Artificial Intelligence-Based 3D Point Cloud Data Analysis Method and System

Abstract

This invention belongs to the field of target classification technology, and particularly relates to a method and system for analyzing 3D point cloud data based on artificial intelligence. It includes steps such as data acquisition, data preprocessing, model building, model training, and model prediction. By acquiring raw point clouds and performing intelligent denoising and completion operations, the integrity and accuracy of the point cloud data are effectively improved. A deep model is constructed using an input layer, a region proposal layer, and classification and regression layers to accurately classify targets and predict 3D bounding boxes. The region proposal layer combines seed point feature extraction and a voting mechanism to generate candidate regions; the classification branch outputs class probabilities, and the regression branch predicts bounding box parameters. This modular system has advantages such as strong noise suppression, high completion accuracy, high target detection accuracy, and applicability to multiple scenarios, making it particularly suitable for the efficient automatic recognition of complex 3D structures.

Description

Technical Field

[0001] This invention belongs to the field of target classification technology, and in particular relates to a method and system for three-dimensional point cloud data analysis based on artificial intelligence. Background Technology

[0002] 3D point cloud data refers to a set of vectors in a three-dimensional coordinate system. 3D laser point cloud data is recorded in the form of "points," each containing "high-precision 3D coordinate information" and other attribute information about that point, such as color, reflectivity, and intensity. The main characteristics of 3D laser point cloud data are its high precision, high resolution, and high-dimensional geometric information, which can intuitively represent the shape, surface, and texture of objects in space.

[0003] 3D laser point cloud data records 3D coordinates and attributes such as color and reflectivity in the form of points. It features high precision, high resolution, and high-dimensional geometric information, and can intuitively present the shape, surface, and texture of objects, making it widely used in fields such as autonomous driving, industrial inspection, and artifact scanning. However, actual point cloud data often suffers from noise interference and data gaps. Furthermore, traditional analysis methods struggle to efficiently and accurately extract features when dealing with the disorder, sparsity, and complex geometric structures of point clouds, leading to insufficient accuracy in target detection and classification. For example, noise affects the accuracy of point cloud feature extraction, data gaps result in incomplete object structures, and traditional models suffer from low computational efficiency and limited ability to detect small objects and complex structures when processing large-scale point cloud data. Summary of the Invention

[0004] To address the technical problems existing in the background art described above, this invention proposes a three-dimensional point cloud data analysis method and system based on artificial intelligence.

[0005] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:

[0006] Data acquisition: Raw 3D point cloud data is collected using LiDAR scanning, and the raw 3D point cloud data is divided proportionally into data for training and data for prediction;

[0007] Data preprocessing: Preprocessing the data used for training, including data denoising and data completion steps;

[0008] The data denoising process involves intelligent denoising of the data used for training to obtain denoised point cloud data.

[0009] The data completion refers to performing data completion on the denoised point cloud data to obtain the point cloud data with complete data.

[0010] Model building: Used to build detection and classification models, including input layer, region proposal layer and classification and regression layers;

[0011] Input layer: Receives preprocessed point cloud data and formats the point cloud data into a unified format;

[0012] Region Proposal Layer: Used to generate candidate regions that may contain objects; the specific implementation includes seed point extraction and feature encoding, voting vector prediction, voting aggregation and candidate center generation, and candidate region feature enhancement;

[0013] Seed point extraction and feature encoding: A sampling algorithm is used to extract several seed points from the point cloud. A local feature representation is constructed for the neighborhood region of each seed point, and a multilayer perceptron is used to extract its local spatial geometric features.

[0014] Voting vector prediction: Based on the extracted seed point features, a three-dimensional offset vector is predicted through a neural network. The coordinates of the seed point are added to the offset vector to obtain the coordinates of the voting point, i.e., the candidate center point.

[0015] Voting aggregation and candidate center generation: Multiple voting points are aggregated into one or more candidate target centers using a density clustering algorithm. The aggregation results are then weighted and averaged, and the local features of each voting point are used as the feature representation of the center point.

