly restored 3D street view model. This method not only improves the integrity and detail expressiveness of 3D modeling, but also enables more realistic and data-driven street view simulation and scene reconstruction in fields such as smart cities, autonomous driving, and security monitoring, effectively supporting the in-depth application of urban-level spatial information.

[0003] The existing technologies have the following deficiencies:

[0004] During the unified registration process of large-scale point cloud data, due to significant differences in the spatial expressions of regional data from different sources, different accuracies, and different perspectives, the local geometric heterogeneity levels of each region vary greatly. Existing registration technologies usually adopt a unified parameter strategy and cannot intelligently adjust the selection strategy of matching points according to the point cloud structure complexity of different regions. In this context, if the number of selected matching points for high-heterogeneity regions with rich geometric structures is insufficient, key feature information cannot be effectively extracted and utilized, resulting in a lack of fine structure constraints in the registration process, prone to local rigid transformation errors, and further spreading to other regions through the global optimization process, ultimately leading to problems such as misaligned building outlines, distorted structural boundaries, and distorted street view models, seriously affecting the accuracy of navigation path reasoning, spatial recognition, and engineering analysis; conversely, if the number of selected matching points for low-heterogeneity regions with simple geometric structures is excessive, redundant matching pairs or mismatching relationships are likely to be introduced, forming interfering registration-driven solutions, thereby causing non-structural geometric offsets and superimposing errors in multi-region fusion, resulting in a decrease in the spatial consistency of the overall point cloud model and topological structure breaks, seriously restricting the stability and reliability of subsequent high-precision surveying and mapping modeling and intelligent recognition tasks.

[0005] The above information disclosed in the background art section is only used to enhance the understanding of the background of the present disclosure, and therefore it may include information that does not constitute the prior art known to those of ordinary skill in the art. Summary of the Invention

[0006] The objective of the present invention is to provide a method for generating and optimizing 3D street scenes driven by multi-source fused point cloud data. By introducing a heterogeneity evaluation mechanism constructed based on the geometric features of sub-unit regions and combining with a trained support vector machine model, it realizes the quantitative recognition and classification judgment of the local structure complexity, and dynamically adjusts the upper limit of the number of matching points accordingly, so that high-heterogeneity regions can obtain more geometric constraints, and low-heterogeneity regions can avoid redundant matching interference, thereby effectively avoiding the registration error diffusion and structural misalignment problems caused by the unified parameter strategy, improving the overall quality and stability of the street scene model in terms of morphological restoration, boundary alignment, and topological continuity, and providing a more robust, efficient, and intelligent technical support for application scenarios such as 3D urban reconstruction, autonomous driving map construction, and precision mapping, so as to solve the problems in the above-mentioned background technology.

[0007] To achieve the above objective, the present invention provides the following technical solutions: A method for generating and optimizing 3D street scenes driven by multi-source fused point cloud data, including the following steps:

[0008] According to the spatial range of the input large-scale point cloud data and combining its three-dimensional coordinate system distribution characteristics, a structured spatial division algorithm is used to divide the overall registration area into multiple sub-unit regions with equal area and uniform distribution, and the wide-area point cloud with non-uniform geometric features is disassembled into cells that can be locally modeled and analyzed independently;

[0009] For each sub-unit region, obtain the geometric structure features of the internal point cloud, construct a geometric feature descriptor containing multiple dimensions, and provide multi-source input variables for subsequent heterogeneity analysis;

[0010] Through feature engineering techniques, extract the key indicators reflecting the high heterogeneity of the sub-unit region from the obtained geometric structure features. After in-depth analysis of the extracted key indicators, quantify the geometric heterogeneity level of each sub-unit region based on the analysis results;

[0011] Input the analyzed indicators as feature vectors into a support vector machine model that has been trained in advance with a large number of labeled samples to intelligently predict the geometric heterogeneity degree corresponding to the sub-unit region;

[0012] Based on the heterogeneity degree of the sub-unit region predicted by the support vector machine model, construct an adaptive mapping relationship between the heterogeneity degree and the registration parameters, and dynamically allocate the corresponding upper limit of the number of matching points for sub-unit regions with different heterogeneity levels to achieve area-specific matching point numbers;

[0013] After setting the range of the number of matching points based on the regulation strategy, input the matching point control parameter and other registration parameters (such as matching radius, tolerance threshold) into the subsequent point cloud registration module to achieve the adaptive registration of the sub-unit region.

[0014] Preferably, the overall registration area is divided into multiple sub-unit areas with equal areas and uniform distribution. The specific division strategy adopted is the recursive space segmentation method based on octree, fixed resolution grid division (VoxelGrid), or region-growing spatial adaptive segmentation algorithm.

[0015] Preferably, the overall registration area is divided into multiple sub-unit areas with equal areas and uniform distribution by the recursive space segmentation method based on octree. The specific steps are as follows:

[0016] First, an overall bounding cube is established for the input large-range point cloud data, which is used as the root node to represent the entire spatial area;

[0017] Then, the cube is recursively divided into eight equal parts according to the three-dimensional coordinate range (X, Y, Z) of the point cloud. Each division evenly divides the parent node space into eight sub-cube areas, which are respectively added to the octree structure as child nodes;

[0018] In each division stage, it is judged whether the number of point clouds in each sub-cube meets the minimum segmentation threshold. If it exceeds the threshold, continue to recursively subdivide until the preset minimum voxel size or the lower limit condition of the number of points is met;

[0019] The finally generated leaf nodes are multiple sub-unit areas with equal areas (volumes) and uniform spatial distribution. These areas not only retain the spatial continuity of the original point cloud but also provide a structured spatial basis for subsequent local geometric feature extraction and adaptive registration strategies.

