Multi-source point cloud adaptive optimization and feature adaptation method

Abstract

The present application belongs to the field of surveying and mapping science and technology, and discloses a multi-source point cloud adaptive optimization and feature adaptation method. After the multi-source point cloud basis is combined, an adaptive buffer zone is dynamically generated based on the collection track and local density, and a splicing area and a general area are divided. Multidimensional prior analysis is performed on the two types of areas, and multi-scale adaptive division of ground objects is realized based on the space distribution continuity, instance-point number balance and geometric shape complexity criteria. Then, adaptive resampling is performed on the general area to preserve features, and reinforced resampling is performed on the splicing area to balance density uniformity and feature preservation. Finally, a structured point cloud sample with area type, scale coding, category proportion and density feature label is fused and generated. The present application can effectively solve the problem that fixed parameter sampling in the existing point cloud preprocessing cannot adapt to multi-scale ground objects, long-tail distribution of categories, and density distortion of splicing areas, and significantly improves the quality of point cloud samples and the segmentation accuracy of downstream AI models.

Description

Technical Field

[0001] This invention belongs to the field of surveying and mapping science and technology, and in particular relates to a method for adaptive optimization and feature matching of multi-source point clouds. Background Technology

[0002] As a core data source for producing 3D spatiotemporal information of roads, laser point cloud data requires intelligent processing. This necessitates first completing 3D scene semantic segmentation and then extracting geometric vectors. High-quality preprocessing and sample preparation are crucial prerequisites for supporting subsequent AI training and ensuring model performance. However, existing point cloud preprocessing technologies have significant shortcomings. Currently, road point cloud acquisition is mainly conducted via vehicle-mounted and motorcycle-mounted methods. Due to significant differences in data density and quality, the heterogeneity of the data increases the difficulty of data processing and preparation. Directly using multi-source point cloud data for deep learning model training still faces multiple challenges, failing to meet the high-precision 3D data requirements of fields such as autonomous driving and smart cities. Ultimately, this results in low model segmentation accuracy and weak generalization ability.

[0003] Specifically, the core limitations of existing point cloud preprocessing technologies and the core challenges of training data are reflected in four aspects: 1) Insufficient resampling and data scale adaptation. Existing methods use fixed parameters and uniform sampling rates, which cannot adapt to multi-scale ground feature features, easily causing loss of large ground feature details or submersion of small ground feature features. At the same time, a single road point cloud can reach billions of points, and direct loading and computation place extremely high demands on hardware. Fixed-ratio random sampling, regular grid sampling, and other methods are difficult to reduce the amount of data while taking into account feature preservation and efficiency improvement; 2) The density and category balance problem is prominent. The spatial density of the data is uneven and the number of points in each category varies greatly. For example, ground points account for more than 70%, while key small objects account for more than 10%. The current point cloud data preprocessing method has several shortcomings. First, the proportion of data points in the preprocessed cloud is less than 0.1%. Existing sample preparation lacks multi-dimensional analysis, and class balancing relies solely on simple data augmentation, which fails to address the long-tail distribution problem. Traditional uniform random downsampling also disrupts the structure of small features and exacerbates imbalances. Second, density imbalances and feature distortions easily occur in multi-source point cloud stitching regions. Existing methods do not perform differentiated processing for different regions, resulting in poor sample quality in stitched regions and affecting the overall model training effect. Third, the preprocessed point cloud features have poor adaptability to downstream 3D semantic segmentation models. Existing preprocessing only focuses on data format and scale, without establishing a connection mechanism between preprocessed features and the model network, leading to low model feature extraction efficiency and poor utilization. Therefore, a point cloud data preprocessing method with multi-scale adaptive resampling capabilities, dynamic class balancing, and the ability to establish a connection between preprocessed features and downstream AI model features is urgently needed. Specifically, a differentiated processing strategy should be designed for multi-source point cloud stitching regions, while simultaneously achieving deep adaptation between preprocessed features and the model, breaking through existing technical bottlenecks. Summary of the Invention

[0004] This invention primarily addresses the challenge of unifying the basic features and structural quality of data before large-scale model training, providing support for mapping multi-source data to the same dimensional space. Specifically, it aims to solve the core bottleneck problems of existing point cloud preprocessing and sample preparation techniques for AI training, including: 1) Fixed parameters and uniform sampling rate resampling cannot adapt to multi-scale ground features, making it difficult to balance data reduction, processing efficiency, and key feature retention; 2) The single category balancing strategy and lack of data feature support in sample preparation fail to alleviate the long-tail distribution problem, resulting in low training sample quality; 3) Density imbalance and feature distortion are prone to occur in multi-source point cloud stitching regions. Existing methods do not perform regional differentiation processing, resulting in poor sample quality in stitching regions and affecting the overall model training effect.

[0005] The purpose of this invention is to provide a multi-source point cloud adaptive optimization and feature adaptation method, overcoming the limitations of existing technologies to achieve high-precision, high-efficiency, and highly adaptable point cloud sample preparation. 1) By designing a multi-scale adaptive resampling mechanism, parameters are dynamically adjusted to reduce data scale and computational costs while preserving key features of ground objects; 2) An intelligent dynamic category balancing system is constructed, combining data feature optimization partitioning logic and employing targeted strategies to alleviate long-tail distribution of categories and improve sample quality; 3) A regional quantification analysis and differentiated resampling strategy is designed to address the feature differences between multi-source point cloud splicing regions and general regions, alleviating density imbalance and feature distortion problems in splicing regions. Ultimately, through the above design, the quality of training samples and model adaptability are improved, semantic segmentation accuracy and generalization ability are enhanced, providing core technical support for fields such as autonomous driving and smart cities, and meeting the high-standard requirements of AI training data.