[0016] Candidate region feature enhancement: A three-dimensional neighborhood voxel region is established for each generated candidate target center point. Features are extracted from the neighborhood points using convolution. The center features obtained by voting are fused with the neighborhood features to form a target region feature vector with strong expressive power.

[0017] Classification and Regression Layer: Detects the enhanced candidate region features, classifies each candidate region, and refines its bounding box;

[0018] The classification branch is used to extract features within the candidate region and outputs the class probability using a fully connected layer; the regression branch is used to predict the bounding box parameters.

[0019] Model training: Input the preprocessed point cloud data used for training into the constructed model for training;

[0020] Model prediction: Using a trained model, the point cloud data used for prediction is input to obtain the final predicted target classification result.

[0021] Preferably, intelligent denoising is performed during the data preprocessing to obtain denoised point cloud data. The operation is as follows:

[0022] First, a dynamic graph neural network (DGNN) is used to construct the neighborhood topology, initialize the point cloud nodes, and calculate the initial neighborhood with each point as the center. Then, the dynamic graph neural network model is used to learn the dynamic edge weights between nodes by taking the point cloud coordinates as input and outputting the updated neighborhood topology.

[0023] Then, global and local noise reduction are performed.

[0024] The dual paths at the global and local levels are fused, and the least squares method is used to optimize the smoothness of the point cloud surface, outputting denoised point cloud data.

[0025] Preferably, the update function for the dynamic edge weights is: ,in Represents the coordinates of points i and j. Let the curvatures be represented by i and j. To control the weight of distance.

[0026] Preferably, the global-level denoising includes global feature extraction, global coarse segmentation, and multi-filter scheduling;

[0027] Global feature extraction: Based on the DGNN neighborhood, PCA is used to fit the local plane, calculate the normal, and calculate the curvature using the local point cloud height difference; then the geometric moments are calculated, including the zeroth moment and the second moment;

[0028] Global coarse segmentation: Set a threshold based on normal and curvature features to segment the point cloud, and set different filtering strategies based on the threshold segmentation.

[0029] Multi-filter scheduling: Anisotropic diffusion filtering is applied to non-flat regions, preserving gradients along the normal direction, smoothing noise along the tangent direction, and iteratively updating point coordinates; statistical filtering is used for flat regions.

[0030] Preferably, the local-level denoising includes: local feature clustering, local fine classification, and adaptive repair;

[0031] Local feature clustering: Calculate the local density of each point, and then calculate the minimum distance from each point to higher density points;

[0032] Local fine-grained classification: Clustering is performed based on the two calculation results, and three categories of labels are output: outliers, edge points, and smooth points;

[0033] Adaptive inpainting: For outliers, a generative adversarial network is used to learn the normal distribution of the neighborhood and complete the inpainting; for edge points, anisotropic diffusion filtering is used to preserve details along the normal and denoise along the tangent; for smooth points, statistical filtering is used.

[0034] Preferably, the steps for performing data completion on the denoised point cloud data to obtain the completed point cloud data are as follows:

[0035] First, read the denoised point cloud, downsample it to a uniform density using a voxel grid, and mark the missing regions. If an RGB image is available, register it with the point cloud to extract the color and texture features of the neighborhood of the missing regions. Construct the bounding box of the missing regions and divide it into multiple sub-blocks.