[0020] Preferably, key indicators reflecting the high heterogeneity of the sub-unit area are extracted from the obtained geometric structure features through feature engineering techniques. Among them, the extracted key indicators include the degree of dispersion of the principal curvature value distribution of all points in the sub-unit area and the degree of spatial gradient change of the point density within the sub-unit area. After in-depth analysis of the extracted key indicators, the curvature dispersion reference value and the density gradient reference value are respectively generated, and the geometric heterogeneity level of each sub-unit area is quantified based on the curvature dispersion reference value and the density gradient reference value.

[0021] Preferably, the analyzed curvature dispersion reference value and density gradient reference value are used as feature vectors and input into a support vector machine model that has been trained in advance with a large number of labeled samples. The support vector machine model generates a heterogeneity level coefficient, and the geometric heterogeneity degree corresponding to the sub-unit area is intelligently predicted based on the heterogeneity level coefficient.

[0022] Preferably, the corresponding matching point number range is dynamically allocated for sub-unit areas with different heterogeneity levels. The specific steps are as follows:

[0023] The heterogeneity level coefficient of each sub-unit area predicted by the support vector machine model is obtained, and a matching point adjustment factor for controlling the intensity of matching point quantity distribution is constructed based on the heterogeneity level coefficient. The matching point adjustment factor is calculated in an exponential mapping manner, and the calculation expression is:

[0024]

[0025] ,in: Represents the heterogeneity-driven matching point adjustment factor, and its value range is ; is the maximum amplification coefficient, which controls the upper limit of the matching point adjustment factor and is used to adjust the resource tilt degree of highly heterogeneous areas. ; It is a nonlinear response adjustment factor, which is used to control the response speed of the matching point adjustment factor curve to the heterogeneous level coefficient; is the coefficient of heterogeneity level in the subunit area;

[0026] Match point adjustment factor based on construction

[0027] , the upper limit of the final number of matching points in the sub-unit area is dynamically allocated. The expression of dynamic allocation is as follows:

[0028]

[0029] ,in: Indicates the upper limit of the number of matching points allocated to the current sub-unit area; Indicates the minimum number of matching points set by the system, which is used in low heterogeneity areas (such as flat ground); Indicates the upper limit of the maximum number of matching points, which is applicable to highly heterogeneous areas (such as complex buildings or mixed edge areas).

[0030] Preferably, the specific steps of generating the curvature discrete reference value after in-depth analysis of the discrete degree of the distribution of the principal curvature values of all points in the subunit area are as follows:

[0031] First, for all points in the subunit area, the k nearest neighbor method (k is generally 10~30) is used to calculate the principal curvature value of each point , and sort the principal curvatures of all points in the subunit region according to spatial order or density gradient to construct a curvature sequence Then, the relative jump ratio factor is introduced, which is defined as the logarithmic ratio of the change between adjacent curvature points to highlight the nonlinear mutation characteristics between the sharp jump points. The calculation expression of the relative jump ratio factor is:

[0032]

[0033] ,in: is the principal curvature value of the i th point; is a tiny constant used to avoid a zero denominator (e.g., ); is the relative jump ratio factor, representing the logarithmic jump amplitude between the i th curvature point and the th curvature point; N is the total number of points in the sub-unit area;

[0034] Construct a jump ratio sequence based on the relative jump ratio factor . After obtaining the jump ratio sequence, strengthen the weight of the locally discrete prominent area by constructing a rate of change enhancement function, and finally generate a curvature discrete reference value. The generation expression of the curvature discrete reference value is:

[0035]

[0036] , where: is the curvature discrete reference value; represents the local rate of change of the curvature jump ratio sequence in the neighborhood of the i th curvature point, and is used to capture the trend fluctuation information of the degree of curvature change in the sub-unit area; is the hyperbolic tangent function, used to compress excessive jumps and strengthen the response in the medium and high curvature change intervals; is the discrete amplification factor (e.g., ), used for non-linear enhancement of the overall heterogeneity trend.

[0037] Preferably, the specific steps for generating a density gradient reference value after in-depth analysis of the degree of spatial gradient change of the point density in the sub-unit area are as follows:

[0038] Divide the sub-unit area into equally spaced voxel grids. Let the side length of each voxel be l . For each voxel, calculate the sum of the density gradient transition intensities with its 6 face-adjacent voxels. The calculation expression is:

[0039]

[0040] , where: is the local density transition intensity between the voxel and its 6 face-adjacent voxels; is the index set of the 6 face-adjacent voxels adjacent to the voxel ; d represents the index of a face-adjacent voxel adjacent to the current voxel , used to enumerate all face-adjacent voxels adjacent to the voxel Adjacent voxels in 6 directions of contact (up and down, front and back, left and right); , respectively represent the number of points in the current voxel and adjacent voxels; is a non-linear enhancement index used to enhance the response to high-density jumps (e.g., taking values from 1.5 to 2.5);

[0041] After obtaining the local density transition intensity of all voxels, a structure-sensitive non-linear weighted aggregation method is used to generate the density gradient reference value of the entire subunit region, and the generated expression is as follows:

[0042]

[0043] , where: is the density gradient reference value, used to characterize the heterogeneity level of the subunit region; w represents the voxel index within the subunit region; M is the total number of voxels within the subunit region; is the non-linear aggregation weight factor, used to further suppress weak gradients and highlight significant gradients, ; is used for compressing large values and enhancing small values, strengthening the discrimination ability of the density gradient reference value for high-variation regions.