[0006] To achieve the above objectives, this invention provides a method for adaptive optimization and feature adaptation of multi-source point clouds, comprising the following steps: Step 1: Collect vehicle-mounted and motorcycle-mounted laser point cloud data, perform basic merging of multi-source point clouds, define a full point cloud sample space, and divide the spliced ​​region and general region; Step 2: Based on the basic merged full point cloud sample space, perform multi-dimensional prior analysis on the labeled point cloud dataset according to the general region and the spliced ​​region, and generate a full region multi-dimensional analysis summary table; Step 3: Based on the multi-dimensional analysis results, perform multi-scale adaptive classification of land cover categories, and output scale encoding and continuous / discrete land cover classification table; Step 4: To address class imbalance, perform differentiated adaptive resampling on the spliced ​​region and the general region respectively, and generate labeled structured point cloud samples for downstream AI model training.

[0007] Furthermore, step 1 specifically includes:

[0008] Step 1.1: Merge the vehicle-mounted and motorcycle-mounted laser point clouds in the same coordinate system and LAS format;

[0009] Step 1.2: Extract the acquisition trajectory coordinates and acquisition device parameters from the vehicle / motorcycle point cloud acquisition trajectory file:

[0010] Step 1.3: Obtain basic information using a single acquisition trajectory from a vehicle / motorcycle: calculate the effective laser point cloud coverage width on the road surface as the basic value for the buffer radius;

[0011] Step 1.4: For each data collection trajectory Perform segmented local density calculations;

[0012] Step 1.5: Utilize multiple acquisition trajectories from vehicle-mounted and motorcycle-mounted equipment to calculate the overlapping intervals between trajectories and adaptively determine the buffer radius;

[0013] Step 1.6: The intersection area of ​​all vehicle / motorcycle local adaptive buffer zones is the stitching area, and the remaining non-intersection area is defined as the general area;

[0014] Step 1.7: For locations where the buffer radius of adjacent trajectory segments changes abruptly, linear interpolation is used for transition.

[0015] Furthermore, step 2 specifically includes:

[0016] Step 2.1: Multi-dimensional basic analysis of general areas:

[0017] For general areas with no density abrupt changes and relatively uniform point cloud distribution, basic quantitative analysis is carried out from three dimensions: spatial distribution, proportion of category points, and geometric morphology, to calculate core indicators and determine characteristic attributes;

[0018] Step 2.2: Multi-dimensional adaptation analysis of the splicing area:

[0019] For stitched areas where local density fluctuations and slight distortions in geometric features are prone to occur after multi-source point cloud stitching, density uniformity supplementary analysis is added to the framework of general area three-dimensional basic analysis, and the core index formulas and judgment thresholds are adapted and modified.

[0020] Step 2.3: Integration of all regional indicators:

[0021] The quantitative analysis results of the general region and the stitched region are merged to generate a multi-dimensional analysis summary table of the full point cloud region, which marks the feature differences, specific indicators and correction parameters between the general region and the stitched region.

[0022] Furthermore, step 3 specifically includes:

[0023] Based on the analysis results of step 2, the spatial distribution continuity criterion, instance point balance criterion, and geometric morphology complexity criterion are used to divide the general area and the spliced ​​area into land feature scales, respectively, and output scale codes and continuous or discrete land feature classification tables; among them, the judgment threshold of the spliced ​​area is adapted and corrected compared with that of the general area.

[0024] Furthermore, step 4 specifically includes: implementing a full-process differentiation strategy based on adaptive downsampling of the general region and adaptive downsampling of the stitched region, as follows:

[0025] Step 4.1: Adaptive downsampling and smoothing in general regions:

[0026] For general areas with uniform point cloud distribution and no abrupt density changes, a basic smoothing + conventional adaptive downsampling strategy is implemented. The parameter configuration fits the features of general areas without additional correction.

[0027] Step 4.2: Adaptive downsampling and smoothing of the stitching region:

[0028] For splicing regions that still exhibit local density fluctuations and slight geometric feature distortions after multi-source splicing, an enhanced smoothing + adaptive downsampling strategy is implemented. Based on the general region process, core parameters are modified, density constraints are increased, and verification thresholds are relaxed.

[0029] Step 4.3: Region fusion after resampling:

[0030] The point cloud is fused with the completed regional adaptive downsampling and smoothed general and spliced ​​regions to form the full point cloud space after resampling. Region type label, scale encoding label, category proportion label and density feature label are added to each point cloud feature.

[0031] Furthermore, step 4.1 includes the following specific steps:

[0032] Step 4.1.1: Point cloud smoothing and denoising: The original point cloud position is corrected using the moving least squares method;

[0033] Step 4.1.2: Local region segmentation: The point cloud is segmented and smoothed using an unbalanced octree, and the voxel size is fixed based on the general regional spatial distribution analysis results;

[0034] Step 4.1.3: Local Continuity Measurement: For each point, calculate the continuity score by combining the scale code of its associated land cover with the normal consistency variance;

[0035] Step 4.1.4: Sampling density weight allocation: The sampling weights are dynamically adjusted based on the continuity score, the proportion of class points, and the variance of normal consistency.

[0036] Step 4.1.5: Weighted farthest point sampling: Calculate the weighted distance between points, iteratively select the point with the largest weighted distance, until the preset number is reached;

[0037] Step 4.1.6: Sampling effect verification: The sampling effect is verified using the mean absolute error of curvature consistency.

[0038] The threshold for determining the mean absolute error of continuously distributed, structurally simple ground features is... Mean absolute error of discretely distributed and structurally complex ground features If the condition is not met, return to step 1 for resampling to ensure that the geometric features of the point cloud are not significantly distorted after sampling.

