Point cloud identification method based on dynamic feature fusion and full-process dynamic parameter adjustment

Abstract

The invention discloses a point cloud identification method based on dynamic feature fusion and full-process dynamic parameter adjustment. The method comprises the following steps: step 1, collecting an original point cloud of an industrial part and removing invalid points, constructing a training set in combination with a CAD model point cloud, calculating an average point spacing based on a k-d tree, dynamically adjusting the leaf size of a voxel grid, and generating a standardized point cloud; 2, calculating a point cloud normal vector, extracting a CVFH feature descriptor and an SHOT feature descriptor, dynamically fusing the two types of features based on the average point spacing and the spatial range of the point cloud, and carrying out smoothing processing; step 3, using the fusion features to train a KNN classification model, establishing a mapping relation between the features and target categories, and storing model parameters; 4, after the test point cloud is processed in the step 1 and the step 2, model parameters are input, and a prediction result with the highest confidence coefficient is output and output in a log. According to the method, the problem of low recognition precision caused by insufficient global and local feature capture of a complex industrial part by a single feature descriptor in the prior art is solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of three-dimensional point cloud processing and object recognition, and particularly relates to a point cloud recognition method based on dynamic feature fusion and full-process dynamic parameter adjustment. BACKGROUND

[0002] In the field of three-dimensional point cloud processing, point cloud object recognition technology is the core of computer vision and robotics, and is widely used in industrial part recognition, scene understanding and target detection fields. Point cloud recognition extracts the geometric and topological features of point clouds, providing important support for industrial automation. However, existing methods still face challenges in feature extraction, dynamic parameter adjustment and real-time adaptability, limiting their comprehensive application in complex industrial scenarios.

[0003] Specifically, in terms of feature extraction, traditional single descriptors such as Clustered Viewpoint Feature Histogram (CVFH) and Fast Point Feature Histogram (FPFH) are difficult to accurately capture both global shape and local details of point clouds, resulting in insufficient discrimination ability in complex scenarios. Existing feature fusion methods rely on fixed weights, which cannot effectively balance global and local features, and have poor adaptability to changes in point cloud density and target size, restricting the performance of the recognition method. In terms of dynamic parameter adjustment, existing technologies usually use fixed parameter configurations, which lack flexibility and are difficult to adapt to different point cloud acquisition devices (such as depth cameras or laser radars), object view changes, size, etc., limiting the robustness and recognition accuracy of the system. In terms of real-time and adaptability, the high computational complexity of large-scale point cloud processing makes it difficult for existing methods to meet the real-time needs of industrial applications, especially in multi-view point cloud scenarios, existing technologies lack adaptability and are difficult to efficiently handle the complex needs of diversified application scenarios.

[0004] Although existing technologies attempt to improve performance through Point Cloud Librar (PCL) CVFH, SHOT descriptor, and Random Forest, Support Vector Machine (SVM) classifiers, they still face challenges such as insufficient dynamic parameter adjustment, poor feature fusion robustness, and difficulty in balancing global and local features, and the recognition rate needs to be improved. Therefore, there is an urgent need for an efficient, robust and adaptable point cloud object recognition method.

[0005] The application proposes a point cloud recognition method based on dynamic feature fusion and full-process dynamic parameter adjustment, which combines the excellent performance of CVFH in describing obvious geometric structures with the strong adaptability of SHOT to complex surfaces, solves the problem that the existing technology cannot simultaneously accurately capture the global shape and local details of point clouds, resulting in insufficient discrimination ability in complex scenes, especially when there are large occlusions, defects, and missing key features, and the recognition accuracy is low. At the same time, it also solves the problem of fixed parameter adjustment and single thread processing time being too long, which cannot truly meet the requirements of generalization and real-time performance, significantly improving the accuracy, generalization, and real-time performance of point cloud object recognition. The research of this comprehensive recognition model will provide innovative ideas and reliable solutions for the further development of three-dimensional point cloud processing technology in the field of industrial automation. SUMMARY

[0006] The purpose of the application is to provide a point cloud recognition method based on dynamic feature fusion and full-process dynamic parameter adjustment, which solves the problem of low recognition accuracy caused by insufficient capture of global and local features of complex industrial parts by a single feature descriptor in the prior art.

[0007] The technical solution adopted by the application is a point cloud recognition method based on dynamic feature fusion and full-process dynamic parameter adjustment, and the steps are as follows: Step 1: Collect the original point cloud of the industrial part and remove invalid points, combine with the CAD model point cloud to construct the training set, calculate the average point distance based on the k-d tree, dynamically adjust the size of the voxel grid leaf, and generate the standardized point cloud; Step 2, calculate the point cloud normal vector, extract the CVFH feature descriptor and the SHOT feature descriptor, dynamically fuse the two types of features based on the average point distance and spatial range of the point cloud, and perform smoothing processing; Step 3, use the fused features to train the KNN classification model, establish the mapping relationship between the features and the target categories, and save the model parameters; Step 4, after the test point cloud is processed by step 1 and step 2, input the model parameters, output the prediction result with the highest confidence, and output it in the log.

[0008] The application also has the following characteristics: In step 1, the specific process of generating the standardized point cloud is as follows: Step 1.1, set the initial voxel grid leaf size to 1.0mm, set the initial leaf size to 1.5mm when the point cloud size is greater than 50k points, and directly output the standardized point cloud after the first sampling if the point cloud size is between 5k points and 30k points. Step 1.2, if the original point cloud size is greater than 30k points, the average point distance is calculated based on the k-d tree, the leaf size is adjusted by adjusting the factor iteration, and the down-sampling is performed to generate the standardized point cloud; specifically: First, the average point distance is calculated as follows: (1) In formula (1), is the average point distance of the original point cloud, is the distance from the point to its nearest neighbor, is the total number of valid point clouds after removing invalid points; Second, the adjustment factor is calculated as follows: (2) In formula (2), is the adjustment factor, limiting the adjustment factor range to 0.7-1.7, is the adjustment intensity coefficient, the value is 0.5-1.0 according to the proportion of invalid points; Then, the leaf size is adjusted by iteration, and the iteration adjustment formula of the voxel grid leaf size is: (3) In formula (3), is the current leaf size, with the unit of mm, limiting its range to 0.5mm-3mm, is the adjustment factor; is the new leaf size generated after iteration, when the point cloud size is not in the range of 5k points-30k points, the iteration calculation is started, and the iteration is terminated until the point cloud size meets the requirements, according to the generated new leaf size, after down-sampling, the iteration is ended when the point cloud size meets the requirements of 5k-30k points, and the standardized point cloud after down-sampling is output; Step 1.3, if the original point cloud size is less than 5k points, the point cloud density is supplemented by up-sampling, and the specific steps are as follows: the k-d tree is constructed to calculate the nearest neighbor distance of each point, the area with a point distance greater than 3mm is identified as a sparse area, and the points in the area are marked as center points, for each center point, 10 nearest neighbor points are selected, the weight is calculated based on the Gaussian function, and the weight center coordinates of the nearest neighbor points are calculated according to the weight, and then new points are inserted at the weight center coordinates to increase the density of the point cloud, after insertion, if the point cloud size is not less than 5k points, the standardized point cloud is directly output; otherwise, the point distance threshold is reduced by 0.1mm gradient, the center points are re-marked and the insertion process is repeated until the point cloud size is not less than 5k points, and the iteration is terminated, and the standardized point cloud is saved.

[0009] The CVFH feature descriptor in step 2 is 308-dimensional, and the SHOT feature descriptor is 352-dimensional. After normalization processing of the extracted feature descriptors, a new 660-dimensional CVFH-SHOT feature descriptor is obtained by concatenation.

[0010] The specific process of step 2 is as follows: Step 2.1, calculate the normalized point cloud normal vector based on k-d tree and radius search; first, take 5 times the average point distance as the initial search radius, limit its range to 10mm-30mm, if the search radius is less than 10mm, default to 10mm, if the search radius is greater than 30mm, default to 30mm, then accelerate the radius search through k-d tree, get the neighboring points of each point, fit the local plane based on principal component analysis, calculate the covariance matrix of the point cloud local area, the normal vector is the eigenvector corresponding to the smallest eigenvalue of the covariance matrix; Step 2.2, use the CVFHEstimation algorithm provided by the PCL library to extract the 308-dimensional CVFH feature descriptor from the normalized point cloud, specifically: calculate the normal and curvature information of the normalized original point cloud, segment the point cloud based on region growing algorithm, use the smoothness of the normal direction and the curvature threshold, the angle difference between the normal directions is less than 15° and the curvature value is less than 0.05 as the judgment basis of similar geometric properties, divide the point cloud into multiple regions; calculate the angle distribution between the view direction and the normal direction and the shape distribution for each region, construct a multi-dimensional CVFH feature histogram, generate a 308-dimensional global CVFH feature descriptor by k-means clustering and concatenating the feature histograms of all regions in order; Step 2.3, when extracting the 352-dimensional SHOT feature descriptor from the normalized point cloud using the SHOTEstimationOMP algorithm, first select key points from the normalized original point cloud, use an adaptive strategy based on curvature: select the top 30% non-edge points with curvature values as key points, the selected non-edge points are defined as points with a distance to the point cloud boundary of not less than 2mm; for areas with low point cloud density, i.e. areas with an average point distance of the normalized point cloud greater than 3mm, increase the number of key points to the top 40% of curvature, and keep the definition of non-edge points unchanged; define a spherical support region centered on each key point, divide the support region into 32 sub-regions based on the local reference frame (LRF), and count the angle distribution of the normal vector of each point in each sub-region with the LRF axis to generate an 11-dimensional direction histogram, then concatenate the histograms of all sub-regions in order to generate a 352-dimensional SHOT feature descriptor; Step 2.4, generate a 660-dimensional fusion feature and smooth the feature.