[0036] For the point cloud within each sub-block, a neighborhood is constructed using a dynamic graph, and self-attention weights are calculated for each point: Where d is the feature dimension, and MLP is a multilayer perceptron; the output is a local block feature vector. ; All local block features The input is fed into a global Transformer layer, where inter-block dependencies are computed to obtain global features. By fusing global and local features, local block fusion features are obtained: ;

[0037] For each missing point in the region to be completed Collect known neighborhood points N within a range of r, and then calculate the neighborhood geometric features, including the normal n, curvature c, and neighborhood centroid. , Generate geometric constraint vectors And constrain the position of the completion point to be consistent with the normal;

[0038] Next, the local block fusion features are... and geometric constraint vector Generate the coordinates of the completion points using a multi-layer MLP. : For each missing sub-block, generate M completion points, where M is 1.5 times the number of missing points;

[0039] The generated completion points are deduplicated using a voxel mesh, retaining points with reasonable density; points are sorted by geometric constraint error, and those with a constraint error less than the set threshold are selected to form a preliminary completion point cloud;

[0040] The initially completed point cloud and the denoised point cloud are merged and input into the global Transformer to calculate the cross-regional structural loss:

[0041] Where K represents the number of point cloud data, For a global Transformer network, The point cloud after denoising. To initially complete the point cloud;

[0042] For the completion point Calculate its Laplace smoothing loss relative to known neighboring points:

[0043] ;

[0044] Construct a joint optimization objective function: ,in This is represented by the weights of the two.

[0045] The coordinates of the completed point cloud are iteratively adjusted using the gradient descent method to minimize the total loss. When the set number of iterations is reached, the iteration stops and the completed point cloud is output.

[0046] Preferably, the specific implementation of the regression branch of the classification and regression layers in the model construction for predicting bounding boxes is as follows:

[0047] The bounding box parameters are decoupled into multiple branches for prediction, including center coordinates, size and orientation angle, and different feature sub-channels are established for each branch.

[0048] Multi-level perceptrons with multiple parallel paths are used to model different parameter types separately; the original regression results of each sub-channel are output through their respective models.

[0049] In the output of the original regression results, the relative offset between the center point coordinates and the seed point coordinates is calculated, and then normalized before output. The calculation method is as follows: ,in The coordinates of the center point of the real target. Here, 'a' represents the coordinates of the corresponding seed point, where 'a' represents the x, y, and z coordinates, and 'y' is the normalization factor. For the size parameters, logarithmic encoding is used for output, and the calculation method is as follows: ,in, The actual location of the target. For reference dimension values, b represents the length, width, and height; for orientation angle parameters, the regression results are directly retained.

[0050] Preferably, the system is applicable to the method described in any one of claims 1-7, including a data acquisition module, a data preprocessing module, a model building module, a model training module, and a model prediction module;

[0051] The data acquisition module is used to acquire point cloud data;

[0052] The data preprocessing module is used to preprocess the acquired data to obtain preprocessed data;

[0053] The model building module is used to build detection and classification models;

[0054] The model training module is used to train the constructed detection and classification models.

[0055] The model prediction module is used to output the final target classification result.

[0056] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0057] 1. In the point cloud preprocessing stage, a dynamic graph neural network (DGNN) is introduced to construct a dynamic neighborhood topology. Combined with global feature extraction and local feature clustering strategies, the accuracy and robustness of point cloud denoising are effectively improved. At the same time, by using the Transformer structure to fuse local and global information and combining geometric constraints for point cloud completion, the structure and texture of missing regions can be better restored, achieving high-quality point cloud reconstruction.

[0058] 2. In the candidate region extraction process, a multi-scale seed point extraction and a three-dimensional offset voting mechanism are used to generate candidate centers. Density clustering and voxel feature enhancement strategies are employed to construct a more expressive target region feature vector. This method improves the recognition ability for small and complex structural targets, solving the problems of high false negative and high false positive rates in sparse scenes encountered by traditional methods.

[0059] 3. The classification and regression layers employ a feature decoupling modeling approach, dividing the bounding box parameters into multiple sub-channels for position, size, and orientation, and establishing independent regression models for each. Simultaneously, the classification branch incorporates a multilayer perceptron and a Softmax structure to output class probabilities, enhancing detection stability. The regression branch combines normalization and logarithmic encoding strategies, effectively improving the bounding box regression accuracy and the detection accuracy of multi-class targets. Attached Figure Description

[0060] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0061] Figure 1 This is an overall flowchart of a method and system for analyzing 3D point cloud data based on artificial intelligence.

