Point cloud data processing method and device of laser radar, equipment and medium

Abstract

The invention provides a point cloud data processing method and device of a laser radar, equipment and a medium, and the method comprises the steps: carrying out the adaptive local neighborhood construction of each query point cloud, and determining a local neighborhood point set of the query point cloud; determining a plurality of geometric features of each query point cloud based on the local neighborhood point set; performing normalization processing and dynamic weight optimization on the plurality of geometric features of each query point cloud, constructing a comprehensive feature of each query point cloud, dividing the plurality of original point cloud data based on the plurality of comprehensive features, and determining a plurality of point cloud categories; and carrying out noise removal processing on the original point cloud data in each point cloud category, determining updated point cloud data, repeatedly carrying out local neighborhood point set determination and point cloud division processing on the updated point cloud data until a convergence condition is met, and outputting the denoised point cloud data. Effectiveness and anti-interference capability of feature fusion are improved, and intelligent identification and denoising of the point cloud and accurate retention of the effective point cloud are realized.

Description

Technical Field

[0001] This application relates to the field of point cloud data processing technology, and in particular to a point cloud data processing method, apparatus, device and medium for lidar. Background Technology

[0002] LiDAR (Light Detection and Ranging) actively emits laser beams and receives reflected signals from the target surface, enabling efficient acquisition of high-density, high-precision three-dimensional spatial coordinate data of objects or features in a scene, forming a point cloud dataset. However, in actual data acquisition, due to limitations in sensor accuracy, measurement distance, target material reflectivity, and complex environmental factors (such as rain, fog, dust, multipath reflection, motion blur, etc.), a large number of noise points inevitably get mixed into the raw point cloud data.

[0003] Traditional point cloud denoising methods mainly include filtering algorithms based on statistical characteristics and threshold discrimination methods based on geometric rules. For example, Statistical Outlier Removal (SOR) identifies and removes outliers that deviate from the main distribution by calculating the average distance and standard deviation between each point and its neighbors. In addition, there are auxiliary methods such as voxel grid downsampling and projection plane fitting for coarse denoising. While the aforementioned methods are simple to implement and computationally efficient, they exhibit significant limitations when dealing with large-scale, non-uniformly distributed, and multi-scale structured real-world point clouds. First, traditional methods typically rely on fixed neighborhood construction strategies (such as k-nearest neighbors or fixed-radius sphere search), making it difficult to adapt to local feature variations in regions of different densities. This can lead to the accidental deletion of valid points in sparse areas and the missed detection of noise in dense areas. Second, relying solely on single geometric or statistical features (such as distance and density) for judgment lacks a comprehensive understanding of the local morphology of point clouds (such as planar, curved, and linear structures), resulting in poor performance in preserving fine structures (such as railings, wires, and edges). Third, most methods rely on manual experience for parameter settings, lacking adaptive adjustment mechanisms, resulting in poor versatility and robustness. In conclusion, there is an urgent need for an efficient method for denoising point cloud data to address the technical challenge of preserving critical structures while dealing with severe noise interference in complex real-world scenarios. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a point cloud data processing method, device, equipment and medium for lidar, which improves the effectiveness and anti-interference ability of feature fusion by the collaborative analysis of multi-dimensional geometric features and the introduction of a dynamic weight optimization mechanism, and realizes intelligent recognition and denoising of point clouds and accurate preservation of effective point clouds.

[0005] This application provides a point cloud data processing method for lidar, the point cloud data processing method including: For each query point cloud in the original point cloud data, an adaptive local neighborhood is constructed to determine the local neighborhood point set of the query point cloud; Based on the local neighborhood point set of each query point cloud, multiple geometric features of each query point cloud are determined; wherein, the geometric features include at least point density features, normal vector consistency features, local plane fitting residual features, curvature change rate features, principal component orientation deviation features, sphericity features, flatness features, and straightness features. Normalization and dynamic weight optimization are performed on multiple geometric features of each query point cloud to construct a comprehensive feature of each query point cloud. Based on the comprehensive feature of multiple query point clouds, multiple original point cloud data are divided to determine multiple point cloud categories. Adaptive noise removal is performed on the original point cloud data in each point cloud category to determine the updated point cloud data. The local neighborhood point set determination and point cloud partitioning are repeated on the updated point cloud data until the convergence condition is met, and then the denoised point cloud data is output.

[0006] In one possible implementation, the adaptive local neighborhood construction for each query point cloud in the original point cloud data, to determine the local neighborhood point set of the query point cloud, includes: The spatial indexing method is dynamically selected based on the local point cloud density of the current queried point cloud region. When the local point cloud density is higher than the first threshold, a neighborhood search is performed using a voxel index. When the local point cloud density is lower than the first threshold or is located in the edge region, a kd-tree index is used for neighborhood search. During the neighborhood search, the number of neighborhood points is adjusted according to the local point cloud density to determine the local neighborhood point set of the query point cloud.