[0011] Step 2.4 is based on the average point distance of the point cloud and the spatial range , dynamically calculates the CVFH and SHOT fusion weight, generates a 660-dimensional fusion feature, and the specific process is as follows: Step 2.4.1, the acquisition of the spatial range is obtained by comparing the coordinates of each point in the normalized point cloud, updating the minimum point ( , , ) and the maximum point ( , , ), calculating the maximum difference of the three-dimensional range as max( , , ), is the maximum range of the point cloud in x, y, z dimensions, unit: mm; Step 2.4.2, calculate the initial weight of CVFH based on the average point distance, the specific calculation method is as follows: If mm, then: (4) If mm, then: (5) In formula (4), (5), is the average point distance of the normalized point cloud, the calculation method is referred to formula (1), 1.5mm is the target point distance, 0.8 is the maximum weight of CVFH, and 0.3 is the adjustment range, is the CVFH weight; Step 2.4.3, based on the fine-tuning weight of the spatial range, further optimize the distribution of CVFH and SHOT weight, the specific implementation process is as follows: Calculate the range factor : (6) In formula (6), is the range factor; is the maximum range of the point cloud in x, y, z dimensions, unit: mm; 100 and 150 are normalization parameters; The clamp function limits in [-0.1, 0.1]; Adjust the CVFH weight: (7) SHOT weight: (8) In formula (7), (8), is the weight of SHOT, which is limited by the clamp function in [0.2, 0.8]; Step 2.4.4, the normalized 308-dimensional CVFH feature descriptor and the 352-dimensional SHOT feature descriptor are spliced by weighting, specifically, the CVFH feature is processed by maximum value normalization, each component of the 308-dimensional histogram is divided by the maximum value, and scaled to the range [0, 1]; the SHOT feature is first averaged for all valid descriptors, and then normalized by the maximum value to scale the result to the range [0, 1]; then the normalized CVFH and SHOT features are spliced in order in series, the first 308 dimensions are CVFH features, and the last 352 dimensions are SHOT features, forming an initial 660-dimensional feature, the dynamic weight calculated based on formulas (4)-(8) is used to weight and fuse the spliced features to generate a CVFH-SHOT 660 dynamic weighted feature.

[0012] The formula for smoothing processing in step 2.4 is as follows: (9) In formula (9), is the feature dimension (with a value of 0 to 659), and the boundary condition is filled with mirror image; smoothing processing is completed by a single iteration to suppress noise and enhance feature consistency.

[0013] In step 3, a KNN classification model is used for training, and a mapping relationship between the feature vector and the target category is established, specifically: Load the training data set: import the training set obtained in step 1, convert it to a CVFH-SHOT 660 fusion feature descriptor training set using the fusion feature vector method generated in step 2, covering several categories of industrial parts, divided into simple geometric class, complex structure class, and small feature class; use the KNN classification model to train the feature descriptor training set, save the model parameters obtained by training to an XML file in a specified path; at the same time, based on the XML file, cross-validation is performed, the training set is used as the test set input, the cross-validation result is used to optimize the model parameters and verify the rationality.

[0014] The test point cloud in step 4 is the original point cloud of the industrial part that did not participate in the training, and steps 1 and 2 are performed on the test point cloud to generate a test dynamic weighted fusion CVFH-SHOT 660 feature, which is compared and evaluated with the KNN model parameter XML file trained in step 3, the confidence level is determined by calculating the distance between the test point cloud feature and the feature of each part in the training set, the class corresponding to the highest confidence level is taken as the prediction result, and the log form is output.

[0015] The beneficial effects of the present application are: (1) Dynamic feature fusion to improve recognition ability: The fusion weight of CVFH and SHOT is dynamically adjusted by the average point distance and spatial range, effectively balancing the global geometric features (CVFH) and local surface details (SHOT), and improving the adaptability of the features to point cloud density and scale changes. The fusion generates 660-dimensional CVFH-SHOT features, which significantly improve the discrimination ability in complex scenes compared to a single descriptor. Experiments show that the recognition rate is improved by about 10-15% in the case of no occlusion, and more than 25% for occluded and incomplete point clouds.

[0016] (2) Adaptive point cloud preprocessing to ensure feature quality: Multi-stage voxel grid downsampling combined with dynamic point distance adjustment strategy reduces the point cloud size from 19,000 to 40,000 points to 5,000 to 3,000 points, and optimizes the point cloud quality by filtering out invalid points to provide stable input for feature extraction; the dynamic weight mechanism can automatically optimize the classifier weight to enhance robustness.

[0017] (3) Dynamic parameter adjustment throughout the process to ensure consistency: The preprocessing, feature extraction and classification stages maintain dynamic parameter adjustment consistency, such as leaf size, radius search, fusion weight, etc., to avoid recognition rate decline due to parameter conflicts, and to ensure processing efficiency, generalization and feature quality.

[0018] (4) Multi-source data enhancement to enhance generalization ability: Fusion of CAD models and real captured point clouds to build a training set, which improves the recognition rate by about 18% compared to a single CAD model training set, enhancing the model's adaptability to multi-angle, occlusion and incomplete point clouds.

[0019] (5) High interpretability for easy optimization: Record feature statistics, classification probability and failure files, support visual analysis, and facilitate system optimization and debugging.

[0020] (7) Targeted optimization of industrial part recognition: Optimizing feature extraction and classification strategies for several types of industrial parts, using CVFH to capture the global geometric shape of the parts, and combining SHOT to extract local details, improving multi-class discrimination ability. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 is the overall flowchart of the point cloud recognition method based on dynamic feature fusion and dynamic parameter adjustment throughout the process of the present application; Figure 2 is the dynamic feature fusion architecture diagram in the present application. DETAILED DESCRIPTION

[0022] The present application will be described in detail below in conjunction with the drawings and specific embodiments.

[0023] The point cloud recognition method based on dynamic feature fusion and dynamic parameter adjustment throughout the process is as follows: Figure 1 , Figure 2The specific process is as follows: Step 1, collect multi-view original point clouds of different industrial parts through a depth camera, remove NaN invalid points using the remove non-numerical point function provided by the point cloud library (PCL), and retain valid point cloud data; combine the single-view point cloud of the CAD model collected by the virtual camera, the complete-view point cloud to form a training set, calculate the average point distance based on the k-d tree, dynamically adjust the size of the voxel grid leaf for downsampling or upsampling, and generate standardized point clouds; the CAD model point cloud is a three-dimensional CAD model of the corresponding industrial part constructed by a three-dimensional drawing software, which is converted into a point cloud format; 154 single-view point clouds and 1 complete-view point cloud are collected by a virtual camera, and the 154 views are generated based on icosahedral subdivision; the rendering resolution is set to 1280 720, the visual angle is fixed at 90°, which can control the viewing direction of the camera and cover a larger observation range to obtain as many single-view point clouds as possible; the camera-to-model center distance is 0.5m, the viewing angle is generated based on icosahedral subdivision, and the initial vertices are 12 vertices on the unit sphere as the initial viewing direction; the triangular faces of the icosahedron are subdivided twice, and the midpoints of the three edges of the triangle are projected to the unit sphere each time, and the original triangle is divided into four small triangles by connecting the new vertices. After two times of subdivision, a total of 162 viewing angles are generated, and after removing the redundant viewing angles (viewing angle direction difference less than a set threshold), 154 effective single-view point clouds are retained. The original point cloud of the industrial part after removing the invalid points, the corresponding 154 single-view CAD point clouds and one complete-view CAD point cloud are combined to form a training set; the average point distance of the point cloud is calculated based on the k-d tree, and the size of the voxel grid leaf is dynamically adjusted for downsampling or upsampling to concentrate the point cloud size in the range of 5k points-30k points, and the standardized point cloud is generated.

[0024] The specific process of generating the standardized point cloud is as follows: Step 1.1, set the initial voxel grid leaf size to 1.0mm, the voxel grid leaf size refers to the edge length of a single cubic grid in voxelization processing, which is used to control the point cloud density after downsampling or upsampling. When the point cloud size is greater than 50k points, set the initial leaf size to 1.5mm to make it faster to reach the target range of 5k points-30k points in subsequent downsampling, and directly output the standardized point cloud after the first sampling if the point cloud size is already within the range of 5k points-30k points; Step 1.2, if the original point cloud size is greater than 30k points, calculate the average point distance based on the k-d tree, and adjust the leaf size through an adjustment factor to perform downsampling, so as to concentrate the original point cloud size within the range of 5k points-30k points, thereby generating the standardized point cloud. Specifically: The dynamic adjustment of the leaf size is limited to a range of 0.5mm-3.0mm, and the target point distance is set to 1.5mm; the specific method for dynamically adjusting the leaf size is as follows (1)-(3): First, the calculation formula of the average point distance is as follows: (1) In formula (1), is the average point distance of the original point cloud before preprocessing, is the distance from the point to its nearest neighbor point, is the total number of valid points after removing invalid points.