[0044] In the above technical solution, the technical effects and advantages provided by the present invention are as follows:

[0045] By introducing a heterogeneity evaluation mechanism constructed based on the geometric features of the subunit region and combining with the trained support vector machine model, the present invention realizes the quantitative recognition and classification judgment of the local structure complexity, and dynamically adjusts the upper limit of the number of matching points accordingly, so that high-heterogeneity regions obtain more geometric constraints, and low-heterogeneity regions avoid redundant matching interference, thereby effectively avoiding the registration error diffusion and structure misalignment problems caused by the unified parameter strategy, improving the overall quality and stability of the street view model in terms of shape restoration, boundary alignment and topological continuity, and providing a more robust, efficient and intelligent technical support for application scenarios such as 3D city reconstruction, autonomous driving map construction and precision mapping. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required in the embodiments. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings.

[0047] Figure 1 is the method flow chart of the method for generating and optimizing 3D street view driven by multi-source fusion point cloud data of the present invention. Detailed Implementation Modes

[0048] Example implementation modes will now be described more fully with reference to the accompanying drawings. However, the example implementation modes can be implemented in various forms and should not be construed as limited to the examples set forth herein; rather, these example implementation modes are provided so that this disclosure will be more thorough and complete, and will fully convey the concept of the example implementation modes to those skilled in the art.

[0049] The present invention provides a multi-source fusion point cloud data-driven three-dimensional street view generation and optimization method as Figure 1 shown, including the following steps:

[0050] According to the spatial range of the input large-scale point cloud data and in combination with its three-dimensional coordinate system distribution characteristics, a structured spatial partitioning algorithm is used to divide the overall registration area into multiple sub-unit areas with equal areas and uniform distributions, and the wide-area point cloud with non-uniform geometric characteristics is disassembled into cells that can be locally modeled and independently analyzed;

[0051] Common partitioning strategies include the recursive spatial segmentation method based on octree, fixed-resolution grid partitioning (voxel grid), or region-growing type spatial adaptive segmentation algorithm. The core objective of this process is to disassemble the wide-area point cloud with non-uniform geometric characteristics into cells that can be locally modeled and independently analyzed, so as to provide an operational basis and spatial boundary conditions for subsequent determination of the heterogeneity level of each sub-unit and allocation of registration parameters.

[0052] The recursive spatial segmentation method based on octree divides the overall registration area into multiple sub-unit areas with equal areas and uniform distributions. The specific steps are as follows:

[0053] First, an overall bounding cube is established for the input large-scale point cloud data to represent the entire spatial area as the root node;

[0054] Then, the cube is recursively divided into eight equal parts according to the three-dimensional coordinate range (X, Y, Z) of the point cloud. Each division evenly divides the parent node space into eight sub-cube areas, which are respectively added to the octree structure as child nodes;

[0055] At each division stage, it is judged whether the number of point clouds in each sub-cube meets the minimum segmentation threshold. If it exceeds the threshold, continue recursive subdivision until the preset minimum voxel size or point number lower limit condition is met;

[0056] The finally generated leaf nodes are multiple sub-unit areas with equal areas (volumes) and uniform spatial distributions. These areas not only retain the spatial continuity of the original point cloud, but also provide a structured spatial basis for subsequent local geometric feature extraction and adaptive registration strategies.

[0057] For each sub - unit region, obtain the geometric structure features of the internal point cloud, construct a geometric feature descriptor containing multiple dimensions, and provide multi - source input variables for subsequent heterogeneity analysis;

[0058] To obtain the geometric structure features of the internal point cloud of each sub - unit region, local neighborhood analysis needs to be performed on all the point clouds within that region. The specific method is as follows: for each point within the sub - unit, use a fixed - radius or k - nearest - neighbor search method to construct its local point set, and calculate its spatial distribution characteristics and geometric properties, such as the principal curvature direction, normal vector, point density, etc.; then statistically summarize the local features of all points, such as calculating the mean, standard deviation, or histogram distribution, and finally form a global descriptor reflecting the structural features of the entire sub - unit. This process can be achieved by algorithms such as PCA (Principal Component Analysis), normal statistics, entropy measurement, or in - voxel structure distribution, ensuring that the geometric characteristics of each sub - unit region are fully expressed.

[0059] The geometric structure features of the sub - unit region usually include the following multi - dimensional indicators: point density, that is, the number of points per unit volume, reflecting the density of spatial distribution; principal curvature values and their statistics, such as mean, extreme values, variance, used to measure the degree of surface undulation; variance of normal vector directions, used to characterize the amplitude of surface normal changes, reflecting the structural consistency or complexity; point cloud information entropy, quantifying the order or randomness of the spatial arrangement of the point cloud; proportion of edge points, which can be extracted by detecting normal mutations or gradient mutations, used to identify structural contour features; voxel uniformity index, measuring the spatial uniformity or aggregation of the point cloud within the sub - unit.

[0060] The above - mentioned indicators are combined to form a multi - dimensional feature vector for further heterogeneity level assessment and matching point parameter regulation.

