Object intelligent fitting pre-marking method and system based on laser radar point cloud data

Abstract

The invention discloses an object intelligent fitting pre-marking method and system based on laser radar point cloud data, and the method comprises the steps: carrying out the self-adaptive downsampling of original point cloud data, extracting local and global features, carrying out the classification clustering iterative optimization, and finally estimating the pose of an object, and carrying out the matching with a preset scene model. Uniformization processing, feature representation optimization, classification clustering precision improvement and geometric pose accurate output of the point cloud data are realized. Complex point cloud data can be effectively processed, the classification clustering precision is improved, the object pose is accurately estimated, the scene adaptability of a point cloud processing system is enhanced, and the efficiency and robustness of point cloud data processing are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of three-dimensional perception, and in particular to an object intelligent fitting pre-labeling method and system based on laser radar point cloud data. BACKGROUND

[0002] In the field of modern intelligent manufacturing and three-dimensional perception, the processing and analysis of point cloud data play a crucial role and become the core technology for promoting industrial automation and intelligentization. Point cloud is a collection of unordered and irregular points. In the face of complex scenes, existing mature 2D image feature extraction methods cannot be directly applied to 3D point cloud. Existing processing methods often show obvious limitations, especially when dealing with objects of different shapes, sizes and poses, which can easily lead to inaccurate feature extraction and poor classification results, resulting in insufficient accuracy in subsequent object contour construction and positioning. SUMMARY

[0003] To solve the above technical problems, the present application provides an object intelligent fitting pre-labeling method and system based on laser radar point cloud data, which can process complex point cloud data and improve the efficiency and robustness of point cloud data processing.

[0004] The present application provides an object intelligent fitting pre-labeling method based on laser radar point cloud data, comprising:

[0005] performing sparse processing on the original point cloud data to obtain uniform first point cloud data;

[0006] obtain the local feature description of the second point cloud data by obtaining the neighborhood density and spatial distribution characteristics of each point in the uniform first point cloud data;

[0007] performing multi-scale analysis on the local features to obtain the third point cloud data representing global features;

[0008] performing category prediction on the third point cloud data, re-extracting features and adjusting the regional distribution of regions with classification confidence lower than a preset threshold to obtain fourth point cloud data containing classification results;

[0009] performing grouping processing on the fourth point cloud data and adjusting the clustering density parameter according to the classification confidence information to obtain the fifth point cloud data of clustering division;

[0010] obtaining the matching degree of the density distribution characteristics and the classification category distribution of the clustering region according to the clustering division of the fifth point cloud data, and if the matching degree is lower than a preset threshold, dynamically adjusting the classification confidence threshold and re-extracting features to obtain the sixth point cloud data of optimized classification and clustering results;

[0011] According to the optimized classification and clustering results of the sixth point cloud data, adaptively optimize the fitting parameters of each cluster area to determine a set of fitting parameters;

[0012] The pose information of each cluster area is calculated according to the fitting parameter set to obtain a pose description, and the pose description is matched and analyzed with the preset scene model to obtain a pose output result.

[0013] In some embodiments, the performing of sparse processing on the original point cloud data to obtain uniform first point cloud data includes:

[0014] Scan the original point cloud data using a pre-established density detection tool to obtain the density range of each area and obtain a first density distribution map;

[0015] Setting an adaptive threshold for a first density region in the first density distribution map where the point cloud density is greater than a density threshold, dividing the first density region into subregions using a meshing tool, and determining a thinning ratio of the subregions;

[0016] If the point cloud density of the sub-region exceeds a preset threshold, the sub-region is thinned according to the thinning ratio using a downsampling tool, point cloud data that meets the uniform distribution condition is extracted from the sub-region to obtain an adjusted point cloud subset;

[0017] The adjusted point cloud subsets are merged using a data integration tool to obtain a merged data set; the merged data set is checked for consistency to obtain homogenized first point cloud data.

[0018] In some embodiments, obtaining the neighborhood density and spatial distribution characteristics of each point in the homogenized first point cloud data to obtain the second point cloud data described by the local characteristics includes:

[0019] Scanning the homogenized first point cloud data using a pre-established density detection tool to obtain point cloud density and spatial distribution characteristics, thereby obtaining a second density distribution map;

[0020] Using a neighborhood search tool to divide each point in the second density distribution map into a local area, calculating neighborhood characteristics and spatial distribution data of each point in the local area, and determining a neighborhood density distribution record for each point;

[0021] If the density difference in the neighborhood density distribution record exceeds the preset neighborhood threshold, a spatial feature extraction tool is used to extract features from the point cloud data in the local area, obtain the description of the local geometric features and spatial features of each point, and obtain a local feature dataset;

[0022] The data integration tool is used to organize the geometric description and feature extraction results of the local feature data set to obtain the second point cloud data of the local feature description.

[0023] In some embodiments, the grouping processing of the fourth point cloud data and adjusting the cluster density parameter according to the classification confidence information to obtain the clustered fifth point cloud data includes:

[0024] Obtaining deviation information from the classification results of the fourth point cloud data, and using a pre-established deviation calibration tool to perform data comparison for the distribution of regions where the classification confidence is lower than a preset threshold, to obtain the deviation distribution of the fourth point cloud data;

[0025] If there are significantly inconsistent areas in the deviation distribution, the local area is re-divided using the spatial partitioning tool to determine the adjusted regional range description;

[0026] According to the adjusted regional range description, the density parameter adjustment tool is used to dynamically update the cluster density to obtain a density parameter set;

[0027] Through the density parameter set, the density-based spatial clustering tool is used to group the classified fourth point cloud data, and calibrate it according to the distribution information corresponding to the area where the classification confidence is lower than the preset threshold to obtain the fifth point cloud data with clustering division results.

[0028] In some embodiments, the clustering of the fifth point cloud data to obtain a degree of match between the density distribution characteristics of the cluster area and the classification category distribution is performed. If the degree of match is lower than a preset threshold, the classification confidence threshold is dynamically adjusted and features are re-extracted to obtain sixth point cloud data with optimized classification and clustering results, including:

[0029] Obtain cluster density and regional distribution information from the clustering results of the fifth point cloud data, use a pre-established density analysis tool to compare regional distribution data, and obtain a uniformity description of the density distribution;

[0030] Based on the uniformity description of density distribution, the category matching tool is used to evaluate the matching degree between regional distribution and classification results, and the quantitative result of matching degree is obtained;

[0031] If the quantification result is lower than the preset threshold, the threshold adjustment tool is used to dynamically update the classification confidence to obtain the adjusted classification confidence threshold range;

[0032] According to the adjusted classification confidence threshold range, the feature extraction tool is used to re-acquire the feature distribution information of the fifth point cloud data, and the sixth point cloud data of the classification and clustering collaborative division result is obtained.