[0039] Furthermore, step 4.2 includes the following specific steps:

[0040] Step 4.2.1: Employ a quadratic MLS smoothing strategy, with the weighting function consistent with that of the general region, combined with the density variation coefficient of the spliced ​​region. Dynamically adjust the radius of influence Simultaneously, density constraints are introduced to ensure the coefficient of variation of mesh density in the splicing region during the smoothing process. ;

[0041] Step 4.2.2: Use an unbalanced octree for segmentation, reducing the voxel size by 20% compared to the general region;

[0042] Step 4.2.3: Based on the general region continuity scoring formula, add a correction term for the density transition smoothness of the splicing region for correction;

[0043] Step 4.2.4: Based on the general regional weight allocation formula, a weight tilt coefficient is added to correct the weight of a few land cover types in the spliced ​​area;

[0044] Step 4.2.5: Weighted farthest point sampling: The weighted distance formula is the same as that for general areas, and double adjustments are made to adapt the stitching area;

[0045] Step 4.2.6: Verify the sampling effect using the mean absolute error of curvature consistency, and appropriately relax the verification threshold. The judgment threshold is: the mean absolute error of continuously distributed and structurally simple ground features. Mean absolute error of discretely distributed and structurally complex ground features If the threshold is not met, return to step 1 for resampling.

[0046] Furthermore, the dual adjustment in step 4.2.5 includes the following steps:

[0047] a) Configure the sampling rate of the stitched area to be 5% higher than that of the general area, specifically: the sampling rate range for continuously distributed features is 25%-35%, and the sampling rate range for discretely distributed features is 55%-65%;

[0048] b) Introduce density uniformity constraints to ensure that the percentage of points in any local grid within the stitched area does not fluctuate by more than 10% during the weighted farthest point sampling process.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] 1. Improved consistency of multi-source data: Merging point clouds from different acquisition methods, balancing and smoothing density, solving the problems of data connection gaps and density imbalances from the source, and laying the foundation for high-quality data.

[0051] 2. Intelligent data compression and feature preservation: By using differentiated downsampling and filtering of salient values ​​of normal features, compared with uniform sampling, it can significantly maintain the geometric integrity and key feature points of complex land features at the same compression rate, providing the model with higher quality feature learning materials.

[0052] 3. Alleviating class imbalance from the source: By adopting a four-pronged strategy of "continuity scoring + class proportion weighting + weighting of minority classes in spliced ​​regions + weighting of feature encoding classes", the total amount of data is reduced while objectively adjusting the proportion of points of different structural features in the dataset. This effectively solves the problem of long-tail distribution of classes and lays a data foundation for the model to learn all classes fairly.

[0053] 4. Order-of-magnitude improvement in preprocessing efficiency: Multi-source point clouds are only merged, reducing ineffective calculations without innovation; combined with multi-threaded parallel processing and efficient algorithms, the preprocessing time of hundreds of GB of urban road multi-source point cloud data is reduced from hours to minutes, improving processing efficiency by more than 90%, and significantly accelerating the AI ​​model development and iteration cycle. Attached Figure Description

[0054] Figure 1 This is a diagram illustrating the overall framework of the multi-source point cloud adaptive optimization and feature adaptation method of the present invention.

[0055] Figure 2 This is a schematic diagram of a local viewpoint in normal downsampling.

[0056] Figure 3 This is a schematic diagram of a local viewpoint for adaptive downsampling. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0058] This invention provides a method for adaptive optimization and feature adaptation of multi-source point clouds. Taking a common scenario of urban road vehicle / motorcycle point clouds (point cloud density 500-1000 points / ㎡, which is the mainstream application scenario in the industry) as an implementation example, the method first completes the basic merging and regional division of multi-source point clouds, and then performs full-process regional differentiated processing according to the core technical solution. The practical parameters, calculation process, threshold settings and adjustments for each step are explained below to ensure the reproducibility and engineering feasibility of the method:

[0059] Step 1: Preliminary Operations: Basic Merging and Region Division of Multi-Source Point Clouds

[0060] Step 1.1: Merge the vehicle-mounted and motorcycle-mounted laser point clouds in the same coordinate system and LAS format, and define the full point cloud sample space as follows. Each point contains Three-dimensional coordinates Basic information on strength values;

[0061] Step 1.2: Extract the acquisition trajectory coordinates and acquisition device parameters from the vehicle / motorcycle point cloud acquisition trajectory file:

[0062] (a) Acquisition trajectory coordinates: Obtain the discrete three-dimensional coordinate point set for each acquisition trajectory. The points are then projected onto a two-dimensional plane (in urban road scenarios, the impact of z-axis elevation differences on the buffer radius is negligible), resulting in a set of surface trajectory points. ;

[0063] (b) Acquisition of equipment parameters: Extracting fixed parameters of the lidar such as scanning field of view (FOV), scanning radius (effective ranging), and installation height / angle (denoted as...). ).

[0064] Step 1.3: Obtain basic information using a single acquisition trajectory from a vehicle / motorcycle: Calculate the effective laser point cloud coverage width on the road surface (i.e., the effective scanning range on one side of the trajectory), as the base value for the buffer radius:

[0065] (a) Calculate the theoretical coverage width based on the lidar parameters. ;

[0066] (b) Correcting theoretical values ​​based on actual point cloud spatial distribution: Statistically analyze the point cloud density distribution on both sides of the trajectory, using "point cloud density ≥ 0.5 points / m² (minimum density for effective identification of urban roads)" as the threshold to determine the actual effective coverage width. (like ,Pick Otherwise take );

[0067] (c) Final value of effective coverage width of a single trajectory (One-sided).