[0025] Second, the calculation formula of the adjustment factor is as follows: (2) In formula (2), is the adjustment factor, and the range of the adjustment factor is limited to 0.7-1.7, is the adjustment intensity coefficient, and the value range of k is 0.5-1.0; the specific value is determined according to the noise level of the point cloud: when the invalid point ratio is greater than 10% (high noise), k takes a value of (0.8, 1.0] to enhance the adjustment sensitivity; when the invalid point ratio is [5%, 10%] (moderate noise), k takes a value of [0.7, 0.8] to balance the adjustment sensitivity and stability; when the invalid point ratio is less than 5% (high quality), k takes a value of [0.5, 0.7); the intermediate ratio is determined by linear interpolation. When the invalid point ratio is 10%, k takes a value of 0.8; when the invalid point ratio is 5%, k takes a value of 0.7, and when the invalid point ratio approaches 0, k approaches 0.5 to maintain stability; Then, the leaf size is adjusted by iteration, and the iteration adjustment formula of the voxel grid leaf size is as follows: (3) In formula (3), is the current leaf size, with a unit of mm, and the range is limited to 0.5mm-3mm, is the adjustment factor; is the new leaf size generated after iteration, and when the point cloud size is not in the range of 5k points-30k points, the iteration calculation is started, and the iteration is terminated when the point cloud size meets the requirements; that is, when the original point cloud size meets the requirements after downsampling according to the generated new leaf size, the iteration is ended, and the standardized point cloud after downsampling is output.

[0026] Step 1.3, if the original point cloud size is less than 5k points, the point cloud density needs to be supplemented by upsampling, the specific steps are: constructing k-d tree to calculate the nearest neighbor distance of each point, identifying the area with a point spacing greater than 3mm as a sparse area, marking the points in the area as center points, for each center point, selecting its 10 nearest neighbor points, calculating the weight based on the Gaussian function, and according to the weight, the nearest neighbor point weight center coordinates are inserted into the new point to increase the density of the point cloud, the Gaussian function is: wherein, is the three-dimensional coordinate of the center point (mean value), σ is the standard deviation, and the default is 1, is the Euclidean distance from the nearest neighbor point to the center point; after insertion, if the point cloud size is not less than 5k points, the standardized point cloud is directly output; otherwise, the point spacing threshold of the sparse area is reduced by 0.1mm gradient, the center points are marked again and the point insertion process is repeated until the point cloud size is not less than 5k points, and the iteration is terminated, and the standardized point cloud is saved.

[0027] Step 2, using the standardized point cloud generated in step 1, calculating the point cloud normal vector based on k-d tree and radius search, extracting 308-dimensional CVFH feature descriptor and 352-dimensional SHOT feature descriptor, normalizing the extracted feature descriptor, and using chain splicing to generate new 660-dimensional CVFH-SHOT feature descriptor; then, combining the point cloud average point spacing and space range to dynamically adjust the fusion weight, generating a fusion feature vector and performing smoothing processing; the specific steps are as follows: Step 2.1, normal vector calculation, The normal vector of the standardized point cloud is calculated based on k-d tree and radius search; the specific steps are as follows: first, take 5 times the average point spacing as the initial search radius, limit its range to 10mm-30mm, if the search radius is less than 10mm, it is defaulted to 10mm, if the search radius is greater than 30mm, it is defaulted to 30mm, then, through k-d tree acceleration radius search, the nearest neighbor points of each point are obtained, based on principal component analysis (PCA) to fit the local plane, calculate the covariance matrix of the local area of the point cloud, the normal vector is the eigenvector corresponding to the smallest eigenvalue of the covariance matrix, that is, the normal vector of the local area, which is perpendicular to the local plane of the point cloud surface, used to describe the orientation of the point cloud surface.

[0028] The judgment of normal vector validity verification needs to meet the following conditions: first, the components of the normal vector on the X-axis, Y-axis and Z-axis are valid values (non-NaN or infinite); second, ensure that the number of adjacent points returned by the radius search is not less than 3 points (three points and above can fit a plane) to ensure that the plane can be fitted; third, verify that the length of the normal vector is between 0.9-1.1 (to ensure that it is a unit normal vector, as a non-unit normal vector may have calculation errors); and fourth, count the proportion of the total sum of invalid normal vectors in the three judgment conditions of a single point cloud, and if it is not greater than 30%, it is judged that the calculated normal vector is valid.

[0029] Step 2.2, extract 308-dimensional CVFH feature descriptor, use the clustering view feature histogram estimation (CVFH Estimation) algorithm provided by the PCL library to extract the 308-dimensional CVFH feature descriptor of the standardized point cloud, the process is as follows: calculate the normal and curvature information of the standardized point cloud, segment the point cloud based on the region growing algorithm, use the smoothness of the normal direction and the curvature threshold, and take the angle difference between the normal directions less than 15° and the curvature value less than 0.05 as the judgment threshold of similar geometric properties, and the points meeting the conditions are classified into the same region, realizing the region segmentation of the point cloud, and laying a foundation for the subsequent construction of the CVFH feature histogram. The point cloud is segmented into multiple regions, the angle distribution between the view direction and the normal direction and the shape distribution (distance distribution and angle distribution of points to the centroid) are calculated for each region, and a multi-dimensional CVFH feature histogram is constructed, and through k-means clustering and concatenating the feature histograms of all regions in order, a 308-dimensional global CVFH descriptor is generated (if less than 308 dimensions, use zero padding). The normal here refers to the vector determined by the eigenvector corresponding to the smallest eigenvalue of the local region covariance matrix after fitting the local plane based on principal component analysis, which is used to describe the orientation of the point cloud surface in the local region.

[0030] Step 2.3, extract 352-dimensional SHOT feature descriptors, use SHOTEstimationOMP algorithm to extract 352-dimensional SHOT descriptors from the normalized point cloud, the process is as follows: select key points from the normalized point cloud, that is, use curvature-based adaptive strategy to select non-edge points from the top 30% of points in curvature value ordering (defined as the distance from the point cloud boundary is not less than 2 mm); for the area with low point cloud density (average point spacing is greater than 3 mm), expand to the top 40% of non-edge points in curvature value; define a spherical support region centered on each key point, divide the support region into 32 sub-regions (2 radial partitions, 8 azimuthal partitions, 2 polar partitions) based on the local reference frame (LRF), and generate an 11-dimensional direction histogram in each sub-region by counting the angle distribution of the normal vector of the point with the LRF axis, and concatenate all sub-region histograms in order to form a 352-dimensional local SHOT descriptor. The local reference frame (LRF) takes the key point as the origin, the x-axis is along the tangent direction of the point cloud surface, the z-axis is along the normal vector direction, and the y-axis is determined by the cross product to ensure the consistency of the direction of sub-region division.

[0031] Step 2.4, dynamic weight fusion, generate 660-dimensional fusion features, and smooth the features.

[0032] According to the average point spacing of the point cloud and the spatial range , dynamically calculate the CVFH and SHOT fusion weights, wherein the CVFH weight ranges from 0.2 to 0.8, and the SHOT weight = 1-CVFH weight, generate 660-dimensional fusion features, and smooth the features to enhance robustness.

[0033] The specific implementation method is as follows: Step 2.4.1, obtain the spatial range by comparing the coordinates of each point in the normalized point cloud, update the minimum point , , ) and the maximum point , , ), calculate the maximum difference of the three-dimensional range as max( , , ), (the maximum range of the point cloud in x, y, z dimensions is mm), and count the maximum difference of each coordinate point in space to obtain the spatial range; Step 2.4.2, calculate the initial weight of CVFH based on the average point spacing, the specific calculation method is as follows: If mm, then: (4) If mm, then: (5) In formula (4), (5), is the average point distance of the pre-processed normalized point cloud (calculation method refers to formula (1)), 1.5mm is the target point distance, 0.8 is the maximum weight of CVFH, and 0.3 is the adjustment range, is the CVFH weight; Step 2.4.3, fine-tuning weight based on spatial range, further optimizes the distribution of CVFH and SHOT weights.

[0034] The purpose of fine-tuning is to further optimize the distribution of CVFH and SHOT weights by introducing spatial range information, to adapt to the geometric scale and density changes of the point cloud, and to ensure the robustness and discrimination ability of feature extraction. The specific benefits are: First, the initial weight is only based on the average point distance , which reflects the local density, but ignores the overall geometric scale of the point cloud. Fine-tuning can make up for the limitations of single point distance weighting; second, the initial weight may be biased towards a certain feature, and fine-tuning ensures that both contribute, avoiding the dominance of a single feature and making the weight unbalanced, enhancing the robustness of the feature; third, the scale of industrial part point cloud varies greatly (from a few centimeters to a few tens of centimeters), and fine-tuning dynamically adjusts the weight according to the spatial range of the point cloud, making it more adaptable to diverse needs.

[0035] The specific implementation process is as follows: Calculate the range factor : (6) In formula (6), is the range factor; is the maximum range of the point cloud in x, y, z dimensions, with units of mm; 100 and 150 are normalization parameters (set according to the middle value of common industrial part sizes, 100mm is the average size of common industrial parts, and 150mm is the size standard deviation); the clamp function limits to [-0.1, 0.1].

[0036] Adjust the CVFH weight: (7)SHOT weight: (8)wherein, is the weight of SHOT, which is limited by function In [0.2, 0.8], it is ensured that both features can effectively contribute, avoiding a single feature from dominating the fusion process, thereby balancing the expression of global and local features.