[0007] In one possible implementation, determining multiple geometric features of each query point cloud based on the local neighborhood point set of each query point cloud includes: A covariance matrix is ​​constructed based on a local neighborhood point set. After performing singular value decomposition on the covariance matrix, a first eigenvalue, a second eigenvalue, and a third eigenvalue are obtained. The first eigenvalue, the second eigenvalue, and the third eigenvalue decrease sequentially. Based on the first feature value, the second feature value, and the third feature value, the curvature change rate feature, principal component direction deviation feature, sphericity feature, flatness feature, and straightness feature are determined.

[0008] In one possible implementation, determining the curvature change rate feature, principal component orientation deviation feature, sphericity feature, flatness feature, and straightness feature based on the first feature value, the second feature value, and the third feature value includes: The sum of the first feature value, the second feature value, and the third feature value is calculated, and the curvature change rate feature is determined based on the ratio between the third feature value and the sum value. Based on the ratio between the first feature value and the sum value, the principal component direction deviation feature is determined; The sphericity feature is determined based on the ratio between the third feature value and the first feature value; The difference between the second feature value and the third feature value is calculated, and the flatness feature is determined based on the ratio between the difference and the first feature value. The difference between the first feature value and the second feature value is calculated, and the straightness feature is determined based on the ratio between the difference and the first feature value.

[0009] In one possible implementation, the normalization and dynamic weight optimization of multiple geometric features of each query point cloud to construct a comprehensive feature for each query point cloud includes: Normalize multiple geometric features of each query point cloud to determine multiple normalized geometric features of each query point cloud. The initial weights of multiple normalized geometric features are dynamically optimized based on the objective function to determine the optimal weight vector. The comprehensive feature is generated by weighting and fusing multiple normalized geometric features based on the optimal weight vector.

[0010] In one possible implementation, the step of dynamically optimizing the initial weights of multiple normalized geometric features based on an optimization objective function to determine the optimal weight vector includes: The discriminative ability of multiple normalized geometric features on a point cloud dataset is quantitatively evaluated to obtain a representation ability score for each normalized geometric feature; wherein the representation ability score is calculated based on at least one of the feature's information gain, variance response, or attention weight. An initial feature weight allocation scheme is generated based on the representation ability score. Based on the initial feature weight allocation scheme, the normalized geometric features of the query point cloud are weighted and fused to determine the initial comprehensive features. One-dimensional clustering is performed using multiple initial integrated features to determine the clustering results; Based on the clustering results, an optimization objective function is constructed, and an optimal weight vector that maximizes the optimization objective function is searched in the probabilistic simplex space using a gradient-free optimization algorithm.

[0011] In one possible implementation, the optimization objective function is:

[0012] in, To optimize the objective function, α and β are balance coefficients. For the first Clustered point cloud dataset of classes, This represents the number of clusters after clustering. For querying point clouds The initial comprehensive characteristics, For the first The category center of the class.

[0013] This application embodiment also provides a point cloud data processing device for lidar, the point cloud data processing device comprising: The neighborhood construction module is used to adaptively construct a local neighborhood for each query point cloud in the original point cloud data, and determine the local neighborhood point set of the query point cloud. The geometric feature determination module is used to determine multiple geometric features of each query point cloud based on the local neighborhood point set of each query point cloud; wherein, the geometric features include at least point density features, normal vector consistency features, local plane fitting residual features, curvature change rate features, principal component direction deviation features, sphericity features, flatness features, and straightness features. The point cloud segmentation module is used to normalize and dynamically optimize the multiple geometric features of each query point cloud to construct the comprehensive features of each query point cloud. Based on the comprehensive features of multiple query point clouds, the module divides multiple original point cloud data to determine multiple point cloud categories. The noise removal module is used to perform adaptive noise removal processing on the original point cloud data in each point cloud category, determine the updated point cloud data, and repeatedly perform local neighborhood point set determination and point cloud partitioning processing on the updated point cloud data until the convergence condition is met, and then output the denoised point cloud data.

[0014] This application embodiment also provides an electronic device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, the steps of the point cloud data processing method for LiDAR described above are performed.

[0015] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the point cloud data processing method for lidar described above.

[0016] This application provides a point cloud data processing method, apparatus, device, and medium for lidar. The point cloud data processing method includes: adaptively constructing a local neighborhood for each query point cloud in the original point cloud data to determine the local neighborhood point set of the query point cloud; determining multiple geometric features of each query point cloud based on the local neighborhood point set of each query point cloud; wherein the geometric features include at least point density features, normal vector consistency features, local plane fitting residual features, curvature change rate features, principal component direction deviation features, sphericity features, flatness features, and straightness features; normalizing and dynamically weighting the multiple geometric features of each query point cloud to construct a comprehensive feature of each query point cloud; dividing the multiple original point cloud data based on the comprehensive features of the multiple query point clouds to determine multiple point cloud categories; adaptively removing noise from the original point cloud data in each point cloud category to determine updated point cloud data; repeatedly determining the local neighborhood point set and dividing the point cloud data for the updated point cloud data until the convergence condition is met, and outputting the denoised point cloud data. By conducting collaborative analysis of multi-dimensional geometric features and introducing a dynamic weight optimization mechanism, the effectiveness and anti-interference ability of feature fusion are improved, enabling intelligent recognition and denoising of point clouds and accurate preservation of effective point clouds.