[0062] Figure 2 This is a flowchart of the data preprocessing and noise reduction part of a method and system for analyzing 3D point cloud data based on artificial intelligence. Detailed Implementation

[0063] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0064] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification.

[0065] Example 1: Using a drone equipped with a lidar or laser scanner, and following the technical specifications for airborne lidar measurement of overhead transmission lines, operational requirements such as the width, point cloud density, point cloud plane, and elevation accuracy of the line channel scan are set. Correspondingly, a laser point cloud data acquisition task for the overhead transmission line channel and equipment is generated, and the data is processed to achieve automatic classification of components such as overhead line towers, conductors, ground wires, insulators, jumpers, ground, and trees. Existing 3D point cloud processing methods suffer from low processing efficiency and are easily affected by acquisition errors and environmental interference when dealing with high-dimensional sparse data, especially in the identification of complex structures and the detection of small targets, where accuracy is significantly insufficient. Therefore, this invention proposes a 3D point cloud data analysis method and system based on artificial intelligence. The overall implementation process is as follows: Figure 1 As shown.

[0066] First, to ensure data quality and scalability, a high-precision lidar acquisition device is used to perform a three-dimensional point cloud scan of the target area. The acquired raw data is divided into training data and prediction data according to the proportion, which are used for the training and prediction stages of the model, respectively.

[0067] Then, to reduce data interference and complete missing data information, the preprocessing steps include data denoising and data completion steps.

[0068] The data denoising process involves intelligently denoising the data used for training to obtain denoised point cloud data. Specifically, the data denoising steps are as follows: Figure 2 As shown, a dynamic graph neural network (DGNN) is first used to construct the neighborhood topology, initialize the point cloud nodes, and calculate the initial neighborhood with each point as the center. The dynamic graph neural network model is then used, where the update function for the dynamic edge weights is: ,in Represents the coordinates of points i and j. Let the curvatures be represented by i and j. To control the distance weights, the dynamic edge weights between nodes are learned by inputting point cloud coordinates, and the updated neighborhood topology is output.

[0069] Then, global and local denoising are performed. The global denoising includes global feature extraction, global coarse segmentation, and multi-filter scheduling. Global feature extraction: Based on the DGNN neighborhood, PCA is used to fit the local plane, calculate the normal, and calculate the curvature using the local point cloud height difference. Then, geometric moments, including zero-order moments and second-order moments, are calculated. Global coarse segmentation: A threshold is set according to the normal and curvature features to perform point cloud segmentation, and different filtering strategies are set according to the threshold segmentation. Multi-filter scheduling: Anisotropic diffusion filtering is performed on non-flat regions, the gradient is preserved along the normal direction, noise is smoothed along the tangent direction, and the point coordinates are iteratively updated. For flat regions, statistical filtering is used. Specifically, firstly, in the global feature extraction stage, based on the neighborhood relationships constructed by the Dynamic Graph Neural Network (DGNN), the local neighborhood of each point is analyzed. Principal Component Analysis (PCA) is used to fit the local point cloud, estimate the normal direction, and calculate curvature information by combining local height differences. Furthermore, geometric moment features, including the zeroth moment (representing point density) and the second moment (reflecting the discreteness of the point cloud distribution), are further extracted to form a complete global geometric attribute description. Secondly, in the global coarse segmentation stage, an adaptive threshold is set based on the extracted normal angles and curvature magnitudes to perform preliminary classification and segmentation of the point cloud, dividing it into different structural types such as flat and non-flat regions. Finally, in the multi-filter scheduling stage, anisotropic diffusion filtering technology is introduced for non-flat regions: edge gradient information is preserved in the normal direction, noise is smoothed in the tangent direction, and the coordinate values ​​of points are iteratively updated to improve continuity. For flat regions, statistical filters (such as median filtering or radius filtering) are used to remove isolated points and error interference, maintaining a smooth and consistent overall structure. This processing mechanism effectively achieves high-precision noise suppression and global structure preservation of point cloud data in complex scenarios.