[0033] In some embodiments, based on the optimized classification and clustering results of the sixth point cloud data, a dynamic adjustment mechanism is used to adaptively optimize the fitting parameters of each cluster area to determine a set of fitting parameters, including:

[0034] According to the optimized classification and clustering results of the sixth point cloud data, a pre-established shape analysis tool is used to extract the distribution characteristics of each cluster area, obtain the shape and posture feature description within the area, and obtain the feature distribution record of the area;

[0035] Based on the feature distribution records, a parameter mapping tool is used to perform a preliminary match on the fitting parameters of the cluster area. If the matching result does not meet the preset threshold of the target match, the fitting parameter value is updated through the adjustment mechanism to determine the adjusted fitting parameter range;

[0036] Based on the adjusted fitting parameter range, the parameter calibration tool is used to perform a secondary adjustment on the fitting parameters of the cluster area to obtain a calibrated fitting parameter configuration. The calibrated fitting parameter configuration is integrated with the distribution characteristics of the cluster area using a data fusion tool to obtain a fitting parameter set.

[0037] In some embodiments, calculating the pose information of each cluster region according to the fitting parameter set to obtain a pose description includes:

[0038] According to the set of fitting parameters, a pre-established geometric calculation tool is used to obtain posture features and shape distribution data within the cluster area to obtain a posture description record;

[0039] Through the posture description record, the fitting parameters and posture features are integrated using data fusion tools to obtain the integrated data;

[0040] If the integrated data does not reach the preset threshold, the posture features are calibrated using the parameter adjustment tool to determine the calibrated posture parameter range;

[0041] According to the calibrated pose parameter range, a distribution analysis tool is used to perform a secondary comparison on the shape distribution of the cluster area to obtain geometric pose data;

[0042] A posture description is obtained according to the geometric posture data and the shape distribution of the cluster area.

[0043] In some embodiments, the present application provides an object intelligent fitting pre-marking system based on lidar point cloud data, including:

[0044] A sparse module is used to perform sparse processing on the original point cloud data to obtain uniform first point cloud data;

[0045] A local feature module is used to obtain the neighborhood density and spatial distribution characteristics of each point in the homogenized first point cloud data to obtain second point cloud data described by local features;

[0046] A global feature module, configured to perform multi-scale analysis on the local features to obtain third point cloud data represented by global features;

[0047] a classification module, configured to perform category prediction on the third point cloud data, re-extract features of regions where the classification confidence is lower than a preset threshold, and adjust the regional distribution to obtain fourth point cloud data containing the classification results;

[0048] a clustering module, configured to perform grouping processing on the fourth point cloud data and adjust a clustering density parameter according to classification confidence information to obtain clustered fifth point cloud data;

[0049] A classification and clustering collaborative module is used to obtain the matching degree between the density distribution characteristics of the cluster area and the classification category distribution based on the cluster division of the fifth point cloud data. If the matching degree is lower than a preset threshold, the classification confidence threshold is dynamically adjusted and features are re-extracted to obtain the sixth point cloud data with optimized classification and clustering results;

[0050] a fitting parameter acquisition module, configured to adaptively optimize the fitting parameters of each cluster area based on the optimized classification and clustering results of the sixth point cloud data, and determine a set of fitting parameters;

[0051] The posture determination module is used to calculate the posture information of each cluster area according to the fitting parameter set to obtain a posture description, and match and analyze the posture description with the preset scene model to obtain a posture output result.

[0052] In some embodiments, the present application also provides an electronic device comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, the method for intelligent object fitting and pre-marking based on lidar point cloud data as described above is implemented.

[0053] In some embodiments, the present application also provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute any one of the above-mentioned object intelligent fitting pre-marking methods based on lidar point cloud data.

[0054] This application has the following beneficial effects:

[0055] (1) The application performs adaptive downsampling on the original point cloud data, then extracts local and global features, and then performs classification and clustering iteration optimization, and finally estimates the object pose and matches the scene, realizing the uniformization processing of point cloud data, optimization of feature representation, improvement of classification and clustering accuracy, and accurate output of geometric pose; it can effectively process complex point cloud data, improve the classification and clustering accuracy, accurately estimate the object pose, enhance the scene adaptability of the point cloud processing system, and improve the efficiency and robustness of point cloud data processing;

[0056] (2) The application uses a pre-established point cloud sparsification algorithm to sparsify the original point cloud data to obtain uniform first point cloud data. Specifically, different density regions are defined according to the density distribution map, and a grid division tool is used to divide multiple sub-regions. By calculating the number of point clouds in each sub-region, the sparsification ratio is determined, which can reduce the computing power of high-density regions and retain sufficient data features. The sparsified point cloud subset is more uniformly distributed in space, effectively avoiding data redundancy and improving processing efficiency.

[0057] (3) The application obtains the neighborhood density and spatial distribution features of each point in the first point cloud data to obtain second point cloud data with local feature description, and performs multi-scale analysis on the local features to obtain third point cloud data with global feature representation. By combining multi-scale division and spatial feature integration, the comprehensiveness of feature description is effectively improved, providing more accurate data support for subsequent point cloud processing.

[0058] (4) The application has a dynamic adjustment mechanism to adapt to point cloud data of different densities and distributions. The fitting parameters can be optimized according to the specific scene to better adapt to complex shape characteristics. On this basis, a secondary calibration scheme is also provided. If the fitting effect is still not ideal after preliminary parameter adjustment, the parameter values are further fine-tuned according to the feature distribution record to form a parameter configuration that better fits the actual distribution. This can effectively deal with the diversity of shapes and poses, improve the applicability of the parameter set, and further ensure accurate feature extraction and classification in complex point cloud data, so that the geometric pose information can be accurately output. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 is the flowchart of the object intelligent fitting pre-marking method based on laser radar point cloud data provided by the first embodiment of the application;

[0060] Figure 2 is the structure diagram of the object intelligent fitting pre-marking system based on laser radar point cloud data provided by the second embodiment of the application. DETAILED DESCRIPTION

[0061] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0062] Due to the complexity and diversity of point cloud data, and the fact that point cloud data usually exhibits disordered and irregular distribution characteristics, traditional feature extraction methods have difficulty in fully capturing the geometric information of objects, which directly leads to insufficient ability to discern the shape and posture of objects. This lack of discrimination ability further affects the clustering effect of point cloud data, making it difficult to effectively cluster point cloud data of the same object and unable to form a clear outline. In addition, there is a lack of dynamic adjustment mechanisms to adapt to point cloud data of different densities and distributions, resulting in the inability to optimize fitting parameters according to specific scenarios, ultimately affecting the accurate output of geometric pose information. Therefore, how to achieve accurate feature extraction and classification in complex point cloud data, while optimizing clustering and fitting parameters through dynamic adjustment mechanisms to output accurate geometric pose information, has become a key problem that needs to be solved urgently.

[0063] To solve the above problems, refer to Figure 1 The embodiment of the present application provides a method for intelligently fitting and pre-marking objects based on lidar point cloud data, comprising the following steps:

[0064] Step 101 : Use a pre-established point cloud thinning algorithm to perform thinning processing on the original point cloud data to obtain uniform first point cloud data.