[0017] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating a point cloud data processing method for a lidar system provided in this application embodiment; Figure 2 A schematic diagram illustrating a point cloud data processing method for a lidar provided in an embodiment of this application; Figure 3 This is a schematic diagram of the point cloud data processing device for the lidar provided in the embodiments of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. Based on the embodiments of this application, every other embodiment obtained by those skilled in the art without inventive effort falls within the scope of protection of this application.

[0021] First, the applicable scenarios for this application will be introduced. This application can be applied to the field of point cloud data processing technology.

[0022] Research has revealed that traditional methods typically rely on fixed neighborhood construction strategies (such as k-nearest neighbors or fixed-radius sphere search), which struggle to adapt to changes in local features across regions of varying density. This can lead to issues like incorrect deletion of valid points in sparse areas and missed noise detection in dense areas. Secondly, relying solely on single geometric or statistical features (such as distance or density) for judgment lacks a comprehensive understanding of the local morphology of point clouds (such as planar, curved, or linear structures), resulting in poor performance in preserving fine structures (such as railings, wires, and edges). Thirdly, most methods rely on manual experience for parameter settings, lacking adaptive adjustment mechanisms and exhibiting poor versatility and robustness. In conclusion, there is an urgent need for an efficient method for denoising point cloud data to address the technical challenge of preserving critical structures while dealing with severe noise interference in complex real-world scenarios.

[0023] Based on this, this application provides a point cloud data processing method for LiDAR. By conducting collaborative analysis of multi-dimensional geometric features and introducing a dynamic weight optimization mechanism, the effectiveness of feature fusion and anti-interference capability are improved, enabling intelligent identification and denoising of point clouds and accurate preservation of effective point clouds.

[0024] Please see Figure 1 , Figure 1 This is a flowchart illustrating a point cloud data processing method for a lidar system provided in an embodiment of this application. Figure 1 As shown in the embodiments of this application, the point cloud data processing method includes: S101: Perform adaptive local neighborhood construction on each query point cloud in the original point cloud data to determine the local neighborhood point set of the query point cloud.

[0025] In this step, an adaptive local neighborhood is constructed for each query point cloud in the original point cloud data to determine the local neighborhood point set of the query point cloud.

[0026] It should be noted that the local neighborhood point set is composed of multiple neighboring point cloud data of the queried point cloud.

[0027] It should be noted that we assume we currently have a point cloud representing a point on the railing. We use this point cloud data as a "query point cloud" and search for the k nearest points around it, forming a small region.

[0028] In one possible implementation, the adaptive local neighborhood construction for each query point cloud in the original point cloud data, to determine the local neighborhood point set of the query point cloud, includes: A: Dynamically select the spatial indexing method based on the local point cloud density of the area where the current queried point cloud is located.

[0029] Here, the local point cloud density of the current query point cloud region is determined, and the spatial indexing method is dynamically selected based on the local point cloud density. It should be noted that the local point cloud density can be determined in the following ways: take the query point cloud as the center, set an initial search radius, and count the number of neighboring points within the radius; or obtain the average number of points in the grid based on a pre-built spatial partitioning structure (such as a coarse-grained voxel grid) as a density reference value.

[0030] B: When the local point cloud density is higher than the first threshold, a neighborhood search is performed using a voxel index.

[0031] Here, when the local point cloud density exceeds the first threshold, it indicates that the point distribution in that area is dense and the data redundancy is high. In this case, voxel grid indexing is used for neighborhood search. Voxel grid indexing divides the 3D space according to rules and assigns the point cloud to the corresponding voxel units. It can quickly locate neighboring points in high-density scenes and effectively suppress the computational noise caused by local oversampling, thereby improving search efficiency and stability.

[0032] C: When the local point cloud density is lower than the first threshold or is located in the edge region, a kd-tree index is used for neighborhood search. During the neighborhood search, the number of neighborhood points is adjusted according to the local point cloud density to determine the local neighborhood point set of the query point cloud.

[0033] Here, when the local point cloud density is below the first threshold or the queried point cloud is located on the edge, outline, or sparse transition region of the target object, the system switches to k-dimensional tree indexing for neighborhood search. k-trees can efficiently support nearest neighbor lookups for irregularly distributed point clouds, accurately capturing a limited but crucial number of neighboring points in low-density or structurally complex regions, avoiding the problem of missing neighborhoods due to voxel holes.

[0034] In addition to selecting the index structure, this application introduces a dynamic adjustment mechanism for the number of neighborhood points: the required number of neighborhood points is adaptively determined based on the current local density level. For example, a larger number of neighborhood points is set in high-density regions to fully represent the local planar or curved surface structure; while a smaller number of neighborhood points is used in low-density regions to prevent the introduction of distant irrelevant points that could cause feature distortion. Finally, combining the selected index structure and the dynamic number of neighborhood points, a neighborhood search operation is performed to obtain a set of nearest neighbors around the query point, constituting the local neighborhood point set of the query point, which is used for subsequent calculation and analysis of multidimensional geometric features.