[0037] Step 2.4.4, the normalized 308-dimensional CVFH feature descriptor and the 352-dimensional SHOT feature descriptor are spliced together with weights; specifically, the CVFH feature is normalized by dividing each component of the 308-dimensional histogram by the maximum value, scaling it to the [0, 1] range; the SHOT feature is first averaged over all valid descriptors, then normalized by the maximum value to scale the result to the [0, 1] range; then the normalized CVFH and SHOT features are spliced in order (the first 308 dimensions are CVFH features, the last 352 dimensions are SHOT features, forming an initial 660-dimensional feature), the dynamic weights calculated based on formulas (4)-(8) are used to weight the spliced features, and the dynamic weights are used to weight the spliced feature components (CVFH feature components multiplied by , SHOT feature components multiplied by ), generating a final 660-dimensional fusion feature; this fusion feature combines the excellent expression of global geometric structure by CVFH and the adaptability to complex surfaces by SHOT, providing a robust and efficient feature input for the KNN model.

[0038] In step 2.4, the fusion feature is smoothed to enhance robustness, and the smoothing formula is: (9) In formula (9), is the feature dimension (values from 0 to 659), and the boundary condition is filled with mirror images (e.g. ); smoothing is completed through a single iteration to suppress noise and enhance feature consistency.

[0039] Based on formula (9), first, ensure that the input fusion feature vector is an array of length 660 containing dimension values from 0 to 659; then, create a temporary array of the same size as the original vector to store the smoothing result; then, iterate through each dimension from 0 to 659 , according to the boundary condition, mirror filling: when , the previous dimension takes the value of ; when , the next dimension takes the value of the values of the adjacent dimensions, and the smoothed feature values are calculated according to formula (9) (the weight of the previous dimension is 0.3, the weight of the current dimension is 0.4, and the weight of the next dimension is 0.3); finally, the results of the temporary array are assigned back to the original feature vector, and the smoothing process is completed. For example, for , the calculation is , and the smoothing process is completed through a single iteration, thereby suppressing noise and enhancing the consistency of the features.

[0040] Step 3, using the fusion feature vector method generated in step 2, the training set obtained in step 1 is converted into a fusion feature CVFH-SHOT660 fusion feature, and a KNN classification model is used for training, and a mapping relationship between the feature vector and the target category is established, specifically: Load the training data set: import the training set obtained in step 1, use the fusion feature vector method generated in step 2 to convert the training set obtained in step 1 into a fusion feature CVFH-SHOT660 feature descriptor training set (covering several categories of industrial parts, divided into simple geometric categories (square, shaft), complex structure categories (double concave circular table, inclined hollow block), and small feature categories (bolt)); use a KNN classification model to train the feature descriptor training set, save the model parameters obtained by training to an XML file in a specified path; at the same time, based on the XML file, cross-validation is performed (using the training set as the test set input, verifying the method rationality), and the model parameters are optimized and verified for rationality through the cross-validation results.

[0041] Step 4, perform the processing of steps 1 and 2 on the test point cloud (i.e. the original point cloud of the industrial part not participating in the training), and generate a test dynamic weighted fusion CVFH-SHOT660 feature; compare and evaluate this feature with the KNN model parameter XML file trained in step 3, determine the matching confidence by calculating the distance between the test point cloud and the feature of each part in the training set, take the class corresponding to the highest confidence as the prediction result, and output it in the form of a log (including the predicted category, the confidence of each category, the feature statistical information, and the processing time).

[0042] The recognition performance of the test object is analyzed through the log output results: The recognition result: records and statistics the prediction category of the test point cloud, the confidence distribution of each category (for example, the confidence percentage of each part for part 1), and the key parameters in the feature extraction process (including the average point distance , the fusion weight of CVFH and SHOT , and , etc.); Performance indicators: calculate the accuracy rate (target value not less than 85%); statistics single sample full process average processing time: the whole process time mean from point cloud input to output prediction result, the target value is not more than 0.5 seconds; the judgment rule is: if the accuracy rate and processing time are up to standard, the method meets the scientificity (recognition accuracy is reliable) and real-time (efficient response) requirements of industrial scene; if not, backtrack and analyze the optimization space of the voxel grid leaf size adjustment in the preprocessing stage, the dynamic weight distribution in the feature extraction stage and the model parameters (such as the K value of KNN), and improve pertinently to improve the performance.

[0043] Embodiment 1 Reference Figure 1 Take the simulation of industrial part identification scene as an example: which includes identifying 9 kinds of industrial parts such as square block, bolt, short boss, etc. Use Lingyun light Y800M camera to obtain part point cloud, point cloud scale 20-40 million points, hardware is Intel i5-13490f CPU and NVIDIA RTX 4060 GPU. Implementation steps: (1) Preprocessing: first filter out the invalid points of the original point cloud collected by the camera and retain 80% of the point cloud, and then perform adaptive voxel grid downsampling to reduce the point cloud scale to 5 thousand-3 million points to obtain standardized point cloud; the adaptive voxel grid downsampling process dynamically adjusts the leaf size of downsampling as shown in equations (1) to (3), and the leaf size is dynamically adjusted by the average point distance and the dynamic adjustment factor to ensure that the point cloud scale is between 5 thousand-3 million points.

[0044] (2) Feature dynamic weighted fusion: First, the normalized point cloud normal vector is calculated based on k-d tree and radius search, the search radius is set to 5 times the average point distance, the range is limited to 10-30 mm, if the search radius is less than 10 mm, the default search radius is 10 mm, if the search radius is greater than 30 mm, the default is 30 mm, then the k-d tree is accelerated to search for the neighboring points of each point, and the local plane is fitted based on principal component analysis, and the normal vector is calculated. After the normal vector is calculated, the CVFH and SHOT feature descriptors are extracted based on the calculated normal vector, and the extracted CVFH (308 dimensions) and SHOT (352 dimensions) descriptors are fused by dynamic weighted fusion to generate a new 660-dimensional dynamic fusion feature CVFH-SHOT660. The calculation of the dynamic weight is based on the geometric characteristics of the point cloud, such as the average point distance and the spatial range, to adaptively balance the contribution of the CVFH global geometric feature and the SHOT local surface detail. The average point distance of the normalized point cloud is calculated by k-d tree to reflect the point cloud density, as shown in formula (1); according to a large number of experiments in the previous work, it is known that the point cloud is best distributed with an average point distance of about 1.5 mm, so the ideal average point distance is set to 1.5 mm, and the preliminary CVFH weight is calculated according to the average point distance by formula (4) and (5). The point cloud spatial range is used to fine-tune the weight to obtain the best fusion effect, as shown in formula (6) and (7). The point cloud spatial range refers to the maximum range of the point cloud in the x, y, and z dimensions. In this way, different sizes of parts are considered, so that the global shape and local details of the point cloud can be accurately captured, and the discrimination ability in complex scenes is greatly improved. After obtaining the optimal CVFH weight, the SHOT weight is obtained as shown in formula (8). Then, smoothing processing is performed as shown in formula (9) to enhance its robustness.

[0045] (3) Feature training: 1845 point cloud files are used for feature training, including several types of parts, each type of part includes 154 single-view point clouds of CAD models collected by a virtual camera and one complete point cloud of a CAD model, plus 50 visual point clouds of different angles and distances captured by Lingyun light Y800M depth camera, which are trained together by KNN to ensure that each orientation feature of the target part can be detected, while the feature distribution of the real camera captured point cloud is integrated to make it have more scientific discrimination ability, and cross-validation is added to ensure the rationality of the training result. The average processing time of each link during training is shown in Table 1.

[0046] (4) Test: The test is performed using 458 test files of different views and different distances of several types of parts outside the test set. The test point cloud is put into the same folder and transmitted to the test code for comparison and prediction with the trained training result model parameter XML file. The predicted object result of each test point cloud is output in the form of a log. The number of specific parts tested is shown in Table 2.

[0047] (5) Result analysis: log output is added at each step for later analysis. The log result output shows that the cross-validation accuracy is 98.89%, proving that the trained model has high reliability; the average accuracy under the condition of no occlusion is higher than 96%. In this example, several types of parts are tested with a ratio of 1:1 of occlusion and no occlusion. The number of several types of parts tested and the accuracy are shown in Table 2. The overall average accuracy is 89.52%, and the effects of bolts, blocks, and double concave round table features are better. The average processing time of each link during training of 1845 parts is shown in Table 2, and the total time is 108.116s. The average processing time of each link for the 458 test parts of several types of parts during testing is shown in Table 3. The log result shows that the average time required to identify one part is 0.402s; Table 1 Average processing time of each link during training

[0048] Table 2 Recognition accuracy of each part

[0049] Table 3 Average processing time of each part of each link during testing

[0050] Example 2 The point cloud recognition method based on dynamic feature fusion and full-process dynamic parameter adjustment has the following steps: Step 1: Collect the original point cloud of industrial parts and remove invalid points. Combine the CAD model point cloud to build a training set. Calculate the average point distance based on k-d tree and dynamically adjust the size of the voxel grid leaf to generate a standardized point cloud. Step 2, calculate the point cloud normal vector, extract the CVFH feature descriptor and the SHOT feature descriptor, dynamically fuse the two types of features based on the average point distance and spatial range of the point cloud, and perform smoothing processing. Step 3, use the fused features to train a KNN classification model, establish the mapping relationship between the features and the target categories, and save the model parameters. Step 4, after the test point cloud is processed by steps 1 and 2, input the model parameters, output the highest confidence prediction result, and output it in the log.