[0065] In some embodiments, a pre-established density detection tool is used to scan the original point cloud data, and hierarchical statistics are performed on different areas according to the point cloud density to obtain the density range of each area and obtain a first density distribution map. An adaptive threshold is set for the first density area in the first density distribution map where the point cloud density is greater than the density threshold, and a grid division tool is used to divide the first density area into sub-areas, and the thinning ratio of the sub-areas is determined. If the point cloud density of the sub-area exceeds the preset threshold, the sub-area is thinned according to the thinning ratio using a downsampling tool, and point cloud data that meets the uniform distribution condition is extracted from the sub-area to obtain an adjusted point cloud subset. The adjusted point cloud subsets are merged using a data integration tool to obtain a merged data set. The merged data set is checked for consistency to obtain a uniformed first point cloud data.

[0066] In one possible implementation, the original point cloud data can be first fully scanned using a density detection tool to analyze the distribution characteristics of the original point cloud data in space. The density detection tool can be software such as CloudCompare, or other tools that can detect the density of point cloud data, and there is no specific limitation here. Suppose that the point cloud data of a three-dimensional scene is being processed. After scanning, it is found that the point cloud density in some areas is as high as 1,000 points per cubic meter, while in other areas it is only 100 points per cubic meter. This density difference will lead to uneven subsequent processing, so it is necessary to generate a density distribution map to intuitively display these differences. The density distribution map can be a three-dimensional heat map, where the depth of color represents the density.

[0067] Furthermore, based on the density distribution map, areas with point cloud density exceeding 800 points per cubic meter can be defined as first density areas or high density areas. Use the meshing tool to divide the first density area into multiple sub-areas, each of which can be 1 cubic meter in size. By calculating the number of point clouds in each sub-area, the thinning ratio is determined. For example, for a sub-area with a density of 900 points per cubic meter, the thinning ratio can be set to 50%, that is, half of the point cloud data is retained. This can specifically reduce the computing power of high-density areas while retaining sufficient data features. The sparse point cloud subset is more in line with the uniform distribution requirements in space, effectively avoiding data redundancy and improving processing efficiency.

[0068] The adjusted point cloud subsets need to be merged using a data integration tool. For example, consider 10 point cloud subsets in subregions, each containing approximately 500 points. After merging, a single dataset containing 5,000 points is formed. A consistency check is performed on the merged dataset to determine whether the point cloud is evenly distributed in space. For example, this can be done by counting the number of points within each grid cell to ensure that density differences are within 10%. This verification method ensures the quality of the final data and meets the goal of uniform distribution.

[0069] Step 102: Obtain the neighborhood density and spatial distribution characteristics of each point in the homogenized first point cloud data to obtain second point cloud data described by local characteristics.

[0070] In some embodiments, a pre-established density detection tool is used to scan the homogenized first point cloud data, and the distribution of each point is hierarchically counted to obtain the point cloud density and spatial distribution characteristics, thereby obtaining a second density distribution map. A neighborhood search tool is used to divide each point in the second density distribution map into local areas, and the neighborhood characteristics and spatial distribution data of each point in the local area are calculated to determine the neighborhood density distribution record of each point. If the density difference in the neighborhood density distribution record exceeds the preset neighborhood threshold, a spatial feature extraction tool is used to extract features from the point cloud data in the local area, and a description of the local geometric features and spatial features of each point is obtained to obtain a local feature data set. A data integration tool is used to organize the geometric description and feature extraction results of the local feature data set to obtain second point cloud data with a local feature description.

[0071] In one possible implementation, the first uniform point cloud data can be fully scanned using a density detection tool to obtain the distribution information of the first point cloud data in space. Assuming that the point cloud data processed is a three-dimensional indoor scene, after scanning, it is found that the point cloud density in some areas, such as the corners, is higher, while the density in the open areas is lower. A density detection tool is used to generate a density distribution map to intuitively show the uneven density distribution. For example, there may be 1,200 points per cubic meter in the corner area, while there are only 200 points in the central area. Furthermore, a neighborhood search tool can be used to divide each point in the density distribution map into local areas. Assuming that each point is the center, a spherical area with a radius of 0.5 meters is defined as the local neighborhood, and the number and distribution characteristics of the point cloud in the area are calculated. The neighborhood search tool is used to record the neighborhood density of each point. For example, there are 50 points in the area near a certain point, and only 10 points near another point, forming a neighborhood density distribution record. If the density difference of certain areas in the neighborhood density distribution record exceeds the preset neighborhood threshold, for example, the threshold is set to 800 points per cubic meter, and the density of a local area reaches 1,000 points, the spatial feature extraction tool will be used to analyze the point cloud distribution pattern in the area, such as identifying whether these points are concentrated on a certain plane, or whether they present a specific geometric shape, such as a wall or the edge of an object, to obtain a local feature data set that describes the geometric and spatial characteristics of the area. Based on the local feature data set, the geometric description and feature extraction results of the point cloud can be unified through the data integration tool. Assuming that a local area is identified as a wall feature, the integration tool is used to compare the geometric feature description of the area with the feature description of the homogenized point cloud data to determine whether it conforms to the local geometric feature description. The final feature description will contain the geometric feature information of each local area, such as the plane equation parameters of the wall area or the curvature change of the edge area. The above method ensures the comprehensiveness and consistency of the feature description.

[0072] Step 103: Perform multi-scale analysis on the local features to obtain third point cloud data represented by global features.

[0073] In some embodiments, a pre-established feature extraction tool is used to perform a layered scan on the second point cloud data, and the geometric characteristics of each local area are divided to obtain the feature distribution data in the local area to obtain a local feature set. A multi-scale division tool is used to perform multi-scale regional segmentation on the point cloud data of the local feature set, and the local geometric characteristics at each scale are compared to determine the feature change records of the local area at different scales. If the difference in the feature change record exceeds the preset difference threshold, a spatial feature integration tool is used to organize the feature data of different scales to obtain an integrated spatial feature description set. A data consistency verification tool is used to perform a global comparison of the feature descriptions in the integrated spatial feature description set to determine whether the feature descriptions meet the global feature representation requirements, and obtain the third point cloud data represented by the global feature.

[0074] In one possible implementation, a feature extraction tool is used to identify differences in the geometric shape of local areas. Assuming that the point cloud data processed is of an indoor three-dimensional scene, the feature extraction tool will initially divide areas such as walls, floors, and object surfaces to form a set of local features. For example, the wall area may be identified as a planar feature, while the object surface may exhibit curved surface characteristics. The second point cloud data is further segmented at multiple scales using a multi-scale segmentation tool to capture changes in local geometric characteristics. For example, taking an indoor scene as an example, the area may first be divided at a scale of 1.0 meters to identify larger wall features. Details are then analyzed at a scale of 0.2 meters to identify decorative objects or edge features on the wall. This multi-scale comparison generates a feature change record that records the differences in features at different scales, providing a basis for subsequent analysis. When the difference in the feature change record exceeds a preset difference threshold, a spatial feature integration tool is used to integrate the feature data at different scales to generate a unified spatial feature description. For example, if the preset difference threshold is 0.5 units of feature curvature change, and the curvature change of a certain area at a small scale reaches 0.8 units, the overall planar features of the wall surface are combined with the curvature change features of the edge to form a complete description set for subsequent processing. For another example, if the predetermined global feature representation requires that the wall area account for at least 60%, and the integrated feature description set shows that the wall area accounts for 65%, the verification tool will confirm that it meets the requirement. Conversely, if the proportion is only 50%, the feature extraction parameters may need to be readjusted. This verification ensures the reliability of the global feature representation data.