[0061] Extract key indicators reflecting high heterogeneity in the sub - unit region from the obtained geometric structure features through feature engineering techniques. After in - depth analysis of the extracted key indicators, quantify the geometric heterogeneity level of each sub - unit region based on the analysis results;

[0062] Extract key indicators reflecting high heterogeneity in the sub - unit region from the obtained geometric structure features through feature engineering techniques. Among them, the extracted key indicators include the degree of dispersion of the distribution of principal curvature values of all points in the sub - unit region and the degree of spatial gradient change of point density within the sub - unit region. After in - depth analysis of the extracted key indicators, generate a curvature dispersion reference value and a density gradient reference value respectively, and quantify the geometric heterogeneity level of each sub - unit region based on the curvature dispersion reference value and the density gradient reference value.

[0063] For each sub-unit region, the higher the degree of dispersion of the distribution of the principal curvature values of all points in the sub-unit region, the higher the level of geometric heterogeneity of the region is usually indicated. This is because the principal curvature reflects the degree of curvature of the local shape of the point cloud surface, and a high degree of dispersion indicates that there are various different geometric feature forms in the region, such as the transition from flat to strongly convex, from concave to abrupt edge changes; such drastic curvature changes usually mean complex structural composition, large detail differences, or the coexistence of multiple configurations, which are common in building facades, carved surfaces, natural vegetation, damaged components, or structural junction areas, etc. Therefore, the dispersion of the principal curvature not only quantifies the severity of the geometric undulation, but is also an important sensitive indicator for characterizing the internal heterogeneity of the local sub-unit region.

[0064] The specific steps for generating the curvature dispersion reference value after in-depth analysis of the degree of dispersion of the distribution of the principal curvature values of all points in the sub-unit region are as follows:

[0065] First, for all points in the sub-unit region, the k-nearest neighbor method (k is generally 10 - 30) is used to calculate the principal curvature value of each point , and the principal curvatures of all points in the sub-unit region are sorted in spatial order or density gradient to construct a curvature sequence . Then, a relative jump ratio factor is introduced, which is defined as the logarithmic ratio form of the change between adjacent curvature points, used to highlight the non-linear mutation characteristics between points with drastic jumps. The calculation expression of the relative jump ratio factor is:

[0066]

[0067] , where: is the principal curvature value of the i th point; is a small constant used to avoid the denominator being 0 (for example ); is the relative jump ratio factor, indicating the logarithmic jump amplitude between the i th curvature point and the th curvature point; N is the total number of points in the sub-unit region;

[0068] This ratio sequence is used to capture the density and intensity of curvature mutation events in the sub-unit region, not only considering the local deformation amplitude, but also effectively avoiding the interference of extreme values, and is the basic feature flow for constructing the heterogeneity characterization.

[0069] Based on the relative jump ratio factor, a jump ratio sequence is constructed . After obtaining the jump ratio sequence, the weight of the region with enhanced local discreteness is strengthened by constructing a rate of change enhancement function, and finally the curvature dispersion reference value is generated. The generation expression of the curvature dispersion reference value is:

[0070]

[0071] , where: is the discrete reference value of curvature; represents the local change rate of the curvature jump ratio sequence within the neighborhood of the i -th curvature point, and is used to capture the trend fluctuation information of the degree of drastic curvature change in the sub-unit area; is the hyperbolic tangent function, which is used to compress excessive jumps and strengthen the response in the medium and high curvature change intervals; is the discrete amplification factor (such as ), which is used to non-linearly enhance the overall heterogeneity trend;

[0072] As can be seen from the discrete reference value of curvature, the larger the performance value of the discrete reference value of curvature generated after in-depth analysis of the discrete degree of the principal curvature value distribution of all points in the sub-unit area, the higher the geometric heterogeneity level of the sub-unit area, and vice versa. The reason is that the discrete reference value of curvature essentially measures the fluctuation intensity of the curvature jump ratio in the spatial sequence within the region, which is manifested as the density of drastic morphological changes of geometric surfaces at different positions within a local range. When the discrete reference value of curvature is large, it indicates that there are a large number of local geometric units with structural mutations, normal discontinuities, complex morphologies or mixed multi-configurations in the region, which is a typical region with high expression of heterogeneity; while if the discrete reference value of curvature is small, it means that the curvature change in the region is gentle, the structure is unified, the surface continuity is high, the overall configuration is regular, and the geometric heterogeneity level is naturally low.

[0073] For each sub-unit area, the higher the degree of change in the spatial gradient of the point density within the sub-unit area, usually the higher the geometric heterogeneity level of the sub-unit area. This is because the spatial gradient of the point density essentially reflects the degree of non-uniformity of the point cloud distribution in space, that is, the rate of change of the number of points from one local position to another local position. In a region with uniform or regular structure, such as a flat ground, a wall surface or an open space, the point cloud density is usually relatively consistent and the gradient is close to zero; while in a region with complex geometric structure or boundary mutation, such as a building corner, a structure junction, an occlusion edge or a stacked area of anisotropic components, the point density will change sharply with the shape undulation, occlusion projection or scanning perspective change, forming a significant density gradient, thus revealing the geometric discontinuity and heterogeneity within the region. Therefore, the degree of change in the spatial gradient of the point density can be used as an important auxiliary index for identifying high-heterogeneity regions.