[0035] S102: Determine multiple geometric features of each query point cloud based on the local neighborhood point set of each query point cloud; wherein the geometric features include at least point density features, normal vector consistency features, local plane fitting residual features, curvature change rate features, principal component direction deviation features, sphericity features, flatness features, and straightness features.

[0036] In this step, multiple geometric features of each query point cloud are determined based on the local neighborhood point set of each query point cloud.

[0037] The geometric feature calculation includes various geometric features such as point density, normal vector consistency, local plane fitting residual, curvature change rate, deviation from the principal component direction, sphericity, flatness, and straightness. The purpose is to classify point clouds into different morphological point cloud regions. For example, density can be used to divide sparse and dense point clouds, flatness can be used to divide planar point clouds, and sphericity can be used to divide spherical point clouds.

[0038] In one possible implementation, determining multiple geometric features of each query point cloud based on the local neighborhood point set of each query point cloud includes: a: Construct a covariance matrix based on a local neighborhood point set, and obtain a first eigenvalue, a second eigenvalue, and a third eigenvalue by performing singular value decomposition on the covariance matrix; wherein the first eigenvalue, the second eigenvalue, and the third eigenvalue decrease sequentially.

[0039] Here, the formula for calculating covariance is:

[0040] in, Let covariance be the local neighborhood point set. , , The coordinates of the neighborhood point cloud within a local neighborhood point set. , , Let N be the mean point of the neighborhood, and N be the number of point clouds in the local neighborhood point set. The first eigenvalue of the covariance matrix can be obtained using SVD decomposition. Second eigenvalue and the third eigenvalue .

[0041] b: Based on the first feature value, the second feature value, and the third feature value, the curvature change rate feature, principal component direction deviation feature, sphericity feature, flatness feature, and straightness feature are determined.

[0042] Here, based on the first eigenvalue, the second eigenvalue, and the third eigenvalue, the curvature change rate characteristic, the principal component direction deviation characteristic, the sphericity characteristic, the flatness characteristic, and the straightness characteristic are determined.

[0043] In one possible implementation, determining the curvature change rate feature, principal component orientation deviation feature, sphericity feature, flatness feature, and straightness feature based on the first feature value, the second feature value, and the third feature value includes: The sum of the first feature value, the second feature value, and the third feature value is calculated. Based on the ratio between the third feature value and the sum, the curvature change rate feature is determined. Based on the ratio between the first feature value and the sum, the principal component direction deviation feature is determined. Based on the ratio between the third feature value and the first feature value, the sphericity feature is determined. The difference between the second feature value and the third feature value is calculated. Based on the ratio between the difference and the first feature value, the flatness feature is determined. The difference between the first feature value and the second feature value is calculated. Based on the ratio between the difference and the first feature value, the straightness feature is determined.

[0044] Here, the curvature change rate characteristic, principal component direction deviation characteristic, sphericity characteristic, flatness characteristic, and straightness characteristic are determined by the following formulas: Curvature change rate characteristic = ; Principal component orientation deviation characteristic = ; Sphericity characteristics = ; Flatness feature = ; Straightness characteristics = ; This application comprehensively extracts multiple geometric features, including point density, normal vector consistency, local plane fitting residuals, curvature change rate, principal component orientation deviation, sphericity, flatness, and straightness. A comprehensive feature vector is constructed through a normalization and dynamic weighted fusion strategy, achieving refined classification of local point cloud morphology (such as planes, curved surfaces, linear structures, isolated points, etc.). Compared to traditional methods relying on only a single feature, this scheme can more accurately distinguish between noise points and effective structural points, significantly improving the targeting and reliability of denoising.

[0045] S103: Normalize and dynamically optimize the multiple geometric features of each query point cloud to construct a comprehensive feature for each query point cloud. Based on the comprehensive feature of multiple query point clouds, divide the multiple original point cloud data to determine multiple point cloud categories.

[0046] In this step, multiple geometric features of each query point cloud are normalized and dynamically weighted to construct a comprehensive feature for each query point cloud. Based on the comprehensive feature of multiple query point clouds, multiple original point cloud data are divided to determine multiple point cloud categories.

[0047] It should be noted that point cloud categories include at least high-density planar regions, low-density linear structures, isolated outliers, simple planar regions, and complex curved surface regions.

[0048] In one possible implementation, the normalization and dynamic weight optimization of multiple geometric features of each query point cloud to construct a comprehensive feature for each query point cloud includes: (1): Normalize multiple geometric features of each query point cloud to determine multiple normalized geometric features of each query point cloud.

[0049] Here, the interquartile range (IQR) truncation normalization method is used to normalize multiple geometric features of each query point cloud, thereby determining multiple normalized geometric features of each query point cloud.

[0050] (2): Based on the optimization objective function, the initial weights of multiple normalized geometric features are dynamically optimized to determine the optimal weight vector.

[0051] Here, the initial weights of multiple normalized geometric features are dynamically optimized using an optimization objective function to determine the optimal weight vector.