[0075] For example, from another side, the combination of multi-scale division and spatial feature integration can effectively improve the comprehensiveness of feature description. Assuming that in an indoor scene, a corner area is identified as a complex geometric shape at a small scale, and is classified as part of a wall at a large scale, the integration tool will integrate the features of both to form a more accurate description. The technical effect brought by this multi-angle analysis is significant, which can provide more accurate data support for subsequent point cloud processing.

[0076] At step 104, a preset point cloud classification model is used to perform category prediction on the third point cloud data, and feature re-extraction is performed on the region with a classification confidence lower than a preset threshold, and the region distribution is adjusted, to obtain fourth point cloud data containing a classification result.

[0077] In some embodiments, according to the global feature representation of the third point cloud data, a pre-established classification model is used to perform category division on the third point cloud data, to obtain a classification confidence corresponding to each region of the third point cloud data, to obtain a category division record. If the classification confidence is lower than a preset threshold, a feature re-extraction tool is used to perform feature re-extraction on the region with a classification confidence lower than the preset threshold, to obtain an updated feature description of the region, and to determine a feature set after re-extraction. A spatial feature integration tool is used to adjust the point cloud region distribution of the feature set after re-extraction, and to perform data consistency comparison on the adjusted local region, to obtain an integrated region feature distribution. Through the integrated region feature distribution, a category division record tool is used to compare the feature description of each region with the global feature representation, to obtain a classification result.

[0078] In a possible implementation, the third point cloud data is divided into different categories according to geometric characteristics based on a pre-trained point cloud classification model. For example, in an indoor scene, the data can be initially divided into categories such as wall, floor, and object. For example, the point cloud data processed is an indoor point cloud data, and the point cloud classification model can identify that the wall region accounts for 70% of the total data, and assign a classification confidence value to each region, such as a wall region with a confidence of 0.85, a floor with a confidence of 0.90, and a small object region with a confidence of only 0.60. For the case where the classification confidence is lower than a preset threshold, assuming that the threshold is set to 0.75, the small object region with a confidence of 0.60 needs to perform re-feature extraction on the local data through a feature re-extraction tool. Specifically, the feature re-extraction tool will re-scan the region to obtain geometric information, such as increasing the scanning resolution from 0.1 meters to 0.05 meters, to capture finer surface features. After re-feature extraction, it can be found that the region is actually a table corner, and the feature description is updated from a fuzzy object category to a geometric shape with corners, forming a feature set after re-extraction.

[0079] In one possible implementation, a spatial feature integration tool is used to readjust the point cloud area distribution of the re-extracted feature set. For example, if a table corner was originally misclassified as part of the ground, the spatial feature integration tool will associate it with other object areas based on the updated feature description to form an updated area distribution. Subsequently, through data consistency comparison, it may be found that the features of the table corner area and the nearby table leg area are highly similar, thus integrating them into a complete table feature distribution.

[0080] In one possible implementation, a classification recording tool is used to compare the integrated regional feature distribution with the global feature representation. For example, in the global feature representation, the feature description of the object class includes corners and curved surfaces. The classification recording tool will confirm that the table corners and legs meet the object category and ultimately classify them as a table. This comparison ensures classification accuracy.

[0081] In one possible implementation, reliability can be improved through multi-angle analysis. For example, in the initial classification, the classification confidence of a certain area is 0.72, slightly below the threshold of 0.75. By combining the feature distribution of adjacent areas for auxiliary judgment, for example, finding that this area is highly correlated with wall features, unnecessary re-extraction can be avoided, thus optimizing processing efficiency.

[0082] In one possible implementation, for example, if the integrated regional feature distribution shows that the table region accounts for 10% of the total data, while the object category in the global feature representation accounts for an expected 8% to 12% of the total data, the comparison result is consistent with expectations and the classification result is confirmed. This consistency check improves the stability of the classification.

[0083] Step 105 : performing grouping processing on the fourth point cloud data, and adjusting cluster density parameters according to classification confidence information to obtain fifth point cloud data divided into clusters.

[0084] In some embodiments, the deviation information is obtained from the classification result of the fourth point cloud data. For the region distribution with a classification confidence lower than a preset threshold, a pre-established deviation calibration tool is used for data comparison to obtain a deviation distribution of the fourth point cloud data. If there is a region with significant inconsistency in the deviation distribution, a spatial division tool is used to resegment the local region to determine an adjusted region range description. According to the adjusted region range description, a density parameter adjustment tool is used to dynamically update the clustering density to obtain a density parameter set. Through the density parameter set, a density-based spatial clustering tool is used to group process the classified fourth point cloud data, and according to the distribution information corresponding to the region with a classification confidence lower than the preset threshold, the fifth point cloud data with a clustering division result is obtained. In this embodiment, the deviation distribution information generally reflects the possible misjudgment or data incompleteness problem of the region with a low classification confidence.

[0085] In a possible implementation, assuming that the point cloud data of an indoor scene is processed, the classification result shows that a certain region is classified as the ground, but the classification confidence is only 0.65, which is lower than the preset threshold. Through the pre-established deviation calibration tool, the data of the region can be compared with the global feature representation, and it is found that the geometric characteristics of the region are closer to low objects than to the ground, and the deviation distribution accounts for about 5% of the total data.

[0086] The spatial division tool processes the region with significant inconsistency in the deviation distribution. Assuming that the low object region described above is continuous with other ground regions in space, but the feature difference is obvious, the low object region is segmented from the large range ground and redefined as an independent region, and the range description can be reduced from the original 10 square meters to 2 square meters. This resegmentation helps to more accurately focus on the problem region. Further description of the adjusted region is performed, and a density parameter adjustment tool is used to further optimize data integration. Assuming that the initial value of the point cloud density of the local region is 100 points per cubic meter, but the feature distribution shows that the data is sparse, the density parameter adjustment tool adjusts the density parameter to 200 points per cubic meter to form a suitable density parameter set. This adjustment can better capture the subtle geometric features in the region and improve the accuracy of subsequent clustering. Based on the density parameter set, a density-based spatial clustering tool is used to group process the fourth point cloud data. For example, the adjusted local region is divided into two subregions through clustering, one subregion shows a smooth surface, and the other subregion shows an irregular shape. According to the distribution information of the classification confidence, the two subregions are calibrated by using the density-based spatial clustering tool, and the fifth point cloud data with a clustering division result is finally obtained. This grouping process can effectively distinguish regions with large feature differences and provide support for classification calibration.