[0051] Example 3 The point cloud recognition method based on dynamic feature fusion and full-process dynamic parameter adjustment comprises the following steps: Step 1: Collect the original point cloud of the industrial part and remove invalid points, combine the CAD model point cloud to construct a training set, calculate the average point distance based on the k-d tree, dynamically adjust the size of the voxel grid leaf, and generate a standardized point cloud; In step 1, the specific process of generating a standardized point cloud is as follows: Step 1.1, set the initial voxel grid leaf size to 1.0 mm, set the initial leaf size to 1.5 mm when the point cloud size is greater than 50k points, and directly output the standardized point cloud after the first sampling if the point cloud size is between 5k-30k points; Step 1.2, if the original point cloud size is greater than 30k points, calculate the average point distance based on the k-d tree, adjust the leaf size through an adjustment factor, and perform downsampling to generate a standardized point cloud; specifically: First, the average point distance is calculated as follows: (1) In equation (1), is the average point distance of the original point cloud, is the distance from point to its nearest neighbor, is the total number of valid point clouds after removing invalid points; Second, the adjustment factor is calculated as follows: (2) In equation (2), is the adjustment factor, limited to the range of 0.7-1.7, is the adjustment intensity coefficient, with a value of 0.5-1.0 according to the proportion of invalid points; Then, the leaf size is adjusted through iteration, and the iterative adjustment formula for the voxel grid leaf size is as follows: (3) In equation (3), is the current leaf size, with a unit of mm, limited to the range of 0.5mm-3mm, is the adjustment factor; is the new leaf size generated after iteration, when the point cloud size is not in the range of 5k-30k points, the iteration calculation is started, and the iteration is terminated when the point cloud size meets the requirements, according to the generated new leaf size, after downsampling, the iteration ends when the point cloud size meets the requirements of 5k-30k points, and the standardized point cloud after downsampling is output; Step 1.3, if the original point cloud size is less than 5k points, supplement the point cloud density by upsampling, the specific steps are: construct a k-d tree to calculate the nearest neighbor distance of each point, identify the area with a point spacing greater than 3mm as a sparse area, mark the points in this area as center points, for each center point, select its 10 nearest neighbor points, calculate the weight based on the Gaussian function, and insert new points at the weight center coordinates according to the weight to increase the density of the point cloud. After insertion, if the point cloud size is not less than 5k points, output the standardized point cloud directly; otherwise, decrease the point spacing threshold by 0.1mm gradient, re-mark the center points and repeat the insertion process until the point cloud size is not less than 5k points, and save it as a standardized point cloud.

[0052] Step 2, calculate the point cloud normal vector, extract the CVFH feature descriptor and the SHOT feature descriptor, dynamically fuse the two types of features based on the average point spacing and spatial range of the point cloud, and perform smoothing processing; Step 3, use the fused features to train a KNN classification model, establish the mapping relationship between the features and the target categories, and save the model parameters; Step 4, after the test point cloud is processed by step 1 and step 2, input the model parameters, output the highest confidence prediction result and output it in the log.

[0053] Example 4 The point cloud recognition method based on dynamic feature fusion and full-process dynamic parameter adjustment has the following steps: Step 1: Collect the original point cloud of industrial parts and remove invalid points, combine with the CAD model point cloud to construct the training set, calculate the average point spacing based on the k-d tree, dynamically adjust the size of the voxel grid leaf, and generate a standardized point cloud; In step 1, the specific process of generating a standardized point cloud is as follows: Step 1.1, set the initial voxel grid leaf size to 1.0mm, set the initial leaf size to 1.5mm when the point cloud size is greater than 50k points, and directly output the standardized point cloud after the first sampling if the point cloud size is between 5k-30k points. Step 1.2, if the original point cloud size is greater than 30k points, calculate the average point spacing based on the k-d tree, adjust the leaf size by adjusting the factor, and downsample to generate a standardized point cloud; the specific process is as follows: First, the average point spacing is calculated as follows: (1) In formula (1), is the average point spacing of the original point cloud, is the distance from the point to its nearest neighbor point, is the total number of valid points after removing invalid points; Second, the adjustment factor is calculated as follows: (2) In formula (2), is an adjustment factor, and the range of the adjustment factor is limited to 0.7-1.7, is an adjustment intensity coefficient, According to the proportion of the invalid point, the value is 0.5-1.0; Then, the leaf size is adjusted by iteration, and the iteration adjustment formula of the voxel grid leaf size is: (3) In formula (3), is the current leaf size, with the unit of mm, and the range is limited to 0.5mm-3mm, is an adjustment factor; is the new leaf size generated after iteration, when the point cloud scale is not in the range of 5k points-30k points, the iteration calculation is started, and the iteration is terminated until the point cloud scale meets the requirements, according to the generated new leaf size, after downsampling, the iteration is ended when the point cloud scale meets the requirements of 5k-30k points, and the standardized point cloud after downsampling is output; Step 1.3, if the original point cloud scale is less than 5k points, the point cloud density is supplemented by upsampling, and the specific steps are as follows: a k-d tree is constructed to calculate the nearest neighbor distance of each point, the area with a point spacing greater than 3mm is identified as a sparse area, and the points in the area are marked as center points, for each center point, 10 nearest neighbor points are selected, the weight is calculated based on the Gaussian function, and the weight center coordinates of the nearest neighbor points are calculated according to the weight, and then new points are inserted at the weight center coordinates to increase the density of the point cloud, after insertion, if the point cloud scale is not less than 5k points, the standardized point cloud is directly output; otherwise, the point spacing threshold is reduced by 0.1mm gradient, the center points are re-marked, and the insertion process is repeated until the point cloud scale is not less than 5k points, and the iteration is terminated, and the standardized point cloud is saved.

[0054] Step 2, calculate the point cloud normal vector, extract the CVFH feature descriptor and the SHOT feature descriptor, dynamically fuse the two types of features based on the average point spacing and the space range of the point cloud, and perform smoothing processing; The CVFH feature descriptor in step 2 is 308-dimensional, and the SHOT feature descriptor is 352-dimensional, after normalization processing of the extracted feature descriptor, a new 660-dimensional CVFH-SHOT feature descriptor is obtained by chain splicing; the specific process is as follows: Step 2.1, calculate the normalized point cloud normal vector based on k-d tree and radius search; first, take 5 times the average point spacing as the initial search radius, limit its range to 10-30 mm, if the search radius is less than 10 mm, default to 10 mm, if the search radius is greater than 30 mm, default to 30 mm, then, through k-d tree acceleration radius search, get the neighbor points of each point, based on principal component analysis to fit the local plane, calculate the covariance matrix of the local area of the point cloud, the normal vector is the eigenvector corresponding to the smallest eigenvalue of the covariance matrix; Step 2.2, use the CVFHEstimation algorithm provided by the PCL library to extract 308-dimensional CVFH feature descriptor from the normalized point cloud, specifically: calculate the normal and curvature information of the normalized original point cloud, segment the point cloud based on region growing algorithm, use the smoothness of the normal direction and the curvature threshold, the angle difference between the normal directions is less than 15° and the curvature value is less than 0.05 as the judgment basis of similar geometric properties, the point cloud is divided into multiple regions; Calculate the angle distribution of the view direction and the normal direction and the shape distribution of each region, construct a multi-dimensional CVFH feature histogram, generate a 308-dimensional global CVFH feature descriptor by k-means clustering and concatenating the feature histograms of all regions in order; Step 2.3, when using SHOTEstimationOMP algorithm to extract 352-dimensional SHOT feature descriptor from the normalized point cloud, first select key points from the normalized original point cloud, adopt adaptive strategy based on curvature: select the top 30% of points with curvature value as key points, exclude edge noise points, edge noise points are defined as points with a distance to the point cloud boundary less than 2 mm, regions with an average point spacing greater than 3 mm in the normalized point cloud, increase the number of key points to the top 40% of curvature, define a spherical support region centered on each key point, divide the support region into 32 sub-regions based on the local reference frame (LRF), and generate an 11-dimensional direction histogram in each sub-region by counting the angle distribution of the normal vector of the points in the sub-region with the LRF axis, concatenate the histograms of all sub-regions in order to generate a 352-dimensional SHOT feature descriptor. Step 2.4, generate 660-dimensional fusion features and smooth the features.

[0055] Step 2.4 is to calculate the CVFH and SHOT fusion weight dynamically according to the average point spacing of the point cloud and the spatial range , generate 660-dimensional fusion features, the specific process is as follows: Step 2.4.1, the spatial range is obtained by comparing the coordinates of each point in the normalized point cloud, updating the minimum point , , ) and the maximum point , , ), the maximum difference of the three-dimensional range is calculated as max( , , ), is the maximum range of the point cloud in the x, y, and z dimensions, with units of mm; Step 2.4.2, calculate the initial weight of CVFH based on the average point distance, the specific calculation method is as follows: If mm, then: (4) If mm, then: (5) In formula (4) and (5), is the average point distance of the normalized point cloud, the calculation method is referred to formula (1), 1.5 mm is the target point distance, 0.8 is the maximum weight of CVFH, and 0.3 is the adjustment range, is the CVFH weight; Step 2.4.3, based on the fine-tuning weight of the spatial range, further optimize the distribution of CVFH and SHOT weight, the specific implementation process is as follows: Calculate the range factor : (6) In formula (6), is the range factor; is the maximum range of the point cloud in the x, y, and z dimensions, with units of mm; 100 and 150 are normalization parameters; the clamp function limits to [-0.1, 0.1].