[0070] The local-level denoising includes: local feature clustering, local fine-tuning, and adaptive repair. Local feature clustering calculates the local density of each point and then calculates the minimum distance from each point to higher-density points. Local fine-tuning clusters the points based on the two calculations, outputting three categories: outliers, edge points, and smooth points. Adaptive repair involves using a generative adversarial network to learn the normal distribution of the neighborhood for outliers and then completing the repair. For edge points, anisotropic diffusion filtering is used to preserve details along the normal and denoise along the tangent direction. For smooth points, statistical filtering is used. Specifically, firstly, in the local feature clustering stage, the spatial relationship between a point and its neighborhood is constructed, the local density of each point is calculated, and its nearest distance to all high-density points is further solved to obtain its distribution state in the local structure. Subsequently, in the local fine-tuning stage, based on the feature space constructed using local density and distance dimensions, density peak clustering is performed on the points, automatically classifying them into three categories: outliers, edge points, and smooth points, providing a foundation for subsequent targeted denoising strategies. Finally, in the adaptive repair stage, different processing methods are adopted according to the classification labels of the points: for outliers, a generative adversarial network (GAN) model is introduced to reconstruct and complete the outliers by learning the distribution characteristics of normal neighboring points, thereby enhancing data continuity and consistency; for edge points, anisotropic diffusion filtering technology is used to preserve structural details along the normal direction and smooth interference noise along the tangent direction; for normal points in the smoothed region, statistical filters such as radius filtering or mean filtering are used to remove isolated outliers. This multi-strategy adaptive mechanism makes local denoising more flexible and accurate, improving the overall quality of point cloud data while preserving key structural features.

[0071] The dual paths at the global and local levels are fused, and the least squares method is used to optimize the smoothness of the point cloud surface, outputting denoised point cloud data.

[0072] Then, the denoised point cloud data is augmented to obtain the augmented point cloud data. Specifically, the denoised point cloud is first read, downsampled to a uniform density using a voxel grid, and missing regions are marked. If an RGB image is available, it is registered with the point cloud to extract the color and texture features of the neighborhood of the missing regions. A bounding box for the missing regions is constructed and divided into multiple sub-blocks. For the point cloud within each sub-block, a neighborhood is constructed using a dynamic graph, and self-attention weights are calculated for each point. Where d is the feature dimension, and MLP is a multilayer perceptron; the output is a local block feature vector. ; All local block features The input is fed into a global Transformer layer, where inter-block dependencies are computed to obtain global features. By fusing global and local features, local block fusion features are obtained: For each missing point in the region that needs to be filled in. Collect known neighborhood points N within a range of r, and then calculate the neighborhood geometric features, including the normal n, curvature c, and neighborhood centroid. , Generate geometric constraint vectors It also constrains the position and normal of the completion points to be consistent; then it merges the local block features. and geometric constraint vector Generate the coordinates of the completion points using a multi-layer MLP. : For each missing sub-block, generate M completion points, where M is 1.5 times the number of missing points; use a voxel mesh to deduplicate the generated completion points, retaining points with reasonable density; sort by geometric constraint error, and select points with an error less than a set constraint error threshold to form a preliminary completed point cloud; merge the preliminary completed point cloud with the denoised point cloud, input it into the global Transformer, and calculate the cross-regional structural loss: Where K represents the number of point cloud data, For a global Transformer network, The point cloud after denoising. To initially complete the point cloud; for the completed points Calculate its Laplace smoothing loss relative to known neighboring points: Construct a joint optimization objective function: ,in The coordinates of the completed point cloud are expressed as the weights of the two values. The coordinates of the completed point cloud are iteratively adjusted using the gradient descent method to minimize the total loss. When the set number of iterations is reached, the iteration stops and the completed point cloud is output.