[0068] Step 1.4: For each data collection trajectory Perform piecewise local density calculations:

[0069] (a) Arrange the two-dimensional trajectory according to the order of the trajectory points. Divide the road into several continuous small segments according to a fixed step length (e.g., 1m to 5m, which can be adjusted according to road accuracy). ;

[0070] (b) Establish a path with a side length of centered at the midpoint of the trajectory segment. Square partial window ( (Take a range of 2-5m) and count the number of point clouds within the window. Calculate local point cloud density ;

[0071] (c) Based on an effective identification density of 0.5 points / m² for urban roads, density levels are divided into: sparse areas. Normal area Dense areas .

[0072] Step 1.5: Using multiple acquisition trajectories from vehicle-mounted and motorcycle-mounted systems, calculate the overlapping area between the trajectories and adaptively determine the buffer radius.

[0073] (a) Calculate any two adjacent trajectories , planar distance (Take the shortest Euclidean distance from the set of trajectory points);

[0074] (b) Calculate the sum of the effective coverage widths of the two trajectories. ( The effective width of the vehicle trajectory. (Effective width of the motorcycle track).

[0075] (c) Calculate the overlap and base buffer radius: If (If there is overlap), then the buffer radius (The distance from the centerline of the overlapping region to the trajectory, i.e., the optimal buffer radius); if (No overlap): Take the average of the effective coverage widths of the two types of trajectories. (A reasonable safety margin for road width);

[0076] (d) For each trajectory segment, the base buffer radius is dynamically adjusted according to its local density level to obtain the segment-by-segment local buffer radius. , where the sparse region Appropriately expand the buffer radius to avoid missing effective points in sparse road sections and ensure complete coverage of the splicing area; normal area Maintaining the basic buffer radius unchanged to adapt to the needs of regular road sections; densely populated areas Appropriately reduce the buffer radius, reduce the inclusion of redundant points in dense road sections, and reduce the computational load of subsequent registration;

[0077] (e) Global calibration: Traverse all The median smoothing method is used to remove interference from outlier segments (such as density abrupt changes caused by local noise) to obtain the final global adaptive buffer radius sequence, which takes into account both local adaptation and overall robustness, and avoids the influence of single outlier segments on the global partition.

[0078] Step 1.6: Using the trajectory centerline as a reference, divide the trajectory into segments and use the corresponding... A two-dimensional strip buffer with adaptive width is generated to replace the traditional global uniform-width buffer, realizing the dynamic change of buffer width with local point cloud density. The intersection region of all vehicle / motorcycle local adaptive buffers is the stitching region. The remaining non-intersecting regions are defined as general regions. ,satisfy , This is the complete set of point cloud sample space, representing the merging of vehicle-mounted and motorcycle-mounted laser point clouds, and encompassing all point cloud data. All subsequent core processes are based on this region and perform differentiated processing accordingly.

[0079] Step 1.7: For locations where the buffer radius of adjacent trajectory segments changes abruptly, linear interpolation is used for transition to avoid hard corners and jagged edges in the buffer zone, ensuring the continuity of the splicing area boundary, solving the boundary discontinuity problem of traditional methods, and improving the stability of subsequent regional processing.

[0080] Step 2: Multi-dimensional prior analysis of point clouds:

[0081] Based on the merged sample space The labeled point cloud dataset is divided into general regions. and splicing area Conduct differentiated quantitative analysis and output key indicators for subsequent processes:

[0082] Step 2.1: General Area Multi-dimensional fundamental analysis:

[0083] For general areas with no abrupt density changes and relatively uniform point cloud distribution Basic quantitative analysis was conducted from three dimensions: spatial distribution, category point ratio, and geometric morphology. Core indicators were calculated and characteristic attributes were determined to provide a basis for subsequent processing.

[0084] Step 2.1.1: Spatial Distribution Analysis: Using the macroscopic spatial continuity rate (… ) and clustering index ( The dual-index quantification of spatial continuity characteristics of ground features differs from the subsequent calculation of microscopic continuity of local surface areas in point cloud sampling. This step focuses on the overall distribution patterns of ground features:

[0085]

[0086]

[0087] in, This represents the number of points within the smallest bounding box (MBR) for this type of land feature. , These are the standard deviation and mean of the local point cloud density within the bounding box, respectively; , , This represents the standard deviation of the coordinates of the ground feature points. The closer to 1, The smaller the value, the more continuous the distribution of ground features; conversely, the larger the value, the more discrete the distribution, and the more discrete the output categories. , Quantified value.

[0088] Step 2.1.2: Category Point Count Proportion Analysis: Statistically analyze the number and proportion of point clouds of each category within the general area G, identify long-tail distribution characteristics, output a table of category point count proportions for the general area, set the minority class determination threshold to point count proportion < 0.5% (this value is set based on the distribution law of urban road point cloud categories, and can be adjusted between 0.3% and 0.8% according to the actual degree of long tail), and clarify the boundary between "majority class - minority class".

[0089] Step 2.1.3: Geometric Morphology Analysis: Through normal consistency variance ( ), length-width-height ratio ( ), shape regularity ( Three core quantitative indicators are used to extract the geometric attributes of land features, outputting a quantitative table of general regional geometric morphology. The indicator calculation and judgment threshold settings (the threshold is the optimal judgment value of the structural characteristics of ordinary land features along urban roads) are as follows:

[0090] a) Normal consistency variance ( Point cloud collections for each type of land cover Calculate each point k-nearest neighbor set Fitting local normal vectors Then, the concentration of the normal directions of all land features is quantified by variance:

[0091]

[0092] in, The global average normal vector. The angle between the average normals.