[0074] The specific steps for generating the density gradient reference value after in-depth analysis of the degree of change in the spatial gradient of the point density within the sub-unit area are as follows:

[0075] Divide the sub-unit area into equally spaced voxel grids, and assume that each voxel The side length is l For each voxel, calculate the sum of the density gradient transition intensities with its six face-adjacent voxels. The calculation expression is as follows:

[0076]

[0077] where: is the voxel and the local density transition intensity with its six face-adjacent voxels; is the index set of the six face-adjacent voxels adjacent to the voxel ; d represents the index of a face-adjacent voxel adjacent to the current voxel , and is used to enumerate all adjacent voxels in the six directions (up and down, front and back, left and right) in contact with the voxel ; , respectively represent the number of points in the current voxel and the adjacent voxel; is the non-linear enhancement index, used to enhance the response to high-density jumps (such as taking values from 1.5 to 2.5);

[0078] This step constructs the local density transition field, emphasizes the rapid spatial variation characteristics of density, and avoids the interference of smooth regions on the characterization of structural heterogeneity.

[0079] After obtaining the local density transition intensities of all voxels, a structure-sensitive non-linear weighted aggregation method is used to generate the density gradient reference value for the entire subunit region. The generated expression is as follows:

[0080]

[0081] where: is the density gradient reference value, used to characterize the heterogeneity level of the subunit region; w represents the voxel index within the subunit region; M is the total number of voxels within the subunit region; is the non-linear aggregation weight factor, used to further suppress weak gradients and highlight significant gradients, ; is used to compress large values and enhance small values, strengthening the discrimination ability of the density gradient reference value for high-variation regions;

[0082] Through non-linear transformation and aggregation, suppress the contribution of low-structure difference regions, highlight regions with significant structural mutations, and achieve high-sensitivity quantitative characterization of the subunit heterogeneity level.

[0083] It can be seen from the density gradient reference value that the larger the performance value of the density gradient reference value generated after in-depth analysis of the spatial gradient variation of the point density in the sub-unit area, the higher the level of geometric heterogeneity of the sub-unit area, and vice versa. The reason is that the density gradient reference value describes the spatial variation of the point density inside the sub-unit by performing enhanced nonlinear accumulation of the point density difference between each voxel and its adjacent voxels; when a region contains obvious structural transitions, occluded edges, interlaced objects or mutation surfaces, its point cloud density will show sharp changes in different spatial positions, resulting in a large local transition intensity, thereby pushing the overall density gradient reference value up; while in areas with smooth geometric shapes and uniform distribution of point density, the density gradient is weak, and the density gradient reference value also decreases accordingly. Therefore, the high or low density gradient reference value can effectively reflect the continuity or mutation degree of the structure in the region, and is an important measurement basis for geometric heterogeneity.

[0084] The analyzed indicators are input as feature vectors into a support vector machine model that has been trained with a large number of labeled samples in advance, and the degree of geometric heterogeneity corresponding to the sub-unit area is intelligently predicted;

[0085] The analyzed curvature discrete reference value and density gradient reference value are input as feature vectors into the support vector machine model that has been trained in advance with a large number of labeled samples. The heterogeneity level coefficient is generated by the support vector machine model, and the degree of geometric heterogeneity corresponding to the sub-unit area is intelligently predicted based on the heterogeneity level coefficient.

[0086] The trained support vector machine model refers to: based on a large number of known input features and label data, the support vector machine (SVM) algorithm is used for learning and optimization, and finally a set of model parameters that can be used for classification or regression tasks is obtained. In the training phase, the model separates or fits input samples with different output categories (or numerical levels) by constructing the optimal hyperplane; in the prediction phase, new, unlabeled input features (such as curvature discrete reference values and density gradient reference values) can be substituted into the model to obtain the corresponding output results (such as geometric heterogeneity levels or continuous heterogeneity coefficients), realizing intelligent identification and evaluation of the heterogeneity level of local sub-unit regions.

[0087] In this application background, the trained SVM model is a classification or regression model designed for the geometric heterogeneity discrimination task of a large range of point cloud data. Its training process includes multiple stages such as sample collection, feature extraction, label construction, model learning, and optimization. Specifically, first, a large number of representative point cloud samples are selected from multiple sub-unit regions with different scenarios and different structural complexities, and a multi-dimensional feature vector containing the discrete reference value of curvature and the reference value of density gradient is constructed for each sample. Subsequently, according to domain knowledge or expert evaluation results, a true "geometric heterogeneity level" label is assigned to each sample point cloud. For example, it can be divided into three levels: "high heterogeneity", "medium heterogeneity", and "low heterogeneity", or directly quantified as a continuous value (such as a heterogeneity coefficient) to form a standard training sample set. Then, these labeled feature samples are input into the SVM learning framework, and the optimal classification boundary or regression function is found through the principle of structural risk minimization.

[0088] The basic idea of the support vector machine is to find a hyperplane with the maximum margin in the feature space to separate samples of different classes, or to find an optimal function expression that fits within the tolerance error range in the regression problem. In the training process of SVM, by introducing kernel functions (such as RBF kernel, linear kernel, polynomial kernel, etc.), the input non-linear features can be mapped to a high-dimensional space, making the originally inseparable samples linearly separable in the new space. In this application, due to the highly non-linear and structurally coupled characteristics of the discrete reference value of curvature and the reference value of density gradient themselves, the Gaussian radial basis function (RBF) is usually used as the kernel function to enhance the model's ability to identify local change features.

[0089] The output of the trained SVM model includes the following key components: the support vector set, that is, the key training samples that constitute the optimal boundary; the Lagrange multipliers or weight coefficients, indicating the influence degree of each support vector on the decision boundary; the bias term b, indicating the offset of the hyperplane in the feature space; the kernel function type and parameter configuration, used for mapping new samples; the error tolerance and penalty factor C, controlling the balance between the model's fitting ability and generalization performance. After cross-validation and performance evaluation (such as accuracy, recall rate, F1-score, RMSE, etc.), the model is fixed for subsequent prediction tasks.