[0087] Step 106: Based on the clustering division of the fifth point cloud data, the matching degree between the density distribution characteristics of the cluster area and the classification category distribution is obtained. If the matching degree is lower than the preset threshold, the classification confidence threshold is dynamically adjusted and the features are re-extracted to obtain the sixth point cloud data with optimized classification and clustering results.

[0088] In some embodiments, cluster density and regional distribution information are obtained from the clustering results of the fifth point cloud data, and a pre-established density analysis tool is used to compare regional distribution data to obtain a uniformity description of the density distribution. Based on the uniformity description of the density distribution, a category matching tool is used to evaluate the degree of match between the regional distribution and the classification result to obtain a quantitative result of the match. If the quantitative result is lower than a preset threshold, a threshold adjustment tool is used to dynamically update the classification confidence to obtain an adjusted classification confidence threshold range. Based on the adjusted classification confidence threshold range, a feature extraction tool is used to re-acquire feature distribution information of the fifth point cloud data to obtain the sixth point cloud data representing the classification and clustering collaborative division result.

[0089] In this example, cluster density reflects the density of point cloud data in space, while regional distribution describes the clustering characteristics of data within different spatial ranges. For example, in point cloud data processing for an indoor scene, the cluster density of one area may be 150 points per cubic meter, while the density of an adjacent area may be only 50 points per cubic meter. This discrepancy suggests that there may be uneven feature distribution, requiring further analysis.

[0090] In one possible embodiment, a pre-established density analysis tool is used to compare regional distribution characteristics to obtain a uniformity description of the density distribution. For example, in the aforementioned indoor scene, the comparison reveals that the density distribution of a certain area fluctuates greatly, with some sub-areas having a density as high as 200 points per cubic meter and others having only 30 points per cubic meter. Therefore, the uniformity description is rated as poor.

[0091] In one possible implementation, assume that a certain area is classified as a desktop with a classification confidence of 0.75, but its density distribution uniformity is low. After evaluation, the quantitative result of the degree of match is 0.6, which is lower than the preset threshold of 0.8. This indicates that the classification result may be biased and requires further adjustment. If the degree of match is lower than the threshold, the classification confidence data is dynamically updated using a threshold adjustment tool. For example, if the initial classification confidence threshold is 0.7, the classification confidence threshold is adjusted to 0.85 based on feature integration requirements to screen out more reliable classification results. This adjustment can effectively filter data in low-confidence areas and provide a more accurate basis for subsequent processing.

[0092] Step 107 : Based on the optimized classification and clustering results of the sixth point cloud data, a dynamic adjustment mechanism is used to adaptively optimize the fitting parameters of each cluster area to determine a set of fitting parameters.

[0093] In some embodiments, based on the optimized classification and clustering results of the sixth point cloud data, a pre-established shape parsing tool is used to extract data from the distribution characteristics of each cluster area, obtain the shape and posture feature description within the area, and obtain the feature distribution record of the area. Based on the feature distribution record, a parameter mapping tool is used to perform a preliminary match on the fitting parameters of the cluster area. If the matching result does not reach the preset threshold of the target match, the fitting parameter value is updated through the adjustment mechanism to determine the adjusted fitting parameter range. Based on the adjusted fitting parameter range, a parameter calibration tool is used to perform a secondary adjustment on the fitting parameters of the cluster area to obtain a calibrated fitting parameter configuration. A data fusion tool is used to integrate the calibrated fitting parameter configuration with the distribution characteristics of the cluster area to obtain a fitting parameter set.

[0094] In this embodiment, the fitting parameter is a key quantitative indicator used to describe the degree of spatial matching between the clustering area and the preset scene model (or target geometric model). Its core function is to establish a geometric correspondence between the clustering area of ​​point cloud data and the preset scene model. The fitting parameter is a bridge parameter connecting point cloud clustering and pose estimation. Through the four-step optimization of feature extraction, mapping, dynamic adjustment, and calibration, high-precision spatial matching of complex point cloud data and the preset scene model is achieved, providing a basis for subsequent pose description. The fitting parameter set is a set of dynamically adjusted geometric description parameters used to characterize the following relationships: shape fit, which indicates the degree of matching between the point cloud distribution form (such as surface curvature, plane normal, edge direction) of the clustering area and the geometric features of the preset scene model; posture fit, which indicates the difference between the orientation, position offset, etc. of the clustering area and the target pose. Local matching weight indicates the contribution weight of different sub-areas to the overall matching (for example, the parameter weight of the edge area may be higher than that of the flat area). For example, if the clustering area is a cylindrical object, the fitting parameters may include: the direction vector of the cylinder axis, the coordinates of the bottom circle center, the radius error tolerance threshold, and the side curvature matching degree.

[0095] Furthermore, in a possible implementation, the fitting parameters may be generated through an adaptive optimization mechanism, with the following specific steps:

[0096] First, shape features are extracted. Specifically, shape analysis tools (such as PCA principal component analysis and RANSAC geometric fitting) can be used to extract features from the optimized clustering results of the sixth point cloud data to obtain a geometric feature description of each cluster area. These geometric features include: surface curvature distribution, normal vector set, boundary contour feature points, and center of mass position.

[0097] Then, the extracted geometric features are coarsely matched with a pre-set scene model (e.g., a CAD template) using a parameter mapping tool (e.g., a feature matching-based algorithm such as ICP, Iterative Closest Point) to calculate initial fitting parameters (e.g., a translation vector, a rotation matrix). If the matching error exceeds a threshold (e.g., the mean square error of the point cloud to the surface of the pre-set scene model > 0.1), a parameter adjustment mechanism is triggered. If there is a local mismatched area (e.g., a curvature mutation) in the feature distribution record, the parameter range is dynamically adjusted (e.g., the rotation angle is widened from [0°, 5°] to [0°, 10°]) and the weight distribution is optimized (e.g., the weight of the edge feature points is increased).

[0098] In a possible implementation, a parameter mapping tool (e.g., a regression model or a deep learning model) is used to preliminarily match the feature distribution of the clustered area with the fitting parameters. If the matching result of the initial fitting parameters does not meet the pre-set threshold, the fitting parameter value is updated through an adjustment mechanism to better adapt to complex shape characteristics. This adjustment helps to improve the accuracy of area division and lays a foundation for subsequent calibration. Further, it is particularly important to use a parameter calibration tool to make a secondary adjustment of the fitting parameters. For example, after a preliminary parameter adjustment, the fitting effect of a certain area is still not ideal, and the calibration tool will further fine-tune the parameter value according to the feature distribution record to form a parameter configuration that better fits the actual distribution. This secondary calibration can effectively cope with the diversity of shapes and poses and improve the applicability of the parameter set.