[0052] In one possible implementation, the step of dynamically optimizing the initial weights of multiple normalized geometric features based on an optimization objective function to determine the optimal weight vector includes: I: Quantitatively evaluate the discriminative ability of multiple normalized geometric features on a point cloud dataset to obtain a representational ability score for each normalized geometric feature; wherein the representational ability score is calculated based on at least one of the feature's information gain, variance response, or attention weight.

[0053] Here, the discriminative ability of each normalized geometric feature is evaluated based on the information gain, variance response, and / or attention weight of the feature, generating a representation ability score.

[0054] Information gain measures the amount of information a feature provides when dividing samples into "noise points" and "effective structure points." Differential responsivity reflects the sensitivity of a feature to numerical fluctuations across different structural regions, assessing its responsiveness to changes in local morphology. For each typical structure type (e.g., planes, edges, lines, isolated points), the variance of the feature within each category is calculated, and the inter-class dispersion is also determined. An MLP-based feature attention network automatically evaluates the importance of each feature. This network takes a normalized feature vector as input and outputs attention weights for each dimension. The above metrics are weighted and fused to obtain a comprehensive feature representation score.

[0055] II: Generate an initial feature weight allocation scheme based on the representation ability score, and perform weighted fusion of the normalized geometric features of the query point cloud based on the initial feature weight allocation scheme to determine the initial comprehensive features.

[0056] Here, an initial feature weight allocation scheme is generated based on the representation ability score. Based on the initial feature weight allocation scheme, the normalized geometric features of the query point cloud are weighted and fused to determine the initial comprehensive features.

[0057] It should be noted that the representational ability scores of each feature are normalized to form a weight set in the form of a probability distribution. This weight allocation principle reflects the design idea of ​​"giving higher weight to features with high discriminative power", ensuring that features with strong ability to distinguish between noise and effective structure (such as flatness in planar regions and straightness in linear structures) can play a dominant role in the subsequent calculation of comprehensive feature values.

[0058] III: Perform one-dimensional clustering using multiple initial integrated features to determine the clustering results.

[0059] It should be noted that clustering results refer to dividing the original point cloud data into several sets (i.e., categories or clusters) of points with similar local geometric characteristics, resulting in category classification { , ,..., }

[0060] IV: Construct an optimization objective function based on the clustering results, and use a gradient-free optimization algorithm in the probabilistic simplex space to search for the optimal weight vector that maximizes the optimization objective function.

[0061] Here, an optimization objective function is constructed based on the clustering results, and an optimal weight vector that maximizes the optimization objective function is searched in the probabilistic simplex space using a gradient-free optimization algorithm.

[0062] The optimization objective function is:

[0063] in, To optimize the objective function, α and β are balance coefficients. For the first Clustered point cloud dataset of classes, This represents the number of clusters after clustering. For querying point clouds The initial comprehensive characteristics, For the first The category center of the class.

[0064] Here, the category center is determined using the following formula:

[0065] (3): Based on the optimal weight vector, multiple normalized geometric features are weighted and fused to generate the comprehensive feature.

[0066] Here, the optimal weight vector is used to weight and fuse multiple normalized geometric features to generate a comprehensive feature.

[0067] It should be noted that the optimal weight vector refers to the set of optimal weight coefficients obtained in the process of weighted fusion of multidimensional normalized geometric features to make the classification result most discriminative, where each weight value corresponds to a normalized geometric feature. In this application, to address the issue of uneven point cloud density, an adaptive neighborhood search mechanism is employed. This mechanism dynamically selects between kd-trees, octrees, or voxel indexes based on different region densities, and combines this with a self-adjusting strategy for the number of neighboring points to avoid insufficient or redundant neighbors, thus ensuring the stability of feature computation. Simultaneously, a dynamic weight optimization mechanism based on clustering results is introduced, utilizing gradient-free optimization to solve for the optimal feature weight combination, further enhancing the effectiveness and anti-interference capability of feature fusion. S104: Perform adaptive noise removal processing on the original point cloud data in each point cloud category to determine the updated point cloud data. Repeat the local neighborhood point set determination and point cloud partitioning processing on the updated point cloud data until the convergence condition is met, and then output the denoised point cloud data.

[0068] In this step, adaptive noise removal processing is performed on the original point cloud data in each point cloud category to determine the updated point cloud data. The above steps are repeated on the updated point cloud data until the convergence condition is met, and then the denoised point cloud data is output.

[0069] It should be noted that the convergence condition is to calculate the mean value of features for each type of point cloud after each iteration and calculate its standard deviation. If the standard deviation is less than the threshold, the processing stops and the denoised point cloud data is output.

[0070] Here, based on the structural characteristics and noise sensitivity of different categories, the denoising algorithm and its parameter configuration are dynamically selected: For point sets identified as high-density planar regions, the Statistical Outlier Removal (SOR) method is adopted, and a relatively strict statistical threshold (such as the average distance multiple) is set to efficiently remove small noise points that deviate from the main distribution; For low-density linear structures or small edge regions, Radius Outlier Removal (ROR) is preferred, or the sensitivity of SOR is reduced to avoid the false deletion of valid points due to sparse neighborhoods; At the same time, topological connectivity constraints can be introduced to ensure that the continuity of linear structures is not destroyed; For points identified as isolated points or low-neighborhood density floating points, they are directly marked as noise and removed; For complex curved surface regions, a flexible threshold is set by combining the local curvature change rate and the plane fitting residual to filter out obvious outliers while preserving surface details.