[0090] In the point cloud registration system, the trained SVM model plays a core role as a "local structure complexity recognizer". Its prediction process is as follows: For each local subunit region to be processed, the system first calculates its discrete reference value of curvature and reference value of density gradient, and constructs a corresponding two-dimensional feature vector; then, this vector is input into the trained SVM model, and the model judges the heterogeneity level (such as high, medium, low) to which the sample should belong through the learned decision function, or outputs a continuous heterogeneity level coefficient. It should be noted that the training effect of the SVM model highly depends on the diversity of samples and the accuracy of labels. In high-precision point cloud processing tasks, the training samples should cover various urban landform types (such as high-rise dense areas, open spaces, transition zones, structurally dense facilities, etc.), and take into account the differences in feature representations brought by different sensor sources (such as LiDAR, structured light, camera fusion data, etc.); in addition, the process of constructing labels requires the participation of domain experts or automatic annotation through an auxiliary rule system to ensure that the heterogeneity level reflects the true geometric complexity state.

[0091] In summary, the "trained support vector machine model" is not only a static classifier or regressor, but also a fundamental core module for realizing local structure cognition and parameter adaptive adjustment in the entire intelligent registration control system. By introducing this model, it can effectively replace manual rules and empirical judgments, realize the intelligent matching strategy control under "structure perception drive", and fundamentally solve the problems of insufficient adaptability, parameter solidification, and easy error diffusion in the current point cloud registration technology when dealing with local heterogeneity. This solution has high generalization ability and can be extended to other applications such as 3D modeling, SLAM, and map reconstruction, providing a solid intelligent recognition basis for building a highly robust and high-precision multi-source point cloud fusion system.

[0092] The support vector machine model is not specifically limited here, and any machine learning model that can realize the comprehensive analysis of the discrete reference value of curvature and the reference value of density gradient to generate a heterogeneity level coefficient is acceptable. To implement the technical solution of the present invention, the present invention provides a specific implementation method; the expression for generating the heterogeneity level coefficient is: , where , are the weight coefficients of the discrete reference value of curvature and the reference value of density gradient respectively, and , are both greater than 0. The weight coefficients (i.e., and in the formula))(It) refers to the proportional factor used to balance and regulate the relative contribution degrees of two different input indicators (curvature discrete reference value and density gradient reference value) in the process of generating the heterogeneous level coefficient. The setting of the weight coefficient reflects the system's judgment on the importance of different feature dimensions: when a certain feature is more significant in characterizing heterogeneity, its corresponding weight should be higher, and vice versa. By weighted combination of these two indicators, the heterogeneous level coefficient can more accurately reflect the true geometric heterogeneity state of the local subunit region. Therefore, the weight coefficient is not only a mathematical parameter, but also an implicit "feature importance trade - off mechanism" in the model.

[0093] From the heterogeneous level coefficient, it can be seen that the larger the performance value of the curvature discrete reference value generated by in - depth analysis of the discrete degree of the principal curvature value distribution of all points in the subunit region, and the larger the performance value of the density gradient reference value generated by in - depth analysis of the spatial gradient change degree of the point density within the subunit region, that is, the larger the performance value of the heterogeneous level coefficient generated by the intelligent prediction of the geometric heterogeneity degree corresponding to the subunit region by the trained support vector machine model, the higher the geometric heterogeneity level of the subunit region, and vice versa, it indicates that the geometric heterogeneity level of the subunit region is lower.

[0094] Based on the heterogeneity degree of the subunit region predicted by the support vector machine model, an adaptive mapping relationship between the heterogeneity degree and the registration parameters is constructed, and the corresponding upper limit of the number of matching points is dynamically allocated for subunit regions with different heterogeneity levels, so as to achieve the adaptation of the number of matching points to different regions;

[0095] Dynamically allocate the corresponding range of the number of matching points for subunit regions with different heterogeneity levels. The specific steps are as follows:

[0096] Obtain the heterogeneous level coefficient of each subunit region predicted by the support vector machine model, and construct a matching point adjustment factor for controlling the intensity of the number of matching points allocated based on the heterogeneous level coefficient. This matching point adjustment factor is calculated in the way of exponential mapping, and the calculation formula is:

[0097]

[0098] , where: represents the matching point adjustment factor driven by heterogeneity, and its value range is in ; is the maximum amplification coefficient, which controls the upper limit of the matching point adjustment factor and is used to adjust the resource tilt degree in high - heterogeneous regions, ; is the non - linear response adjustment factor, which is used to control the response speed of the matching point adjustment factor curve to the heterogeneous level coefficient; is the heterogeneous level coefficient of this subunit region;

[0099] The core advantage of the exponential function design lies in that when the heterogeneity coefficient is small, the matching point adjustment factor responds slowly, avoiding excessive resource investment in low-complexity regions; while when the heterogeneity coefficient approaches 1, the matching point adjustment factor rapidly amplifies, so as to concentrate more matching points in high-heterogeneity regions, realizing the control strategy of "focusing on key areas and smooth growth".