[0099] In a possible implementation, a data fusion tool integrates the calibrated parameter configuration with the distribution characteristics of the clustered area. For example, in the same indoor scene, the calibration parameter configuration of a certain area shows a high fitting degree, and after the fusion tool combines it with the distribution characteristics, the fitting parameter set is determined. This integration process ensures the matching of the parameters with the actual data and helps to generate a point cloud data set that better meets business needs.

[0100] Step 108: calculating the pose information of each clustered area according to the fitting parameter set, obtaining a pose description, and matching and analyzing the pose description with a pre-set scene model to obtain a geometric pose output result.

[0101] In some embodiments, according to the fitting parameter set of step 107 above, the pose information of each clustered area is calculated using a geometric pose estimation algorithm to obtain an eighth point cloud data of the pose description.

[0102] In this embodiment, based on the set of fitting parameters, a pre-established geometric calculation tool is used to extract the pose information of the cluster area, from which the posture features and shape distribution data within the cluster area are obtained to obtain a pose description record. Based on the pose description record, a data fusion tool is used to integrate the fitting parameters and the pose features to obtain integrated data. If the integrated data does not reach a preset threshold, the pose features are calibrated using a parameter adjustment tool to determine the calibrated pose parameter range. Based on the calibrated pose parameter range, a distribution analysis tool is used to perform a secondary comparison of the shape distribution of the cluster area to obtain geometric pose data. Based on the geometric pose data and the shape distribution of the cluster area, the pose description content is obtained.

[0103] In one possible implementation, a geometric calculation tool is used to parse the point cloud data within the clustering area to obtain posture features, such as direction and tilt, and shape distribution data, such as boundary range and geometric center.

[0104] In this embodiment, geometric calculation tools such as principal component analysis (PCA) and minimum bounding box (OBB-Oriented Bounding Box) can be used to calculate the local posture in the local coordinate system. For example, the implementation process of principal component analysis (PCA) is as follows: calculate the eigenvalues ​​and eigenvectors of the point cloud covariance matrix, and obtain the local posture in the local coordinates based on the eigenvectors. Specifically, the coordinates of the center point of the point cloud are usually the mean point, the eigenvector V1 corresponding to the maximum eigenvalue is usually used as the main axis direction (such as the long axis direction of the object), and the V2 corresponding to the second largest eigenvalue and the V3 corresponding to the minimum eigenvalue constitute the local coordinate system. The posture can be expressed by (V1, V2, V3) or by a rotation matrix.

[0105] At the same time, geometric calculation tools are used to calculate the shape distribution data describing the cluster area. These shape distribution data specifically include: size, surface area (estimated area of ​​the point cloud surface), volume (estimated volume of the space occupied by the point cloud), compactness (such as (area^3) / (volume^2) or based on the bounding sphere), anisotropy (based on the ratio of the eigenvalues ​​calculated by the above PCA (such as λ1 / λ3)), and point distribution histograms describing the distribution of points in different directions or distances in space.

[0106] Furthermore, data fusion tools are used to integrate the fitting parameters with the pose features. Fitting parameters can be expressed as rigid transformation parameters that align the clustered regions with the pre-set scene model. They are typically expressed as a rotation matrix (R): a 3x3 matrix describing rotation; a translation vector (T): a 3x1 vector describing displacement; or a combination of these into a 4x4 homogeneous transformation matrix.

[0107] The essence of data fusion is to use fitting parameters to transform the local pose of each cluster region in its own local coordinate system into a unified and meaningful global coordinate system or reference coordinate system. The specific implementation process is as follows:

[0108] First, we obtain the fitting parameters of the cluster regions and perform pose and shape data conversion based on these fitting parameters. Each cluster region has a specific fitting parameter, which implicitly defines the position of the cluster region in the reference coordinate system. This transformation consists of two parts: pose transformation and shape distribution data transformation, as follows:

[0109] The local posture of the cluster area is transformed into the target coordinate system (the coordinate system pointed to by the fitting parameters, usually the global coordinate system or the CAD coordinate system) to obtain the global posture feature to achieve posture feature transformation.

[0110] Transform the shape distribution data from the local coordinate system to the target coordinate system to achieve shape distribution data conversion. If the shape distribution data does not change with the coordinate system transformation (size, volume, point distribution statistics, etc. are inherent properties of the object), the shape data can be directly retained.

[0111] Finally, the transformed global posture features are combined with the shape distribution data (and cluster ID) to obtain the integrated data.

[0112] In one possible implementation, a distribution analysis tool (such as kernel density estimation) performs a secondary comparison of the shape distribution. For example, in the same indoor environment, the initial analysis of the shape distribution data for a certain area reveals unclear boundaries, but after secondary comparison, it is found that its geometric center is offset by approximately 0.2 units. This comparison allows the tool to obtain more accurate geometric pose data and determine whether it meets the target accuracy requirements.

[0113] Furthermore, in some embodiments, a pre-established comparison tool is used to perform a preliminary comparison between the pose description and the scene model, obtaining matching analysis data and determining a scene matching degree. If the matching degree is below a preset threshold, a correction tool is used to adjust the pose description to obtain adjusted pose correction data. A data integration tool is used to perform a secondary matching of the adjusted pose correction data with the scene model to obtain a matching result. Based on the matching result, a formatting tool is used to organize the adjusted pose correction data to obtain a final geometric pose output result.

[0114] In one possible implementation, a pre-established comparison tool is employed to extract key feature points from the pose description and the pre-set scene model, and perform matching analysis to determine the spatial correspondence between the two. For example, in an indoor environment, the pose information of a certain area shows a directional deviation of about 10 degrees from the standard value, while the scene model requires a directional error within 5 degrees. After analysis by the comparison tool, the preliminary matching degree value is 0.6, which is lower than the pre-set threshold value of 0.8. Further adjustment of the pose description is made using a correction tool. This adjustment process ensures that the pose information is closer to the requirements of the scene model, laying a foundation for subsequent matching.

[0115] In one possible implementation, a data integration tool is employed to perform secondary matching between the corrected pose description and the pre-set scene model. The secondary matching takes into account multiple parameters in the corrected data, such as direction, position, and boundary features, to ensure that the matching result is more in line with the adaptive requirements.

[0116] In one possible implementation, according to the requirements of the final output result, a formatting tool is employed to organize the corrected pose description and determine the geometric pose output content. The main function of the formatting tool is to arrange the data according to a specific structure or standard, facilitating subsequent application. For example, the formatting tool may organize key parameters such as directional error, positional offset, etc. into a unified format, ensuring the standardization of the output content. This organization method helps to improve the readability and usability of the data, providing convenience for subsequent analysis or application.