[0071] In this application, a closed-loop iterative denoising process is designed. After each iteration, features are recalculated, classifications are updated, and denoising intensity is adjusted to gradually approach the optimal state. By introducing a fusion convergence criterion—including a joint judgment mechanism of multiple indicators such as the standard deviation of point cloud feature values ​​of each category, structural similarity, and geometric detail retention rate—premature termination or over-denoising is effectively avoided. At the same time, the denoising intensity adaptively decays with the iteration process, enhancing the stability of the algorithm.

[0072] For further details, please refer to Figure 2 , Figure 2 This is a schematic diagram illustrating a point cloud data processing method for a lidar system provided in an embodiment of this application. Figure 2 As shown, adaptive local neighborhood construction is performed on each query point cloud in the original point cloud data to determine the local neighborhood point set of the query point cloud and determine multiple geometric features of each query point cloud; feature normalization processing, weighted fusion, and point cloud classification processing are performed; adaptive noise removal processing is performed on the original point cloud data in each point cloud category; it is checked whether the convergence condition is met; if yes, the denoised point cloud data is output; otherwise, the above steps are repeated.

[0073] This application provides a point cloud data processing method for lidar, comprising: adaptively constructing a local neighborhood for each query point cloud in the original point cloud data to determine a local neighborhood point set for the query point cloud; determining multiple geometric features for each query point cloud based on the local neighborhood point set; wherein the geometric features include at least point density features, normal vector consistency features, local plane fitting residual features, curvature change rate features, principal component direction deviation features, sphericity features, flatness features, and straightness features; normalizing and dynamically weighting the multiple geometric features of each query point cloud to construct a comprehensive feature for each query point cloud; dividing the original point cloud data into multiple point cloud categories based on the comprehensive features of the multiple query point clouds; adaptively removing noise from the original point cloud data in each point cloud category to determine updated point cloud data; repeatedly determining the local neighborhood point set and dividing the point cloud data until the convergence condition is met, and outputting the denoised point cloud data. By conducting collaborative analysis of multi-dimensional geometric features and introducing a dynamic weight optimization mechanism, the effectiveness and anti-interference ability of feature fusion are improved, enabling intelligent recognition and denoising of point clouds and accurate preservation of effective point clouds.

[0074] Please see Figure 3 , Figure 3 This is a schematic diagram of the point cloud data processing device for a lidar provided in an embodiment of this application. Figure 3 As shown, the point cloud data processing device 300 of the lidar includes: The neighborhood construction module 310 is used to perform adaptive local neighborhood construction on each query point cloud in the original point cloud data, and determine the local neighborhood point set of the query point cloud; The geometric feature determination module 320 is used to determine multiple geometric features of each query point cloud based on the local neighborhood point set of each query point cloud; wherein, the geometric features include at least point density features, normal vector consistency features, local plane fitting residual features, curvature change rate features, principal component direction deviation features, sphericity features, flatness features, and straightness features. The point cloud segmentation module 330 is used to normalize and dynamically optimize the multiple geometric features of each query point cloud to construct the comprehensive features of each query point cloud, and to segment the multiple original point cloud data based on the comprehensive features of multiple query point clouds to determine multiple point cloud categories. The noise removal module 340 is used to perform adaptive noise removal processing on the original point cloud data in each point cloud category, determine the updated point cloud data, and repeatedly perform local neighborhood point set determination and point cloud partitioning processing on the updated point cloud data until the convergence condition is met, and then output the denoised point cloud data.

[0075] Furthermore, the neighborhood construction module 310 is used to perform adaptive local neighborhood construction on each query point cloud in the original point cloud data, and determine the local neighborhood point set of the query point cloud: The spatial indexing method is dynamically selected based on the local point cloud density of the current queried point cloud region. When the local point cloud density is higher than the first threshold, a neighborhood search is performed using a voxel index. When the local point cloud density is lower than the first threshold or is located in the edge region, a kd-tree index is used for neighborhood search. During the neighborhood search, the number of neighborhood points is adjusted according to the local point cloud density to determine the local neighborhood point set of the query point cloud.

[0076] Furthermore, the geometric feature determination module 320 is used to determine multiple geometric features of each query point cloud based on the local neighborhood point set of each query point cloud: A covariance matrix is ​​constructed based on a local neighborhood point set. After performing singular value decomposition on the covariance matrix, a first eigenvalue, a second eigenvalue, and a third eigenvalue are obtained. The first eigenvalue, the second eigenvalue, and the third eigenvalue decrease sequentially. Based on the first feature value, the second feature value, and the third feature value, the curvature change rate feature, principal component direction deviation feature, sphericity feature, flatness feature, and straightness feature are determined.