[0056] Adjust the CVFH weight: (7) SHOT weight: (8) In formula (8), is the weight of SHOT, which is limited by the clamp function to [0.2, 0.8]; Step 2.4.4, the normalized 308-dimensional CVFH feature descriptor and the 352-dimensional SHOT feature descriptor are spliced by weighting, specifically, the CVFH feature is normalized by maximum value, each component of the 308-dimensional histogram is divided by the maximum value, scaled to the range of [0, 1]; the SHOT feature is first averaged over all valid descriptors, then normalized by maximum value to scale the result to the range of [0, 1]; then the normalized CVFH and SHOT features are spliced in order in series, the first 308 dimensions are CVFH features, and the last 352 dimensions are SHOT features, forming an initial 660-dimensional feature. Based on the dynamic weight calculated by formula (4)-(8), the spliced feature is weighted and fused to generate the CVFH-SHOT 660 dynamic weighted feature.

[0057] The formula for smoothing in step 2.4 is as follows: (9) In formula (9), is the feature dimension (value is 0 to 659), the boundary condition is filled with mirror image; smoothing is completed by single iteration to suppress noise and enhance feature consistency.

[0058] Step 3, use the fused feature to train a KNN classification model, establish the mapping relationship between the feature and the target category and save the model parameters; Step 4, after the test point cloud is processed by step 1 and step 2, input the model parameters, output the prediction result with the highest confidence and output in the log.

[0059] Example 5 The point cloud recognition method based on dynamic feature fusion and full-process dynamic parameter adjustment has the following steps: Step 1: Collect the original point cloud of industrial parts and remove invalid points, combine with the CAD model point cloud to construct the training set, calculate the average point distance based on k-d tree, dynamically adjust the size of the voxel grid leaf, and generate standardized point cloud; In step 1, the specific process of generating standardized point cloud is as follows: Step 1.1, set the initial voxel grid leaf size to 1.0 mm, set the initial leaf size to 1.5 mm when the point cloud size is greater than 50k points, and directly output the standardized point cloud if the point cloud size is between 5k-30k points after the first sampling; Step 1.2, if the original point cloud size is greater than 30k points, calculate the average point distance based on k-d tree, adjust the leaf size by iteration through adjustment factor, and downsample to generate standardized point cloud; specifically: First, the average point distance is calculated as follows: (1) In formula (1), the average point distance of the original point cloud, the distance to its nearest neighbor point, the total number of valid point clouds after removing invalid points; Secondly, the adjustment factor is calculated as follows: (2) In formula (2), is the adjustment factor, limiting the adjustment factor range to 0.7-1.7, is the adjustment intensity coefficient, According to the proportion of invalid points, the value is 0.5-1.0; Then, the leaf size is adjusted by iteration, and the iteration adjustment formula of the voxel grid leaf size is: (3) In formula (3), is the current leaf size, unit: mm, limiting its range to 0.5mm-3mm, is the adjustment factor; is the new leaf size generated after iteration, when the point cloud size is not in the range of 5k points-30k points, the iteration calculation is started, and the iteration is terminated when the point cloud size meets the requirements, and after downsampling according to the generated new leaf size, the iteration is ended when the point cloud size meets the requirements of 5k-30k points, and the standardized point cloud after downsampling is output; Step 1.3, if the original point cloud size is less than 5k points, the point cloud density is supplemented by upsampling, the specific steps are: constructing k-d tree to calculate the nearest neighbor distance of each point, identifying the region with a point distance greater than 3mm as a sparse region, marking the points in the region as center points, selecting 10 nearest neighbor points for each center point, calculating the weight based on the Gaussian function, and inserting new points at the weight center coordinates according to the weight to increase the density of the point cloud. After insertion, if the point cloud size is not less than 5k points, the standardized point cloud is directly output; otherwise, the point distance threshold is reduced by 0.1mm gradient, the center points are marked again and the insertion process is repeated until the point cloud size is not less than 5k points, and the iteration is terminated. Save as a standardized point cloud.

[0060] Step 2, calculate the point cloud normal vector, extract the CVFH feature descriptor and SHOT feature descriptor, dynamically fuse the two types of features based on the average point distance of the point cloud and the spatial range, and perform smoothing processing; The CVFH feature descriptor in step 2 is 308-dimensional, and the SHOT feature descriptor is 352-dimensional. After normalizing the extracted feature descriptors, a new 660-dimensional CVFH-SHOT feature descriptor is obtained by concatenating them in series; The specific process is as follows: ​Step 2.1, calculate the normalized point cloud normal vector based on k-d tree and radius search; first, take 5 times the average point spacing as the initial search radius, limit its range to 10-30 mm, if the search radius is less than 10 mm, default to 10 mm, if the search radius is greater than 30 mm, default to 30 mm, then, through k-d tree acceleration radius search, get the neighbor points of each point, based on principal component analysis to fit the local plane, calculate the covariance matrix of the point cloud local area, the normal vector is the eigenvector corresponding to the smallest eigenvalue of the covariance matrix; Step 2.2, use the CVFHEstimation algorithm provided by the PCL library to extract 308-dimensional CVFH feature descriptor from the normalized point cloud, specifically: calculate the normal and curvature information of the normalized original point cloud, segment the point cloud based on region growing algorithm, use the smoothness of the normal direction and the curvature threshold, the angle difference between the normal directions is less than 15° and the curvature value is less than 0.05 as the judgment basis of similar geometric properties, the point cloud is divided into multiple regions; Calculate the angle distribution of the view direction and the normal direction and the shape distribution of each region, construct a multi-dimensional CVFH feature histogram, generate a 308-dimensional global CVFH feature descriptor by k-means clustering and concatenating the feature histograms of all regions in order; Step 2.3, when using SHOTEstimationOMP algorithm to extract 352-dimensional SHOT feature descriptor from the normalized point cloud, first select key points from the normalized original point cloud, adopt adaptive strategy based on curvature: select the top 30% of points with curvature value as key points, exclude edge noise points, edge noise points are defined as points with a distance to the point cloud boundary less than 2 mm, regions with an average point spacing greater than 3 mm in the normalized point cloud, increase the number of key points to the top 40% of curvature, define a spherical support region centered on each key point, divide the support region into 32 sub-regions based on the local reference frame (LRF), and generate an 11-dimensional direction histogram in each sub-region by counting the angle distribution of the normal vector of the points in the sub-region with the LRF axis, concatenate the histograms of all sub-regions in order to generate a 352-dimensional SHOT feature descriptor. Step 2.4, generate 660-dimensional fusion features and smooth the features.

[0061] Step 2.4 is to calculate the CVFH and SHOT fusion weight dynamically according to the average point spacing of the point cloud and the spatial range , generate 660-dimensional fusion features, the specific process is as follows: Step 2.4.1, the spatial range is obtained by comparing the coordinates of each point in the normalized point cloud, updating the minimum point , , ) and the maximum point , , ), the maximum difference of the three-dimensional range is calculated as max( , , ), is the maximum range of the point cloud in the x, y, and z dimensions, with a unit of mm; Step 2.4.2, calculate the initial weight of CVFH based on the average point distance, the specific calculation method is as follows: If mm, then: (4) If mm, then: (5) In formula (4) and (5), is the average point distance of the normalized point cloud, the calculation method is referred to formula (1), 1.5 mm is the target point distance, 0.8 is the maximum weight of CVFH, and 0.3 is the adjustment range, is the weight of CVFH; Step 2.4.3, based on the fine-tuning weight of the spatial range, further optimize the distribution of CVFH and SHOT weight, the specific implementation process is as follows: Calculate the range factor : (6) In formula (6), is the range factor; is the maximum range of the point cloud in the x, y, and z dimensions, with a unit of mm; 100 and 150 are normalization parameters; the clamp function limits in [-0.1, 0.1].

[0062] Adjust the CVFH weight: (7) SHOT weight: (8) In formula (7) and (8), is the weight of SHOT, which is limited by the clamp function in [0.2, 0.8]; Step 2.4.4, the normalized 308-dimensional CVFH feature descriptor and the 352-dimensional SHOT feature descriptor are spliced by weighting, specifically, the CVFH feature is normalized by maximum value, each component of the 308-dimensional histogram is divided by the maximum value, scaled to the range of [0, 1]; the SHOT feature is first averaged over all valid descriptors, then normalized by maximum value to scale the result to the range of [0, 1]; then the normalized CVFH and SHOT features are spliced in order in series, the first 308 dimensions are CVFH features, and the last 352 dimensions are SHOT features, forming an initial 660-dimensional feature, based on the dynamic weights calculated by formulas (4)-(8), the spliced features are weighted and fused to generate CVFH-SHOT660 dynamic weighted features.

[0063] The formula for smoothing in step 2.4 is as follows: (9) In formula (9), is the feature dimension (with a value of 0 to 659), and the boundary condition is filled with mirror image; smoothing is completed by a single iteration to suppress noise and enhance feature consistency.

[0064] Step 3, use the fused features to train a KNN classification model, establish the mapping relationship between the features and the target categories and save the model parameters; In step 3, the KNN classification model is used for training, and the mapping relationship between the feature vector and the target category is established, specifically: Load the training data set: import the training set obtained in step 1, use the fused feature vector method generated in step 2 to convert it into a CVFH-SHOT660 fused feature descriptor training set, covering several categories of industrial parts, divided into simple geometric class, complex structure class, and small feature class; use the KNN classification model to train the feature descriptor training set, save the model parameters obtained by training to an XML file in a specified path; at the same time, based on the XML file, cross-validation is performed, the training set is used as the test set input, the cross-validation result is used to optimize the model parameters and verify the rationality.