[0100] Based on the constructed matching point adjustment factor , the upper limit of the final number of matching points in this subunit region is dynamically allocated, and the expression of the dynamic allocation is as follows:

[0101]

[0102] , where: represents the upper limit of the number of matching points allocated to the current subunit region; represents the minimum lower limit of the number of matching points set by the system, which is used for low-heterogeneity regions (such as flat ground); represents the maximum upper limit of the number of matching points, which is applicable to high-heterogeneity regions (such as complex buildings or edge mixing areas);

[0103] The formula design reflects the regulation logic of "adapting to local conditions": when the heterogeneity level is low (i.e., ), the matching point adjustment factor , and at this time ; while when the heterogeneity level is high (i.e., ), , and at this time is close to or equal to . This allocation mechanism based on the exponential regulation factor can effectively avoid mis-matching caused by "excessive matching points", and can also provide sufficient geometric constraint support in complex regions, enhancing the local adaptability and global stability of point cloud registration.

[0104] By introducing a parameter regulation mechanism driven by heterogeneity, an intelligent control strategy of "dynamically adjusting parameter configuration due to regional structure differences" in the point cloud registration process is realized, thus effectively solving problems such as insufficient accuracy, high risk of misregistration, and unreasonable calculation resource allocation in the existing unified registration strategy when dealing with spatially heterogeneous regions. Specifically, for the geometric structure differences presented by different sub-unit regions in the large-scale point cloud registration task, this step first uses a support vector machine model to classify or quantitatively judge the heterogeneity degree of each region, and establishes a mapping relationship between the "heterogeneity degree" and the "registration control parameter (such as the upper limit of the number of matching points)" based on the model output results. Through this mapping mechanism, regions with high structural complexity and significant heterogeneity will obtain more support for the number of matching points to strengthen the local rigid transformation constraint and feature capture ability; while regions with single geometric features and low heterogeneity can reduce the matching point configuration to reduce the risk of redundant matching and error propagation. This strategy not only improves the registration accuracy and robustness of complex regions, but also optimizes the resource allocation efficiency of the entire system, which is the key support link for realizing high-precision and multi-scenario adaptive point cloud registration. Through the implementation of this step, the adaptability of the registration system to multi-source and heterogeneous environments can be significantly enhanced, effectively ensuring the geometric accuracy and global consistency of subsequent tasks such as 3D modeling, environmental perception, and navigation and positioning.

[0105] After setting the range of the number of matching points based on the regulation strategy, this matching point control parameter and other registration parameters (such as the matching radius, tolerance threshold) are jointly input into the subsequent point cloud registration module to achieve the adaptive registration of sub-unit regions;

[0106] Each sub-unit region can independently execute the establishment of initial matching pairs, local rigid transformation estimation, and perform consistency coordination through a global optimizer (such as pose graph or BA). This strategy not only improves the local accuracy of registration, but also significantly reduces the risk of global error conduction and accumulation, and finally a high-precision 3D point cloud model with topological continuity, geometric accuracy, and heterogeneous compatibility can be constructed.

[0107] Through the above method for generating and optimizing 3D street scenes driven by multi-source fusion point cloud data, it is possible to achieve intelligent local adaptive processing of large-scale heterogeneous point cloud data based on structural perception, significantly improving the registration accuracy and overall modeling consistency. By introducing a heterogeneity evaluation mechanism constructed based on the geometric features of sub-unit regions and combining with the trained support vector machine model, this method realizes the quantitative recognition and classification judgment of the local structure complexity, and dynamically adjusts the upper limit of the number of matching points accordingly, so that high-heterogeneity regions obtain more geometric constraints, and low-heterogeneity regions avoid redundant matching interference, thus effectively avoiding the registration error diffusion and structural misalignment problems caused by the unified parameter strategy, and improving the overall quality and stability of the street scene model in terms of shape restoration, boundary alignment, and topological continuity, providing a more robust, efficient, and intelligent technical support for application scenarios such as 3D city reconstruction, autonomous driving map construction, and precision surveying and mapping.

[0108] The above formulas are all dimensionless and take their numerical calculations. The formulas are obtained by collecting a large amount of data for software simulation to get a formula closest to the actual situation. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0109] Only some exemplary embodiments of the present invention have been described above by way of illustration. Undoubtedly, for those of ordinary skill in the art, without departing from the spirit and scope of the present invention, the described embodiments can be modified in various different ways. Therefore, the above drawings and description are illustrative in nature and should not be construed as limiting the protection scope of the claims of the present invention.

[0110] The above is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present application can easily think of changes or substitutions, which should all be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for generating and optimizing 3D street scenes driven by point cloud data with multi-source fusion, characterized in that Including the following steps: According to the spatial range of the input large-scale point cloud data and combining its three-dimensional coordinate system distribution characteristics, the overall registration area is divided into multiple sub-unit areas with equal area and uniform distribution by using a recursive spatial segmentation method based on octree, a fixed-resolution grid division, or a region-growing type of spatial adaptive segmentation algorithm; For each sub-unit area, obtain the geometric structure characteristics of the internal point cloud and construct a geometric feature descriptor containing multiple dimensions; Extract the key indicators reflecting the high heterogeneity of the sub-unit area from the obtained geometric structure characteristics through feature engineering techniques. Among them, the extracted key indicators include the degree of dispersion of the distribution of the principal curvature values of all points in the sub-unit area and the degree of spatial gradient change of the point density within the sub-unit area. After in-depth analysis of the extracted key indicators, a curvature dispersion reference value and a density gradient reference value are respectively generated, and the geometric heterogeneity level of each sub-unit area is quantified based on the curvature dispersion reference value and the density gradient reference value; Input the analyzed curvature dispersion reference value and density gradient reference value as feature vectors into a support vector machine model that has been trained in advance with a large number of labeled samples. Generate a heterogeneity level coefficient through the support vector machine model, and intelligently predict the geometric heterogeneity degree corresponding to the sub-unit area based on the heterogeneity level coefficient; Based on the heterogeneity degree of the sub-unit area predicted by the support vector machine model, construct an adaptive mapping relationship between the heterogeneity degree and the registration parameters, and dynamically allocate the corresponding upper limit of the number of matching points for sub-unit areas with different heterogeneity levels to achieve region-specific allocation of the number of matching points; After setting the range of the number of matching points based on the regulation strategy, input the matching point control parameter and other registration parameters into the subsequent point cloud registration module together to achieve the adaptive registration of the sub-unit area.