[0117] Through the object intelligent fitting pre-marking method based on lidar point cloud data of the above embodiment, the original point cloud data is sparsely processed to obtain the first uniform point cloud data; the neighborhood density and spatial distribution characteristics of each point in the uniform first point cloud data are obtained to obtain the second point cloud data described by the local features; the local features are subjected to multi-scale analysis to obtain the third point cloud data represented by the global features; the category prediction is performed on the third point cloud data, and the features are re-extracted and the regional distribution is adjusted for the areas where the classification confidence is lower than the preset threshold to obtain the fourth point cloud data containing the classification results; the fourth point cloud data is grouped and adjusted according to the classification confidence information. The cluster density parameters are adjusted to obtain the fifth point cloud data of cluster division; according to the cluster division of the fifth point cloud data, the matching degree between the density distribution characteristics of the cluster area and the classification category distribution is obtained. If the matching degree is lower than the preset threshold, the classification confidence threshold is dynamically adjusted and the features are re-extracted to obtain the sixth point cloud data of optimized classification and clustering results; according to the optimized classification and clustering results of the sixth point cloud data, the fitting parameters of each cluster area are adaptively optimized, a fitting parameter set is determined, and the posture information of each cluster area is calculated according to the fitting parameter set to obtain a posture description, and the posture description is matched and analyzed with the preset scene model to obtain a geometric posture output result. In the above embodiment, the original point cloud is first adaptively downsampled, and then local and global features are extracted, and then classification and clustering iterative optimization is performed, and finally the object posture is estimated and matched with the scene, thereby achieving uniform processing of point cloud data, feature representation optimization, classification and clustering accuracy improvement, and geometric posture accurate output. It can effectively process complex point cloud data, improve classification and clustering accuracy, accurately estimate object pose, enhance the scene adaptability of the point cloud processing system, and improve the efficiency and robustness of point cloud data processing.

[0118] Reference Figure 2 The present application also provides an object intelligent fitting pre-marking system based on lidar point cloud data, including:

[0119] A sparseness module 201 is used to perform sparseness processing on the original point cloud data to obtain uniform first point cloud data;

[0120] A local feature module 202 is configured to obtain neighborhood density and spatial distribution characteristics of each point in the homogenized first point cloud data to obtain second point cloud data described by local features;

[0121] A global feature module 203 is configured to perform multi-scale analysis on the local features to obtain third point cloud data represented by global features;

[0122] A classification module 204 is configured to perform category prediction on the third point cloud data, re-extract features of regions where the classification confidence is lower than a preset threshold, and adjust the regional distribution to obtain fourth point cloud data containing the classification results;

[0123] A clustering module 205 is configured to perform grouping processing on the fourth point cloud data and adjust a clustering density parameter according to classification confidence information to obtain clustered fifth point cloud data;

[0124] The classification and clustering collaborative module 206 is configured to determine the degree of match between the density distribution characteristics of the clustered area and the classification category distribution based on the clustering division of the fifth point cloud data. If the degree of match is lower than a preset threshold, the classification confidence threshold is dynamically adjusted and features are re-extracted to obtain sixth point cloud data with optimized classification and clustering results.

[0125] A fitting parameter acquisition module 207 is configured to adaptively optimize the fitting parameters of each cluster region based on the optimized classification and clustering results of the sixth point cloud data to determine a set of fitting parameters;

[0126] The pose determination module 208 is used to calculate the pose information of each cluster area according to the fitting parameter set to obtain a pose description, and perform matching analysis between the pose description and the preset scene model to obtain a pose output result.

[0127] It should be noted that the object intelligent fitting and pre-marking system based on lidar point cloud data provided in the embodiment of the present application is used to execute all the process steps of the object intelligent fitting and pre-marking method based on lidar point cloud data in the above embodiment. The working principles and beneficial effects of the two correspond one to one, so they will not be repeated here.

[0128] The present application also provides an electronic device. The electronic device includes: a processor, a memory, and a computer program stored in the memory and executable on the processor, such as a data acquisition program. When the processor executes the computer program, the steps in the above-mentioned method for intelligent object alignment and pre-marking based on laser radar point cloud data are implemented, such as Figure 1 Alternatively, when the processor executes the computer program, the functions of the modules / units in the above-mentioned system embodiments are realized, such as the data acquisition module.

[0129] Exemplarily, the computer program may be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present application. The one or more modules / units may be a series of computer program instruction segments capable of completing specific functions, which are used to describe the execution process of the computer program in the electronic device.

[0130] The electronic device may be a computing device such as a desktop computer, notebook, PDA, or smart tablet. The electronic device may include, but is not limited to, a processor and memory. Those skilled in the art will appreciate that the aforementioned components are merely examples of electronic devices and do not constitute a limitation of the electronic device. The electronic device may include more or fewer components than those described above, or a combination of certain components, or different components. For example, the electronic device may also include input / output devices, network access devices, buses, and the like.

[0131] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the electronic device, connecting various parts of the entire electronic device using various interfaces and lines.

[0132] The memory can be used to store the computer programs and / or modules, and the processor realizes various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application required for a function (such as a sound playback function, an image playback function, etc.); the data storage area can store data created based on the use of the mobile phone (such as audio data, a phone book, etc.). In addition, the memory can include a high-speed random access memory and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage device.

[0133] Wherein, if the module / unit integrated in the electronic device is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application implements all or part of the process in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and when the computer program is executed by the processor, it can implement the steps of the above-mentioned various method embodiments. Wherein, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.

[0134] It should be noted that the device embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed across multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment. In addition, in the drawings of the device embodiments provided in this application, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines. A person of ordinary skill in the art can understand and implement the present invention without inventive work.

[0135] The specific embodiments described above further illustrate the objectives, technical solutions, and beneficial effects of this application. It should be understood that the above descriptions are merely specific embodiments of this application and are not intended to limit the scope of protection of this application. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of this application by those skilled in the art should be included within the scope of protection of this application.

Claims

1. An object intelligent fitting pre-marking method based on lidar point cloud data, characterized in that: include: Performing sparse processing on the original point cloud data to obtain uniform first point cloud data; Obtaining neighborhood density and spatial distribution characteristics of each point in the homogenized first point cloud data to obtain second point cloud data described by local characteristics; Performing multi-scale analysis on the local features to obtain third point cloud data represented by global features; Performing category prediction on the third point cloud data, re-extracting features of areas where the classification confidence is lower than a preset threshold and adjusting the area distribution, to obtain fourth point cloud data containing the classification result; performing grouping processing on the fourth point cloud data, and adjusting a cluster density parameter according to classification confidence information to obtain cluster-divided fifth point cloud data; Based on the clustering of the fifth point cloud data, the matching degree between the density distribution characteristics of the cluster area and the classification category distribution is obtained. If the matching degree is lower than a preset threshold, the classification confidence threshold is dynamically adjusted and features are re-extracted to obtain the sixth point cloud data with optimized classification and clustering results; According to the optimized classification and clustering results of the sixth point cloud data, adaptively optimize the fitting parameters of each cluster area to determine a set of fitting parameters; The pose information of each cluster area is calculated according to the fitting parameter set to obtain a pose description, and the pose description is matched and analyzed with the preset scene model to obtain a pose output result.