[0093] when It was determined to be simple in structure. The structure was determined to be complex at that time;

[0094] b) Length-width-height ratio ( ): Calculate the length (L), width (W), and height (H) of features using the minimum bounding box (MBR), and take the ratio of the maximum to the minimum value:

[0095]

[0096] when It was determined to be linear / slender type. It is then determined to be planar / blocky / irregular in shape;

[0097] c) Shape regularity ( ): Calculate the convex hull volume of ground features Volume of the circumscribed cuboid The ratio:

[0098]

[0099] when When it is determined to be a rule structure, It was determined to be an irregular structure.

[0100] Step 2.2: Multi-dimensional adaptation analysis of the splicing area:

[0101] For stitched areas where local density fluctuations and slight geometric feature distortions are prone to occur after multi-source point cloud stitching. Based on the framework of general three-dimensional basic analysis, a supplementary analysis of density uniformity is added, and the core index formulas and judgment thresholds are adapted and modified (the purpose of the modification is to eliminate the interference of density fluctuations on the analysis results and make the judgment results more consistent with the actual characteristics of the spliced ​​area), as follows:

[0102] Step 2.2.1: Spatial Distribution Analysis (Indicator Correction): In general areas , Based on the calculations, a density correction factor is introduced. The formula for the spatial continuity of the spliced ​​region is corrected as follows: ; The calculation rules are the same as for general regions, and the output is the stitched region. Various , Quantified value.

[0103] Step 2.2.2: Category Point Percentage Analysis (Same as General Area): Statistical Analysis of the Composite Area The table outputs the number and proportion of point clouds of each category within the spliced ​​region, and links it with the general region table to form the distribution characteristics of all region categories, providing a basis for the weight allocation of different regions.

[0104] Step 2.2.3: Geometric Morphology Analysis (Threshold Adaptation): The core indicator calculation formula is the same as for general areas. To address the slight distortion of geometric features that easily occurs in splicing areas, the geometric morphology judgment threshold is appropriately relaxed (by approximately 20%, which both adapts to the actual situation of feature distortion in splicing areas and avoids excessive deviation in the judgment results). A quantification table of the geometric morphology of the splicing area is output. The adjusted threshold is:

[0105] a) Threshold for determining structural simplicity: ;

[0106] b) Linear / Elongated Type Judgment Threshold: ;

[0107] c) Threshold for determining regular structure:

[0108] Step 2.2.4: Supplementary analysis of density uniformity: Calculate the density variation coefficient as a core dimension for the differential analysis of the spliced ​​area. Smoothness of density transition Two dedicated metrics are used to output a "Quantitative Table of Density Uniformity in the Stitching Region," providing a core basis for subsequent smoothing and resampling parameter configuration in the stitching region. The formula and judgment threshold (the threshold is the qualified judgment standard for the density uniformity of the stitching region; if it is not met, it needs to be strengthened in the subsequent smoothing step) are as follows:

[0109] a) Coefficient of density variation : , To ensure uniform density, Due to uneven density, among which To show the degree of deviation of the point cloud density of each grid within the stitched area from the mean, This is the arithmetic mean of the point cloud density of all grids within the stitched area;

[0110] b) Density transition smoothness : (n is the number of grid cells in the splicing region, (Number of point clouds within a single grid / Grid area) The smaller the value, the smoother the density transition.

[0111] Step 2.3: Integration of all regional indicators:

[0112] general area splicing area The quantitative analysis results are integrated to generate the "Full Point Cloud Region Multidimensional Analysis Summary Table", which marks the feature differences, specific indicators and correction parameters of the two types of regions. This provides a refined and differentiated quantitative basis for subsequent multi-scale land cover division and regional resampling, ensuring that subsequent processing strategies are consistent with the point cloud characteristics of different regions.

[0113] Step 3: Adaptive classification of land cover categories at multiple scales:

[0114] Based on the results of multi-dimensional analysis and fusion of different regions, and on the basis of the three-dimensional division criteria of "spatial distribution continuity, instance-point balance, and geometric complexity", the spliced ​​region is further analyzed. The judgment threshold is adjusted for adaptation (the adjustment principle is consistent with geometric morphology analysis, slightly relaxed to adapt to the characteristics of the spliced ​​area), realizing automated and standardized division of land features at the land feature scale; the division results are labeled as "general area land features" and "spliced ​​area land features", and scale codes are output. The classification table for (continuous feature 1, discrete feature 0) and continuous / discrete features is as follows, with specific criteria and threshold settings:

[0115] Step 3.1: Spatial Distribution Continuity Criterion: Generally, continuously distributed land features in a region satisfy the following criteria. , Discretely distributed ground features satisfy , The continuous features in the splicing area meet the requirements. , Discretely distributed ground features satisfy , ; This represents the macroscopic spatial continuity rate for a general region. To ensure the continuity of the macroscopic space in the splicing region;

[0116] Step 3.2: Instance-Point Balance Criterion: Continuous features have a larger total number of points but fewer instances, while discrete features have a sparse total number of points but more instances. The judgment rule is the same in general areas and spliced ​​areas.

[0117] Step 3.3: Geometric morphological complexity criterion: In general areas, continuous land features are mainly characterized by regular and simple structures. , Discrete ground features are mainly characterized by complex and irregular structures. , The continuous features in the spliced ​​area are mainly of regular and simple structure. , Discrete ground features are mainly characterized by complex and irregular structures. , ).