2. The method for generating and optimizing a 3D street view driven by multi-source fusion point cloud data according to claim 1, wherein The overall registration area is divided into multiple sub-unit areas with equal area and uniform distribution by using a recursive spatial segmentation method based on octree. The specific steps are as follows: Establish an overall bounding cube for the input large-scale point cloud data, which is used as the root node to represent the entire spatial area; Recursively divide the cube into eight equal parts according to the three-dimensional coordinate range of the point cloud. Each division evenly divides the parent node space into eight sub-cube areas, which are respectively added as child nodes to the octree structure; At each division stage, judge whether the number of point clouds in each sub-cube meets the minimum segmentation threshold. If it exceeds the minimum segmentation threshold, continue recursive subdivision until the preset minimum voxel size or the lower limit condition of the number of points is met. The finally generated leaf nodes are multiple sub-unit areas with equal area and uniform spatial distribution.

3. The multi-source fusion point cloud data-driven three-dimensional street view generation and optimization method according to claim 1, wherein, Dynamically allocate the corresponding range of the number of matching points for sub-unit areas with different heterogeneity levels. The specific steps are as follows: Obtain the heterogeneity level coefficient of each sub-unit area predicted by the support vector machine model, and construct a matching point adjustment factor for controlling the intensity of the number of matching points allocated based on the heterogeneity level coefficient. The matching point adjustment factor is calculated in an exponential mapping manner, and the calculation expression is: , where: represents the heterogeneity-driven matching point adjustment factor, whose value range is in ; is the maximum amplification factor, which controls the upper limit of the matching point adjustment factor and is used to adjust the resource tilt degree in high-heterogeneity regions, ; is the non-linear response adjustment factor, which is used to control the response speed of the matching point adjustment factor curve to the heterogeneity level coefficient; is the heterogeneity level coefficient of this subunit region; Based on the constructed matching point adjustment factor , the upper limit of the number of final matching points in this subunit area is dynamically allocated, and the expression for dynamic allocation is as follows: , where: represents the upper limit of the number of matching points obtained by the current sub-unit area allocation; represents the lower limit of the minimum number of matching points set by the system; represents the upper limit of the maximum number of matching points.

4. The method for generating and optimizing 3D street views driven by multi-source fusion point cloud data according to claim 1, characterized in that The specific steps for generating a curvature dispersion reference value after in-depth analysis of the degree of dispersion of the distribution of the principal curvature values of all points in the sub-unit area are as follows: For all points within the sub-unit region, the principal curvature value of each point is calculated using the k-nearest neighbor method , and the principal curvatures of all points within the sub-unit region are sorted in spatial order or density gradient to construct a curvature sequence ; A relative jump ratio factor is introduced to highlight the non-linear mutation characteristics between points with intense jumps. The calculation expression of the relative jump ratio factor is as follows: , where: is the principal curvature value of the i -th point; is a small constant; is the relative jump ratio factor, indicating the logarithmic jump amplitude between the i -th curvature point and the -th curvature point; N is the total number of points within the subunit region; Constructing a jump ratio sequence based on a relative jump ratio factor , after obtaining the jump ratio sequence, strengthen the weights of the locally discrete prominent regions by constructing a rate of change enhancement function, and finally generate a curvature discrete reference value. The generation expression of the curvature discrete reference value is as follows: , where: is the discrete reference value of curvature; represents the local change rate of the curvature jump ratio sequence in the neighborhood of the i -th curvature point, and is used to capture the trend fluctuation information of the severity of curvature change in the subunit region; is the discrete amplification factor, which is used to nonlinearly enhance the overall heterogeneity trend.

5. The method for generating and optimizing 3D street scenes driven by multi-source fusion point cloud data according to claim 1, wherein, The specific steps for generating the density gradient reference value after deeply analyzing the degree of spatial gradient change of the point density within the subunit region are as follows: Divide the sub-unit region into a voxel grid with equal spacing, and assume that each voxel has a side length of l . For each voxel, calculate the sum of the density gradient transition intensities with its six face-adjacent voxels. The calculation formula is as follows: , where: is a voxel is the local density transition intensity of the voxel adjacent to the six - face voxels; is the voxel is the index set of the six - face voxels adjacent to the voxel; d represents the index of a face - adjacent voxel adjacent to the current voxel, used to enumerate all adjacent voxels in the six directions in contact with the voxel ; is the number of points in the current voxel and the adjacent voxels respectively; and ; is the non - linear enhancement index; After obtaining the local density transition intensities of all voxels, a structure-sensitive non-linear weighted aggregation method is used to generate the density gradient reference value for the entire subunit region. The generated expression is as follows: , where: is the density gradient reference value, used to characterize the heterogeneity level of the subunit region; w represents the voxel index within the subunit region; M is the total number of voxels within the subunit region; is the non-linear aggregation weight factor, .

Images