[0065] Step 4, after the point cloud is processed by steps 1 and 2, input the model parameters, output the prediction result with the highest confidence and output it in the log.

[0066] Example 6 The point cloud recognition method based on dynamic feature fusion and full-process dynamic parameter adjustment has the following steps: Step 1: Collect the original point cloud of industrial parts and remove invalid points, combine with the CAD model point cloud to build a training set, calculate the average point distance based on k-d tree, dynamically adjust the size of the voxel grid leaf, and generate a standardized point cloud; In step 1, the specific process of generating a standardized point cloud is as follows: Step 1.1, set the initial voxel grid leaf size to 1.0 mm, set the initial leaf size to 1.5 mm when the point cloud size is greater than 50k points, and directly output the standardized point cloud after the first sampling if the point cloud size is between 5k-30k points. Step 1.2, if the original point cloud size is greater than 30k points, calculate the average point distance based on the k-d tree, adjust the leaf size through an adjustment factor, and perform downsampling to generate a standardized point cloud; specifically: First, the average point distance is calculated as follows: (1) In equation (1), is the average point distance of the original point cloud, is the distance from point to its nearest neighbor, is the total number of valid point clouds after removing invalid points; Second, the adjustment factor is calculated as follows: (2) In equation (2), is the adjustment factor, limited to the range of 0.7-1.7, is the adjustment intensity coefficient, with a value of 0.5-1.0 based on the proportion of invalid points; Then, the leaf size is adjusted through iteration, and the iterative adjustment formula for the voxel grid leaf size is as follows: (3) In equation (3), is the current leaf size, with a unit of mm, limited to the range of 0.5mm-3mm, is the adjustment factor; is the new leaf size generated after iteration, when the point cloud size is not in the range of 5k-30k points, the iterative calculation is started, and the iteration is terminated when the point cloud size meets the requirements, and the new leaf size is generated after downsampling, and the iteration is ended when the point cloud size meets the requirements, and the standardized point cloud after downsampling is output; Step 1.3, if the original point cloud size is less than 5k points, supplement the point cloud density by upsampling, the specific steps are: construct a k-d tree to calculate the nearest neighbor distance of each point, identify the area with a point spacing greater than 3mm as a sparse area, mark the points in the area as center points, for each center point, select its 10 nearest neighbors, calculate the weight based on the Gaussian function, and according to the weight, the nearest neighbor point weight center coordinate is inserted into the new point at the weight center coordinate to increase the density of the point cloud. After insertion, if the point cloud size is not less than 5k points, the standardized point cloud is directly output; otherwise, decrease the point spacing threshold by 0.1mm gradient, re-mark the center points and repeat the insertion process until the point cloud size is not less than 5k points, and then terminate the iteration, and save it as a standardized point cloud.

[0067] Step 2, calculate the point cloud normal vector, extract the CVFH feature descriptor and the SHOT feature descriptor, dynamically fuse the two types of features based on the average point spacing and spatial range of the point cloud and perform smoothing processing; The CVFH feature descriptor in step 2 is 308-dimensional, and the SHOT feature descriptor is 352-dimensional. After normalizing the extracted feature descriptors, a new 660-dimensional CVFH-SHOT feature descriptor is obtained by concatenating them in a chain. The specific process is as follows: Step 2.1, calculate the normal vector of the standardized point cloud based on k-d tree and radius search; first, take 5 times the average point spacing as the initial search radius, limit its range to 10mm-30mm, if the search radius is less than 10mm, default to 10mm, if the search radius is greater than 30mm, default to 30mm, then accelerate the radius search through k-d tree, get the neighboring points of each point, fit the local plane based on principal component analysis, calculate the covariance matrix of the local region of the point cloud, and the normal vector is the eigenvector corresponding to the smallest eigenvalue of the covariance matrix; Step 2.2, use the CVFHEstimation algorithm provided by the PCL library to extract the 308-dimensional CVFH feature descriptor from the standardized point cloud, which is: calculate the normal and curvature information of the standardized original point cloud, segment the point cloud based on the region growing algorithm, use the smoothness of the normal direction and the curvature threshold, the angle difference between the normal directions is less than 15° and the curvature value is less than 0.05 as the judgment basis of similar geometric properties, divide the point cloud into multiple regions; calculate the angle distribution and shape distribution of the view direction and the normal direction for each region, construct a multi-dimensional CVFH feature histogram, generate a 308-dimensional global CVFH feature descriptor by k-means clustering and concatenating the feature histograms of all regions in order. Step 2.3, when extracting 352-dimensional SHOT feature descriptors from the standardized point cloud using the SHOTEstimationOMP algorithm, first select key points from the standardized original point cloud, and use a curvature-based adaptive strategy: select the top 30% of points with curvature values as key points, exclude edge noise points defined as points with a distance to the point cloud boundary of less than 2 mm, and regions with an average point spacing of the standardized point cloud of more than 3 mm, increase the number of key points to the top 40% of curvature, define a spherical support region centered on each key point, divide the support region into 32 sub-regions based on the local reference frame (LRF), and generate an 11-dimensional direction histogram based on the distribution of the normal vector of the points in each sub-region with respect to the LRF axis., sequentially concatenate the histograms of all sub-regions to generate a 352-dimensional SHOT feature descriptor; Step 2.4, generate a 660-dimensional fusion feature and smooth the feature.

[0068] Step 2.4 is to dynamically calculate the CVFH and SHOT fusion weights according to the average point spacing of the point cloud and the spatial range , to generate a 660-dimensional fusion feature, the specific process is as follows: Step 2.4.1, the spatial range is obtained by comparing the coordinates of each point in the standardized point cloud, updating the minimum point , , ) and the maximum point , , ), and calculating the maximum difference of the three-dimensional range as max( , , ), is the maximum range of the point cloud in the x, y, z dimensions, with units of mm; Step 2.4.2, calculate the initial CVFH weight based on the average point spacing, the specific calculation method is as follows: If mm, then: (4) If mm, then: (5) In equations (4) and (5), is the average point spacing of the standardized point cloud, calculated as in equation (1), 1.5 mm is the target point spacing, 0.8 is the maximum CVFH weight, and 0.3 is the adjustment amplitude, is the CVFH weight; Step 2.4.3, based on the spatial range of fine-tuning weights, further optimizes the allocation of CVFH and SHOT weights, and the specific implementation process is as follows: Calculate the range factor : (6) In formula (6), is the range factor; is the maximum range of the point cloud in the x, y, z dimensions, with units of mm; 100 and 150 are normalization parameters; the clamp function limits to [-0.1, 0.1].

[0069] Adjust the CVFH weight: (7) SHOT weight: (8) In formulas (7) and (8), is the weight of SHOT, which is limited by the clamp function to [0.2, 0.8]; Step 2.4.4, the normalized 308-dimensional CVFH feature descriptor and the 352-dimensional SHOT feature descriptor are spliced by weighting, specifically, the CVFH feature is processed by maximum value normalization, and each component of the 308-dimensional histogram is divided by the maximum value to scale to the [0, 1] range; the SHOT feature is first averaged for all valid descriptors, and then normalized by the maximum value to scale the result to the [0, 1] range; then the normalized CVFH and SHOT features are sequentially chained and spliced, the first 308 dimensions are CVFH features, and the last 352 dimensions are SHOT features, forming an initial 660-dimensional feature. Based on the dynamic weights calculated by formulas (4)-(8), the spliced features are weighted and fused to generate a CVFH-SHOT 660 dynamic weighted feature.

[0070] The formula for smoothing in step 2.4 is as follows: (9) In formula (9), is the feature dimension (with a value of 0 to 659), and the boundary condition is filled with mirror image; smoothing is completed by a single iteration to suppress noise and enhance feature consistency.

[0071] Step 3, use the fused features to train a KNN classification model to establish the mapping relationship between the features and the target categories and save the model parameters; In step 3, the KNN classification model is used for training, and the mapping relationship between the feature vector and the target category is established, specifically: Load training dataset: import the training set obtained in step 1, use the fusion feature vector method generated in step 2 to convert it into a CVFH-SHOT660 fusion feature descriptor training set, covering several categories of industrial parts, divided into simple geometric class, complex structure class, small feature class; use KNN classification model to train the feature descriptor training set, save the trained model parameters to an XML file in the specified path; at the same time, based on the XML file, cross-validation is carried out, the training set is used as the test set input, the cross-validation result is used to optimize the model parameters and verify the rationality.

[0072] Step 4, after the test point cloud is processed by steps 1 and 2, input the model parameters, output the highest confidence prediction result and output in the log.

[0073] The test point cloud in step 4 is the original point cloud of the industrial part that does not participate in training. Perform steps 1 and 2 on the test point cloud to generate a test dynamic weighted fusion CVFH-SHOT660 feature. Compare and evaluate this feature with the KNN model parameter XML file trained in step 3. Determine the matching confidence by calculating the distance between the test point cloud feature and the feature of each part in the training set. Take the class corresponding to the highest confidence as the prediction result and output it in the form of a log.