[0118] Step 4: Adaptive resampling by region for class imbalance:

[0119] A differentiated strategy is implemented for adaptive downsampling of general regions and stitched regions. Both types of regions include a six-step process: "smoothing and denoising - local segmentation - continuous quantization - weight allocation - weighted sampling - effect verification". The core differences lie in parameter configuration, smoothing strategy, and sampling rules, achieving the dual goals of "preserving features in general regions and preserving uniformity and features in stitched regions", as detailed below:

[0120] Step 4.1: General Area Adaptive downsampling and smoothing:

[0121] For a general region G with uniform point cloud distribution and no abrupt density changes, a basic smoothing + conventional adaptive downsampling strategy is implemented. The parameter configuration fits the characteristics of the general region without additional correction. The core steps are as follows:

[0122] Step 4.1.1: Point Cloud Smoothing and Denoising: The original point cloud positions are corrected using the Moving Least Squares (MLS) method, with the following weighting function: Where d is the distance between points, and the influence radius h is dynamically adjusted according to the scale of the ground features: h=0.5m for continuous ground features and h=0.2m for discrete ground features (this value is the optimal value for smoothing urban road ground features and can be adjusted within ±0.1m according to the scale of the ground features). Only basic smoothing is performed to avoid over-processing and loss of features.

[0123] Step 4.1.2: Local region segmentation: The point cloud is segmented and smoothed using an unbalanced octree. The voxel size is fixed based on the general regional spatial distribution analysis results: 0.3m×0.3m×0.3m for continuous distribution and simple structure (suitable for large-area continuous distribution), and 0.1m×0.1m×0.1m for discrete distribution and complex structure (voxel size matches the scale of ground features to avoid splitting discrete ground features and overly fine segmentation of continuous ground features).

[0124] Step 4.1.3: Local Continuous Measurement: For each point Combined with the scale code of its respective land feature Consistency variance with normal The formula for calculating the continuity score is:

[0125]

[0126] in , , (The core logic is that continuous features emphasize the consistency of curvature and normals, while discrete features emphasize surface roughness.) For curvature, For surface roughness, These represent the minimum and maximum curvature values ​​for the corresponding land cover categories. These represent the minimum and maximum normal consistency variances for the corresponding land cover categories. These represent the minimum and maximum surface roughness values ​​for the corresponding land cover categories.

[0127] Step 4.1.4: Sampling density weight allocation: based on continuity score , and the percentage of points by category Consistency variance with normal Dynamically adjust sampling weights to avoid oversampling of a minority of discrete land cover types:

[0128]

[0129] in , , (Percentage adjustment coefficient) (Complexity correction factor) (Adaptive sensitive parameter) The smaller, The larger, For point cloud single point The sampling weight value, A higher baseline value allows for increased weighting of minority and highly complex features, thus preventing oversampling. For a single point The continuity score power term, continuity score of Power of 1 To fix the adaptive sensitivity parameter.

[0130] Step 4.1.5: Weighted Farthest Point Sampling: Perform improved FPS, the weighted distance between points is defined as:

[0131]

[0132] in For point and points The three-dimensional Euclidean distance between them (Scaling factor) Candidate points The sampling weight value, continuous ground features ( Appropriately increase the effective sampling distance and reduce redundant points; discrete ground features ( ) When dealing with large (complex) features, reduce the distance correction margin and prioritize retaining structural points. Iterate through the selection of the maximum weighted distance points until a preset number is reached (20%-30% sampling rate for continuous features, 50%-60% sampling rate for discrete features, based on...). (Dynamically adjusted).

[0133] Step 4.1.6: Sampling effect verification: Verification is performed using the mean absolute error (MAE) of curvature consistency, with the following formula:

[0134]

[0135] in, Let i be the i-th point in the sampled point cloud set Q. The total number of points in the sampled point cloud set Q. For the i-th point after sampling The calculated value of curvature, Before sampling, with Corresponding origin point The true value of curvature. The judgment threshold is for continuously distributed, structurally simple ground features. (Relax thresholds, prioritize preserving overall structure) Discretely distributed, structurally complex ground features: (Strict threshold to avoid loss of detail) If not met, return to step 1 to resample to ensure that the geometric features of the point cloud are not significantly distorted after sampling.

[0136] Step 4.2: Joining Area Adaptive downsampling and smoothing:

[0137] For stitched region I, which still exhibits local density fluctuations and slight geometric feature distortions after multi-source stitching, an enhanced smoothing + adaptive downsampling strategy is implemented. This involves modifying core parameters, increasing density constraints, and relaxing the validation threshold based on the general region processing procedure (all modifications are adapted to the stitched region features, enhancing density uniformity while compensating for slight feature loss during smoothing). The core steps are as follows:

[0138] Step 4.2.1: Enhanced Smoothing and Denoising (Secondary MLS Smoothing + Density Constraint): A secondary MLS smoothing strategy is adopted (secondary smoothing enhances the density uniformity of the splicing region, which is different from the single basic smoothing of the general region). The weighting function is the same as that of the general region, combined with the density variation coefficient of the splicing region. Dynamically adjust the radius of influence The formula is ,in This represents the radius of influence of the corresponding land features in a typical area. The larger (the more uneven the density). The larger the value, the smoother the density transition zone; at the same time, density constraints are introduced to ensure the coefficient of variation of mesh density in the splicing area during the smoothing process. To avoid excessive smoothing that could lead to the loss of ground features.

[0139] Step 4.2.2: Local Region Segmentation (Vocet Reduction): Unbalanced octree segmentation is adopted, and the voxel size is reduced by 20% based on the general region (the purpose of reduction is to improve the segmentation accuracy of the splicing region, adapt to local density fluctuations, and ensure that small features are not split and the density is more uniform). The configuration is as follows: continuous distribution and simple structure features 0.24m×0.24m×0.24m, discrete distribution and complex structure features 0.08m×0.08m×0.08m.