[0077] Furthermore, the geometric feature determination module 320 is used to determine the curvature change rate feature, principal component direction deviation feature, sphericity feature, flatness feature, and straightness feature based on the first feature value, the second feature value, and the third feature value. The sum of the first feature value, the second feature value, and the third feature value is calculated, and the curvature change rate feature is determined based on the ratio between the third feature value and the sum value. Based on the ratio between the first feature value and the sum value, the principal component direction deviation feature is determined; The sphericity feature is determined based on the ratio between the third feature value and the first feature value; The difference between the second feature value and the third feature value is calculated, and the flatness feature is determined based on the ratio between the difference and the first feature value. The difference between the first feature value and the second feature value is calculated, and the straightness feature is determined based on the ratio between the difference and the first feature value.

[0078] Furthermore, the point cloud segmentation module 330 is used to normalize and dynamically optimize the multiple geometric features of each query point cloud to construct a comprehensive feature for each query point cloud: Normalize multiple geometric features of each query point cloud to determine multiple normalized geometric features of each query point cloud. The initial weights of multiple normalized geometric features are dynamically optimized based on the objective function to determine the optimal weight vector. The comprehensive feature is generated by weighting and fusing multiple normalized geometric features based on the optimal weight vector.

[0079] Furthermore, the point cloud segmentation module 330 is used to dynamically optimize the initial weights of multiple normalized geometric features based on the optimization objective function, and determine the optimal weight vector: The discriminative ability of multiple normalized geometric features on a point cloud dataset is quantitatively evaluated to obtain a representation ability score for each normalized geometric feature; wherein the representation ability score is calculated based on at least one of the feature's information gain, variance response, or attention weight. An initial feature weight allocation scheme is generated based on the representation ability score. Based on the initial feature weight allocation scheme, the normalized geometric features of the query point cloud are weighted and fused to determine the initial comprehensive features. One-dimensional clustering is performed using multiple initial integrated features to determine the clustering results; Based on the clustering results, an optimization objective function is constructed, and an optimal weight vector that maximizes the optimization objective function is searched in the probabilistic simplex space using a gradient-free optimization algorithm.

[0080] This application provides a point cloud data processing device for a lidar system. The device includes: a neighborhood construction module for adaptively constructing a local neighborhood for each query point cloud in the original point cloud data, determining a local neighborhood point set for the query point cloud; and a geometric feature determination module for determining multiple geometric features for each query point cloud based on its local neighborhood point set. The geometric features include at least point density features, normal vector consistency features, local plane fitting residual features, curvature change rate features, principal component direction deviation features, sphericity features, and flatness features. The system includes a linearity feature module; a point cloud segmentation module, which normalizes multiple geometric features of each query point cloud and performs dynamic weight optimization to construct a comprehensive feature for each query point cloud. Based on the comprehensive feature of multiple query point clouds, it segments multiple original point cloud data to determine multiple point cloud categories; and a noise removal module, which performs adaptive noise removal processing on the original point cloud data in each point cloud category to determine the updated point cloud data. The updated point cloud data undergoes repeated local neighborhood point set determination and point cloud segmentation processing until convergence is met, at which point the denoised point cloud data is output. Through the collaborative analysis of multi-dimensional geometric features and the introduction of a dynamic weight optimization mechanism, the effectiveness and anti-interference ability of feature fusion are improved, achieving intelligent point cloud recognition and denoising while accurately preserving effective point clouds.

[0081] Please see Figure 4 , Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 4 As shown, the electronic device 400 includes a processor 410, a memory 420, and a bus 430.

[0082] The memory 420 stores machine-readable instructions executable by the processor 410. When the electronic device 400 is running, the processor 410 communicates with the memory 420 via the bus 430. When the machine-readable instructions are executed by the processor 410, they can perform the operations described above. Figure 1 as well as Figure 2 The steps of the point cloud data processing method of the lidar in the method embodiment shown are described in detail in the method embodiment, and will not be repeated here.

[0083] This application also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, can perform the above-described actions. Figure 1 as well as Figure 2 The steps of the point cloud data processing method of the lidar in the method embodiment shown are described in detail in the method embodiment, and will not be repeated here.

[0084] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0085] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the shown or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0086] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0087] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0088] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0089] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The scope of protection of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for processing point cloud data of a laser radar, characterized in that, The point cloud data processing method comprises: Adaptive local neighborhood construction is performed on each query point cloud in the original point cloud data to determine a local neighborhood point set of the query point cloud; A plurality of geometric features of each query point cloud are determined based on the local neighborhood point set of the query point cloud; wherein the geometric features at least include point density features, normal vector consistency features, local plane fitting residual features, curvature change rate features, principal component direction deviation degree features, sphericity features, planeness features and straightness features; Normalization processing and dynamic weight optimization are performed on the plurality of geometric features of each query point cloud to construct a comprehensive feature of each query point cloud, and the original point cloud data is divided based on the comprehensive features of the plurality of query point clouds to determine a plurality of point cloud categories; Adaptive noise removal processing is performed on the original point cloud data in each point cloud category to determine updated point cloud data, and the local neighborhood point set determination and point cloud division processing are repeatedly performed on the updated point cloud data until a convergence condition is met, and the denoised point cloud data is output.