Claims

1. A point cloud recognition method based on dynamic feature fusion and full-process dynamic parameter tuning, characterized in that, The steps are as follows: Step 1: Collect the original point cloud of the industrial parts and remove invalid points. Combine the point cloud of the CAD model to build a training set. Calculate the average point spacing based on the kd tree, dynamically adjust the leaf size of the voxel mesh, and generate a standardized point cloud. Step 2: Calculate the point cloud normal vector, extract the CVFH feature descriptor and SHOT feature descriptor, dynamically fuse the two types of features based on the average point spacing and spatial range of the point cloud, and perform smoothing processing. Step 3: Train the KNN classification model using fused features, establish the mapping relationship between features and target categories, and save the model parameters; Step 4: After processing the point cloud in Steps 1 and 2, input the model parameters, output the prediction result with the highest confidence, and output it in the log.

2. The point cloud recognition method based on dynamic feature fusion and full-process dynamic parameter tuning according to claim 1, characterized in that, In step 1, the specific process of generating the standardized point cloud is as follows: Step 1.1: Set the initial voxel mesh leaf size to 1.0 mm. When the point cloud size is greater than 50k points, set the initial leaf size to 1.5 mm. If the point cloud size is between 5k and 30k points after the first sampling, directly output the standardized point cloud. Step 1.2: If the original point cloud size is greater than 30k points, calculate the average point spacing based on the kd-tree, adjust the leaf size iteratively by adjusting the adjustment factor, and perform downsampling to generate a standardized point cloud; specifically: First, the average point spacing is calculated as follows: （1） In equation (1), The average point spacing of the original point cloud. For point Distance to its nearest neighbor This represents the total number of valid point clouds after removing invalid points. Secondly, the adjustment factor is calculated as follows: （2） In equation (2), As an adjustment factor, the adjustment factor range is limited to 0.7~1.

7. To adjust the strength coefficient, The percentage of invalid points is set between 0.5 and 1.

0. Then, by iteratively adjusting the leaf size, the formula for iteratively adjusting the voxel mesh leaf size is: （3） In equation (3), This refers to the current leaf size, in mm, with a range limited to 0.5mm to 3mm. For adjustment factors; To determine the size of the new leaf generated after iteration, if the point cloud size is not within the range of 5k-30k points, the iterative calculation is started and terminated until the point cloud size meets the requirements. After downsampling based on the size of the generated new leaf, the iteration ends when the point cloud size is between 5k-30k points, and the downsampled standardized point cloud is output. Step 1.3: If the original point cloud size is less than 5k points, the point cloud density is supplemented by upsampling. The specific steps are as follows: construct a kd-tree to calculate the nearest neighbor distance of each point, identify the region with a point spacing greater than 3mm as a sparse region, mark the points in the region as the center point, select 10 nearest neighbor points for each center point, calculate the weight based on the Gaussian function, and insert new points at the weight center coordinates to increase the density of the point cloud. If the point cloud size is not less than 5k points after insertion, directly output the standardized point cloud. Otherwise, the point spacing threshold is reduced by a gradient of 0.1mm, the center point is remarked, and the point interpolation process is repeated until the point cloud size is not less than 5k points, at which point the iteration is terminated and the point cloud is saved as a standardized point cloud.

3. The point cloud recognition method based on dynamic feature fusion and full-process dynamic parameter tuning according to claim 1, characterized in that, In step 2, the CVFH feature descriptor is 308-dimensional and the SHOT feature descriptor is 352-dimensional. After normalization, the extracted feature descriptors are chained together to form a new 660-dimensional CVFH-SHOT feature descriptor.

4. The point cloud recognition method based on dynamic feature fusion and full-process dynamic parameter tuning according to claim 3, characterized in that, The specific process of step 2 is as follows: Step 2.1: Calculate the normal vector of the standardized point cloud based on kd-tree and radius search. First, use 5 times the average point spacing as the initial search radius, limiting it to between 10mm and 30mm. If the search radius is less than 10mm, the default value is 10mm; if the search radius is greater than 30mm, the default value is 30mm. Then, accelerate the radius search using kd-tree to obtain the neighboring points of each point. Fit the local plane based on principal component analysis and calculate the covariance matrix of the local region of the point cloud. The normal vector is the eigenvector corresponding to the smallest eigenvalue of the covariance matrix. Step 2.2: Use the CVFHEstimation algorithm provided by the PCL library to extract a 308-dimensional CVFH feature descriptor from the standardized point cloud. Specifically, calculate the normal and curvature information of the standardized original point cloud, segment the point cloud based on the region growing algorithm, and use the smoothness of the normal direction and curvature threshold. The angle difference between normal directions is less than 15° and the curvature value is less than 0.05 as the criteria for similar geometric attributes to segment the point cloud into multiple regions. For each region, calculate the angle distribution and shape distribution between the view direction and the normal direction, construct a multi-dimensional CVFH feature histogram, and generate a 308-dimensional global CVFH feature descriptor by k-means clustering and sequential chain concatenation of the feature histograms of all regions. Step 2.3: When extracting 352-dimensional SHOT feature descriptors from the standardized point cloud using the SHOTEstimationOMP algorithm, key points are first selected from the standardized original point cloud. An adaptive curvature-based strategy is adopted: the top 30% of the points with the highest curvature values ​​are selected as key points. The non-edge points are defined as areas that are not less than 2 mm from the point cloud boundary and whose average point spacing in the standardized point cloud is greater than 3 mm. The number of key points is increased to the top 40% of the curvature. A spherical support region is defined with each key point as the center. The support region is divided into 32 sub-regions based on the Local Reference Frame (LRF). The angle distribution between the normal vector of the point and the LRF axis is statistically analyzed in each sub-region to generate an 11-dimensional orientation histogram. The histograms of all sub-regions are sequentially chained together to generate the 352-dimensional SHOT feature descriptor. Step 2.4: Generate 660-dimensional fused features and smooth the features.

5. The point cloud recognition method based on dynamic feature fusion and full-process dynamic parameter tuning according to claim 4, characterized in that, Step 2.4 is based on the average point spacing of the point cloud. and spatial range The CVFH and SHOT fusion weights are dynamically calculated to generate 660-dimensional fusion features. The specific process is as follows: Step 2.4.1, obtaining the spatial range involves comparing the coordinates of each point in the standardized point cloud and updating the minimum point ( , , ) and maximum point ( , , ), calculate the maximum difference in the three-dimensional range. max( , , ), This represents the maximum extent of the point cloud in the x, y, and z dimensions, in mm. Step 2.4.2: Calculate the initial weights of CVFH based on the average point spacing. The specific calculation method is as follows: like mm, then: （4） like mm, then: （5） In equations (4) and (5), The average point spacing of the standardized point cloud is calculated using formula (1), where 1.5 mm is the target point spacing, 0.8 is the maximum weight of CVFH, and 0.3 is the adjustment range. For CVFH weights; Step 2.4.3: Based on the spatial range, fine-tune the weights to further optimize the allocation of CVFH and SHOT weights. The specific implementation process is as follows: Calculate the range factor : （6） In equation (6), Range factor; The maximum extent of the point cloud in the x, y, z dimensions, in mm; 100 and 150 are normalization parameters; the clamp function limits... In [-0.1, 0.1]; Adjust CVFH weights: （7） SHOT weight: （8） In equations (7) and (8), The weights of SHOT are limited by the clamp function. In [0.2, 0.8]; Step 2.4.4: The normalized 308-dimensional CVFH feature descriptor and the 352-dimensional SHOT feature descriptor are weighted and concatenated. Specifically, the CVFH feature is normalized by the maximum value, and each component of the 308-dimensional histogram is divided by the maximum value and scaled to the range of [0,1]. The SHOT feature is first averaged over all valid descriptors, and then normalized by the maximum value to scale the result to the range of [0,1]. Then, the normalized CVFH and SHOT features are concatenated in a chain in sequence, with the first 308 dimensions being the CVFH feature and the last 352 dimensions being the SHOT feature, forming an initial 660-dimensional feature. Based on the dynamic weights calculated by formulas (4)-(8), the concatenated features are weighted and fused to generate the CVFH-SHOT660 dynamic weighted feature.

6. The point cloud recognition method based on dynamic feature fusion and full-process dynamic parameter tuning according to claim 5, characterized in that, The formula for smoothing in step 2.4 is as follows: （9） In equation (9), The feature dimension is 0-659, and the boundary cases are filled with mirror images. Smoothing is completed in a single iteration to reduce noise and enhance feature consistency.

7. The point cloud recognition method based on dynamic feature fusion and full-process dynamic parameter tuning according to claim 6, characterized in that, Step 3 uses the KNN classification model for training, and simultaneously establishes a mapping relationship between feature vectors and target categories, specifically as follows: Load the training dataset: Import the training set obtained in step 1, and use the fusion feature vector method generated in step 2 to convert it into a CVFH-SHOT660 fusion feature descriptor training set, covering several categories of industrial parts, divided into simple geometric, complex structural, and small feature categories; train the feature descriptor training set using the KNN classification model, and save the trained model parameters to an XML file in a specified path; at the same time, perform cross-validation based on the XML file, using the training set as the test set input, and optimize the model parameters and verify their rationality through the cross-validation results.

8. The point cloud recognition method based on dynamic feature fusion and full-process dynamic parameter tuning according to claim 7, characterized in that, The test point cloud in step 4 is the original point cloud of industrial parts that were not used in training. Steps 1 and 2 are performed on the test point cloud to generate CVFH-SHOT660 features with dynamic weighted fusion for testing. This feature is compared and evaluated with the XML file of KNN model parameters trained in step 3. The matching confidence is determined by calculating the distance between the test point cloud features and the features of each type of part in the training set. The category corresponding to the highest confidence is taken as the prediction result and output in log form.

Images