[0140] Step 4.2.3: Local Continuity Measurement (Adding a Density Smoothness Correction Term): Based on the general region continuity scoring formula, add a density transition smoothness correction term for the splicing region. , To ensure smooth density transition in the splicing area, This represents the maximum density transition smoothness across the entire region, with a value ranging from 0 to 1. A smoother density transition indicates a more consistent and smoother transition. The closer it is to 1, the more precise the correction formula becomes:

[0141]

[0142] in For the i-th point within the splicing region The corrected continuity score is the general region continuity score. The splicing area version, , , Consistent with general areas, this ensures that the continuity score better reflects the actual characteristics of the land features in the spliced ​​area.

[0143] Step 4.2.4: Sampling density weight allocation (minority class weight tilt): Based on the general area weight allocation formula, a weight tilt coefficient is added to the minority land cover in the stitched area. (The skewness coefficient is 1.2, which both increases the sampling weight of the minority class to prevent minority class features from being overwhelmed due to density fluctuations, and avoids excessive weighting leading to an imbalance in sample distribution.) The corrected formula is:

[0144]

[0145] in This is the dedicated sampling weight value for the i-th point within the splicing region, which is the sampling weight for the general region. The version with corrected splicing areas. , , , , This is consistent with the general area. Let be the continuity score power term for the i-th point within the splicing region, where To fix the adaptive sensitivity parameter.

[0146] Step 4.2.5: Weighted Farthest Point Sampling (Sampling Rate Improvement + Density Uniformity Constraint): The weighted distance formula is the same as that for general areas, with dual adjustments to adapt to the stitching area: a) The sampling rate is increased by 5% compared to general areas, configured as 25%-35% for continuous features and 55%-65% for discrete features (the purpose of the improvement is to compensate for the slight feature loss during the secondary smoothing process and ensure that the features of the stitching area are fully preserved); b) A density uniformity constraint is added to ensure that the percentage of points in any grid in the stitching area fluctuates by ≤10% during the sampling process, maintaining the density consistency of the stitching area.

[0147] Step 4.2.6: Sampling Effect Verification (Threshold Relaxation): Verification is performed using the mean absolute error (MAE) of curvature consistency. Considering the enhanced smoothing and high sampling rate characteristics of the stitched area, the verification threshold is appropriately relaxed (by 0.01 to match the processing results of the stitched area and avoid over-verification leading to duplicate sampling). The judgment threshold is: continuously distributed, structurally simple ground features. Discretely distributed and structurally complex land features If the threshold is not met, return to step 1 for resampling.

[0148] Step 4.3: Region fusion after resampling:

[0149] This will complete the regional adaptive downsampling and smoothing of the general region. and splicing area Perform point cloud fusion to form a resampled full point cloud space. Add region type labels to each point cloud feature ( Marked as 0 Labeled as 1), scale encoding label, category proportion label and density feature label provide labeled structured feature input for subsequent AI model training, ensuring that the model can accurately identify point cloud features of different regions and scales.

[0150] like Figures 2-3 The image shows a comparison of local viewpoints between normal downsampling and adaptive downsampling. It can be seen that the overall and local viewpoint accuracy is higher under the adaptive downsampling of this invention.

[0151] Table 1 Comparison of feature encoding module embedding effects between existing methods and the method of the present invention:

[0152]

[0153] This invention:

[0154] 1) By designing a multi-scale adaptive resampling mechanism, the parameters were dynamically adjusted to reduce the data size and computational cost, while preserving key features of ground features;

[0155] 2) An intelligent dynamic category balancing system was constructed, which combined data features to optimize the classification logic and adopted targeted strategies to alleviate the long-tail distribution of categories, thereby improving sample quality;

[0156] 3) The core of the design is to address the differences in features between the stitched area of ​​multi-source point cloud and the general area. A regional quantitative analysis and differentiated resampling strategy is designed to alleviate the problems of density imbalance and feature distortion in the stitched area.

[0157] In summary, this invention improves the quality of training samples and model adaptability, enhances semantic segmentation accuracy and generalization ability, provides core technical support for fields such as autonomous driving and smart cities, and meets the high-standard requirements of AI training data.

[0158] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A method for adaptive optimization and feature matching of multi-source point clouds, characterized in that, Includes the following steps: Step 1: Collect vehicle-mounted and motorcycle-mounted laser point cloud data, perform basic merging of multi-source point clouds, define the full point cloud sample space, and divide the spliced ​​region and general region; Step 2: Based on the basic merged full point cloud sample space, perform multi-dimensional prior analysis on the labeled point cloud dataset according to the general region and the spliced ​​region, and generate a summary table of multi-dimensional analysis of the full region; Step 3: Based on the multi-dimensional analysis results, perform multi-scale adaptive classification of land cover categories, and output scale codes and continuous / discrete land cover classification tables; Step 4: To address class imbalance, perform differentiated adaptive resampling on the spliced ​​region and the general region respectively, and generate labeled structured point cloud samples for downstream AI model training.

2. The multi-source point cloud adaptive optimization and feature adaptation method according to claim 1, characterized in that, Step 1 specifically includes: Step 1.1: Merge the vehicle-mounted and motorcycle-mounted laser point clouds in the same coordinate system and LAS format; Step 1.2: Extract the acquisition trajectory coordinates and acquisition device parameters from the vehicle / motorcycle point cloud acquisition trajectory file: Step 1.3: Obtain basic information using a single acquisition trajectory from a vehicle / motorcycle: calculate the effective laser point cloud coverage width on the road surface as the basic value for the buffer radius; Step 1.4: For each data collection trajectory Perform segmented local density calculations; Step 1.5: Utilize multiple acquisition trajectories from vehicle-mounted and motorcycle-mounted equipment to calculate the overlapping intervals between trajectories and adaptively determine the buffer radius; Step 1.6: The intersection area of ​​all vehicle / motorcycle local adaptive buffer zones is the stitching area, and the remaining non-intersection area is defined as the general area; Step 1.7: For locations where the buffer radius of adjacent trajectory segments changes abruptly, linear interpolation is used for transition.