2. The point cloud data processing method of claim 1, wherein, The adaptive local neighborhood construction performed on each query point cloud in the original point cloud data to determine a local neighborhood point set of the query point cloud comprises: A spatial index mode is dynamically selected according to the local point cloud density of the region where the current query point cloud is located; When the local point cloud density is higher than a first threshold, voxel indexing is adopted for neighborhood search; When the local point cloud density is lower than the first threshold or located at an edge region, k-d tree indexing is adopted for neighborhood search, and the number of neighborhood points is adjusted according to the local point cloud density in the neighborhood search process to determine the local neighborhood point set of the query point cloud.

3. The point cloud data processing method of claim 1, wherein, The plurality of geometric features of each query point cloud are determined based on the local neighborhood point set, comprising: A covariance matrix is constructed based on the local neighborhood point set, and first, second and third eigenvalues are obtained after singular value decomposition of the covariance matrix; wherein the first, second and third eigenvalues decrease in turn; Based on the first, second and third eigenvalues, curvature change rate features, principal component direction deviation degree features, sphericity features, planeness features and straightness features are determined.

4. The point cloud data processing method of claim 3, wherein, Based on the first, second and third eigenvalues, curvature change rate features, principal component direction deviation degree features, sphericity features, planeness features and straightness features are determined, comprising: The sum of the first, second and third eigenvalues is calculated, and the curvature change rate features are determined based on the ratio between the third eigenvalue and the sum; The principal component direction deviation degree features are determined based on the ratio between the first eigenvalue and the sum; The sphericity features are determined based on the ratio between the third eigenvalue and the first eigenvalue; The difference between the second and third eigenvalues is calculated, and the planeness features are determined based on the ratio between the difference and the first eigenvalue; A difference between the first feature value and the second feature value is calculated, and the straightness feature is determined based on a ratio between the difference and the first feature value.

5. The method of claim 1, wherein, The normalization processing and the dynamic weight optimization are performed on the multiple geometric features of each query point cloud to construct a comprehensive feature of each query point cloud, and the multiple original point cloud data are divided based on the comprehensive features of the multiple query point clouds to determine multiple point cloud categories. The normalization processing is performed on the multiple geometric features of each query point cloud to determine multiple normalized geometric features of each query point cloud. The initial weight of each normalized geometric feature is dynamically optimized based on an optimization objective function to determine an optimal weight vector. The multiple normalized geometric features are weighted and fused based on the optimal weight vector to generate the comprehensive feature.

6. The point cloud data processing method of claim 5, wherein, The initial weight of each normalized geometric feature is dynamically optimized based on an optimization objective function to determine an optimal weight vector, including: The discrimination ability of each normalized geometric feature on the point cloud dataset is quantitatively evaluated to obtain a representation ability score of each normalized geometric feature, wherein the representation ability score is calculated based on at least one of information gain, variance response, or attention weight; An initial feature weight distribution scheme is generated according to the representation ability score, and the multiple normalized geometric features of each query point cloud are weighted and fused based on the initial feature weight distribution scheme to determine an initial comprehensive feature; One-dimensional clustering is performed using the multiple initial comprehensive features to determine a clustering result; An optimization objective function is constructed according to the clustering result, and a gradient-free optimization algorithm is used in a probability simplex space to search for an optimal weight vector that maximizes the optimization objective function.

7. The point cloud data processing method of claim 6, wherein, The optimization objective function is: wherein, are balance coefficients for optimizing the objective function, is the first class of clustered point cloud data sets, is the number of classes after clustering, is the initial comprehensive feature of the query point cloud is the initial comprehensive feature of the query point cloud is the class center of the first class. 8.A point cloud data processing apparatus of a laser radar, characterized by comprising: The point cloud data processing device includes: A neighborhood construction module is configured to perform adaptive local neighborhood construction on each query point cloud in the original point cloud data to determine a local neighborhood point set of the query point cloud. A geometric feature determination module is configured to determine multiple geometric features of each query point cloud based on the local neighborhood point set of the query point cloud, wherein the geometric features include at least point density feature, normal vector consistency feature, local plane fitting residual feature, curvature change rate feature, principal component direction deviation feature, sphericity feature, flatness feature, and straightness feature. A point cloud division module is configured to perform normalization processing and dynamic weight optimization on the multiple geometric features of each query point cloud to construct a comprehensive feature of each query point cloud, and divide the multiple original point cloud data based on the comprehensive features of the multiple query point clouds to determine multiple point cloud categories. A noise removal module is configured to perform adaptive noise removal processing on the original point cloud data in each point cloud category to determine updated point cloud data, and repeatedly perform local neighborhood point set determination and point cloud division processing on the updated point cloud data until a convergence condition is met, and output denoised point cloud data.

9. An electronic device, comprising: The point cloud data processing device includes: A processor, a memory, and a bus, the memory storing machine readable instructions executable by the processor, when the electronic device is running, the processor and the memory communicating through the bus, the machine readable instructions being executed by the processor to perform the steps of the point cloud data processing method of the laser radar according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, A computer readable storage medium storing a computer program, the computer program being executed by the processor to perform the steps of the point cloud data processing method of the laser radar according to any one of claims 1 to 7.

Images