2. The method according to claim 1, characterized in that The step of performing sparse processing on the original point cloud data to obtain uniform first point cloud data includes: Scan the original point cloud data using a pre-established density detection tool to obtain the density range of each area and obtain a first density distribution map; Setting an adaptive threshold for a first density region in the first density distribution map where the point cloud density is greater than a density threshold, dividing the first density region into subregions using a meshing tool, and determining a thinning ratio of the subregions; If the point cloud density of the sub-region exceeds a preset threshold, the sub-region is thinned according to the thinning ratio using a downsampling tool, point cloud data that meets the uniform distribution condition is extracted from the sub-region to obtain an adjusted point cloud subset; The adjusted point cloud subsets are merged using a data integration tool to obtain a merged data set; the merged data set is checked for consistency to obtain homogenized first point cloud data.

3. The method according to claim 1, characterized in that The obtaining of the neighborhood density and spatial distribution characteristics of each point in the homogenized first point cloud data to obtain the second point cloud data described by the local characteristics includes: Scanning the homogenized first point cloud data using a pre-established density detection tool to obtain point cloud density and spatial distribution characteristics, thereby obtaining a second density distribution map; Using a neighborhood search tool to divide each point in the second density distribution map into a local area, calculating neighborhood characteristics and spatial distribution data of each point in the local area, and determining a neighborhood density distribution record for each point; If the density difference in the neighborhood density distribution record exceeds the preset neighborhood threshold, a spatial feature extraction tool is used to extract features from the point cloud data in the local area, obtain the description of the local geometric features and spatial features of each point, and obtain a local feature dataset; The data integration tool is used to organize the geometric description and feature extraction results of the local feature data set to obtain the second point cloud data of the local feature description.

4. The method according to claim 1, wherein The grouping process of the fourth point cloud data and adjusting the cluster density parameter according to the classification confidence information to obtain the fifth point cloud data divided by clustering includes: Obtaining deviation information from the classification results of the fourth point cloud data, and using a pre-established deviation calibration tool to perform data comparison for the distribution of regions where the classification confidence is lower than a preset threshold, to obtain the deviation distribution of the fourth point cloud data; If there are significantly inconsistent areas in the deviation distribution, the local area is re-divided using the spatial partitioning tool to determine the adjusted regional range description; According to the adjusted regional range description, the density parameter adjustment tool is used to dynamically update the cluster density to obtain a density parameter set; Through the density parameter set, the density-based spatial clustering tool is used to group the classified fourth point cloud data, and calibrate it according to the distribution information corresponding to the area where the classification confidence is lower than the preset threshold to obtain the fifth point cloud data with clustering division results.

5. The method according to claim 1, wherein The clustering division of the fifth point cloud data is performed to obtain a matching degree between the density distribution characteristics of the cluster area and the classification category distribution. If the matching degree is lower than a preset threshold, the classification confidence threshold is dynamically adjusted and features are re-extracted to obtain sixth point cloud data with optimized classification and clustering results, including: Obtain cluster density and regional distribution information from the clustering results of the fifth point cloud data, use a pre-established density analysis tool to compare regional distribution data, and obtain a uniformity description of the density distribution; Based on the uniformity description of density distribution, the category matching tool is used to evaluate the matching degree between regional distribution and classification results, and the quantitative result of matching degree is obtained; If the quantification result is lower than the preset threshold, the threshold adjustment tool is used to dynamically update the classification confidence to obtain the adjusted classification confidence threshold range; According to the adjusted classification confidence threshold range, the feature extraction tool is used to re-acquire the feature distribution information of the fifth point cloud data, and the sixth point cloud data of the classification and clustering collaborative division result is obtained.

6. The method according to claim 1, characterized in that Based on the optimized classification and clustering results of the sixth point cloud data, a dynamic adjustment mechanism is used to adaptively optimize the fitting parameters of each cluster area to determine the fitting parameter set, including: According to the optimized classification and clustering results of the sixth point cloud data, a pre-established shape analysis tool is used to extract the distribution characteristics of each cluster area, obtain the shape and posture feature description within the area, and obtain the feature distribution record of the area; Based on the feature distribution records, a parameter mapping tool is used to perform a preliminary match on the fitting parameters of the cluster area. If the matching result does not meet the preset threshold of the target match, the fitting parameter value is updated through the adjustment mechanism to determine the adjusted fitting parameter range; According to the adjusted fitting parameter range, a parameter calibration tool is used to perform a secondary adjustment on the fitting parameters of the cluster area to obtain a calibrated fitting parameter configuration, and a data fusion tool is used to integrate the calibrated fitting parameter configuration with the distribution characteristics of the cluster area to obtain a fitting parameter set.

7. The method according to claim 1, characterized in that Calculating the pose information of each cluster region according to the fitting parameter set to obtain a pose description includes: According to the set of fitting parameters, a pre-established geometric calculation tool is used to obtain posture features and shape distribution data within the cluster area to obtain a posture description record; Through the posture description record, the fitting parameters and posture features are integrated using data fusion tools to obtain the integrated data; If the integrated data does not reach the preset threshold, the posture features are calibrated using the parameter adjustment tool to determine the calibrated posture parameter range; According to the calibrated pose parameter range, a distribution analysis tool is used to perform a secondary comparison on the shape distribution of the cluster area to obtain geometric pose data; A posture description is obtained according to the geometric posture data and the shape distribution of the cluster area.

8. The object intelligent fitting pre-marking system based on laser radar point cloud data is characterized by: include: A sparse module is used to perform sparse processing on the original point cloud data to obtain uniform first point cloud data; A local feature module is used to obtain the neighborhood density and spatial distribution characteristics of each point in the homogenized first point cloud data to obtain second point cloud data described by local features; A global feature module, configured to perform multi-scale analysis on the local features to obtain third point cloud data represented by global features; a classification module, configured to perform category prediction on the third point cloud data, re-extract features of regions where the classification confidence is lower than a preset threshold, and adjust the regional distribution to obtain fourth point cloud data containing the classification results; a clustering module, configured to perform grouping processing on the fourth point cloud data and adjust a clustering density parameter according to classification confidence information to obtain clustered fifth point cloud data; A classification and clustering collaborative module is used to obtain the matching degree between the density distribution characteristics of the cluster area and the classification category distribution based on the cluster division of the fifth point cloud data. If the matching degree is lower than a preset threshold, the classification confidence threshold is dynamically adjusted and features are re-extracted to obtain the sixth point cloud data with optimized classification and clustering results; a fitting parameter acquisition module, configured to adaptively optimize the fitting parameters of each cluster area based on the optimized classification and clustering results of the sixth point cloud data, and determine a set of fitting parameters; The posture determination module is used to calculate the posture information of each cluster area according to the fitting parameter set to obtain a posture description, and match and analyze the posture description with the preset scene model to obtain a posture output result.

9. An electronic device, characterized in that: The method comprises a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, the method for intelligent fitting and pre-marking of objects based on lidar point cloud data as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium includes a stored computer program, wherein, when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the object intelligent fitting pre-marking method based on lidar point cloud data as described in any one of claims 1 to 7.

Images