3. The multi-source point cloud adaptive optimization and feature adaptation method according to claim 1, characterized in that, Step 2 specifically includes: Step 2.1: Multi-dimensional basic analysis of general areas: For general areas with no density abrupt changes and relatively uniform point cloud distribution, basic quantitative analysis is carried out from three dimensions: spatial distribution, proportion of category points, and geometric morphology, to calculate core indicators and determine characteristic attributes; Step 2.2: Multi-dimensional adaptation analysis of the splicing area: For stitched areas where local density fluctuations and slight distortions in geometric features are prone to occur after multi-source point cloud stitching, density uniformity supplementary analysis is added to the framework of general area three-dimensional basic analysis, and the core index formulas and judgment thresholds are adapted and modified. Step 2.3: Integration of all regional indicators: The quantitative analysis results of the general region and the stitched region are merged to generate a multi-dimensional analysis summary table of the full point cloud region, which marks the feature differences, specific indicators and correction parameters between the general region and the stitched region.

4. The multi-source point cloud adaptive optimization and feature adaptation method according to claim 1, characterized in that, Step 3 specifically includes: Based on the analysis results of step 2, the spatial distribution continuity criterion, instance point balance criterion, and geometric morphology complexity criterion are used to divide the general area and the spliced ​​area into land feature scales, respectively, and output scale codes and continuous or discrete land feature classification tables; among them, the judgment threshold of the spliced ​​area is adapted and corrected compared with that of the general area.

5. The multi-source point cloud adaptive optimization and feature adaptation method according to claim 1, characterized in that, Step 4 specifically includes: implementing a full-process differentiation strategy based on adaptive downsampling of the general region and adaptive downsampling of the stitched region, as follows: Step 4.1: Adaptive downsampling and smoothing in general regions: For general areas with uniform point cloud distribution and no abrupt density changes, a basic smoothing + conventional adaptive downsampling strategy is implemented. The parameter configuration fits the features of general areas without additional correction. Step 4.2: Adaptive downsampling and smoothing of the stitching region: For splicing regions that still exhibit local density fluctuations and slight geometric feature distortions after multi-source splicing, an enhanced smoothing + adaptive downsampling strategy is implemented. Based on the general region process, core parameters are modified, density constraints are increased, and verification thresholds are relaxed. Step 4.3: Region fusion after resampling: The point cloud is fused with the completed regional adaptive downsampling and smoothed general and spliced ​​regions to form the full point cloud space after resampling. Region type label, scale encoding label, category proportion label and density feature label are added to each point cloud feature.

6. The multi-source point cloud adaptive optimization and feature adaptation method according to claim 5, characterized in that, Step 4.1 includes the following specific steps: Step 4.1.1: Point cloud smoothing and denoising: The original point cloud position is corrected using the moving least squares method; Step 4.1.2: Local region segmentation: The point cloud is segmented and smoothed using an unbalanced octree, and the voxel size is fixed based on the general regional spatial distribution analysis results; Step 4.1.3: Local Continuity Measurement: For each point, calculate the continuity score by combining the scale code of its associated land cover with the normal consistency variance; Step 4.1.4: Sampling density weight allocation: The sampling weights are dynamically adjusted based on the continuity score, the proportion of class points, and the variance of normal consistency. Step 4.1.5: Weighted farthest point sampling: Calculate the weighted distance between points, iteratively select the point with the largest weighted distance, until the preset number is reached; Step 4.1.6: Sampling effect verification: The sampling effect is verified using the mean absolute error of curvature consistency. The threshold for determining the mean absolute error of continuously distributed, structurally simple ground features is... Mean absolute error of discretely distributed and structurally complex ground features If the condition is not met, return to step 1 for resampling to ensure that the geometric features of the point cloud are not significantly distorted after sampling.

7. The multi-source point cloud adaptive optimization and feature adaptation method according to claim 5, characterized in that, Step 4.2 includes the following specific steps: Step 4.2.1: Employ a quadratic MLS smoothing strategy, with the weighting function consistent with that of the general region, combined with the density variation coefficient of the spliced ​​region. Dynamically adjust the radius of influence Simultaneously, density constraints are introduced to ensure the coefficient of variation of mesh density in the splicing region during the smoothing process. ; Step 4.2.2: Use an unbalanced octree for segmentation, reducing the voxel size by 20% compared to the general region; Step 4.2.3: Based on the general region continuity scoring formula, add a correction term for the density transition smoothness of the splicing region for correction; Step 4.2.4: Based on the general regional weight allocation formula, a weight tilt coefficient is added to the minority land cover types in the splicing area for correction; Step 4.2.5: Weighted farthest point sampling: The weighted distance formula is the same as that for general areas, and double adjustments are made to adapt the stitched area; Step 4.2.6: Verify the sampling effect using the mean absolute error of curvature consistency, and appropriately relax the verification threshold. The judgment threshold is: the mean absolute error of continuously distributed and structurally simple ground features. Mean absolute error of discretely distributed and structurally complex ground features If the threshold is not met, return to step 1 for resampling.

8. The multi-source point cloud adaptive optimization and feature adaptation method according to claim 7, characterized in that, The dual adjustment in step 4.2.5 includes the following steps: a) Configure the sampling rate of the stitched area to be 5% higher than that of the general area, specifically: the sampling rate range for continuously distributed features is 25%-35%, and the sampling rate range for discretely distributed features is 55%-65%; b) Introduce density uniformity constraints to ensure that the percentage of points in any local grid within the stitched area does not fluctuate by more than 10% during the weighted farthest point sampling process.

Images