[0073] To improve target recognition capabilities and bounding box accuracy, the model construction design consists of three parts: an input layer, a region proposal layer, and a classification and regression layer. The input layer receives preprocessed point cloud data and formats the point cloud data into a unified format.

[0074] Region Proposal Layer: Used to generate candidate regions that may contain objects; the specific implementation includes seed point extraction and feature encoding, voting vector prediction, voting aggregation and candidate center generation, and candidate region feature enhancement; specifically, seed point extraction and feature encoding: several seed points are extracted from the point cloud using a sampling algorithm, a local feature representation is constructed for the neighborhood region of each seed point, and its local spatial geometric features are extracted using a multilayer perceptron; voting vector prediction: based on the extracted seed point features, a three-dimensional offset vector is predicted through a neural network, and the coordinates of the seed point are added to the offset vector to obtain the coordinates of the voting point, i.e., the candidate center point; voting aggregation and candidate center generation: multiple voting points are aggregated into one or more candidate target centers using a density clustering algorithm, the aggregation result is weighted and averaged, and the local features of each voting point are combined as the feature representation of the center point; candidate region feature enhancement: a three-dimensional neighborhood voxel region is established for each generated candidate target center point, features are extracted using convolutional neighborhood point set, and the center features obtained from voting are fused with the neighborhood features to form a target region feature vector with strong expressive power.

[0075] Classification and Regression Layers: These layers detect the enhanced candidate region features, classifying each region and refining its bounding box. The classification branch extracts features from the candidate regions and outputs class probabilities using a fully connected layer. The regression branch predicts the bounding box parameters. Specifically, the classification branch determines the category of the candidate region. This branch employs a multi-layer fully connected neural network (MLP) structure. After non-linear mapping using an activation function, it connects to a Softmax layer to output the class probability distribution, ultimately selecting the category with the highest probability as the classification result. This process effectively mines the semantic information of the candidate regions, achieving accurate classification. The regression branch accurately predicts the target bounding box parameters, including center coordinates (x, y, z), dimensions (length, width, height), and orientation angle. To improve prediction accuracy, the bounding box parameters are decoupled into multiple branches for prediction, including center coordinates, size, and orientation angle, and different feature sub-channels are established for each. A multi-path parallel multilayer perceptron is used to model different parameter types separately. The original regression results for each sub-channel are output through their respective models. In the output original regression results, the relative offset between the center point coordinates and the seed point coordinates is calculated, normalized, and then output. The calculation method is as follows: ,in The coordinates of the center point of the real target. Here, 'a' represents the coordinates of the corresponding seed point, where 'a' represents the x, y, and z coordinates, and 'y' is the normalization factor. For the size parameters, logarithmic encoding is used for output, and the calculation method is as follows: ,in, The actual location of the target. For reference dimensions, 'b' represents the length, width, and height. For the orientation angle parameter, the regression result is directly retained. During the training phase, cross-entropy loss and smoothed L1 loss are used to optimize both classification and regression, ensuring a balance between classification accuracy and localization precision.

[0076] After model building, the preprocessed point cloud data used for training is input into the built model for training; after the model is trained, the trained model is used to input the point cloud data used for prediction to obtain the final predicted target classification result.

[0077] Example 2 discloses an artificial intelligence-based 3D point cloud data analysis system, including a data acquisition module, a data preprocessing module, a model building module, a model training module, and a model prediction module;

[0078] The data acquisition module is used to acquire point cloud data by connecting to various point cloud sensing devices, such as lidar.

[0079] The data preprocessing module is used to preprocess the acquired data, obtaining preprocessed data that improves the quality and structural integrity of the original point cloud data, ensuring the accuracy of subsequent model input. This module includes two core functions: denoising and completion. The denoising part uses a dynamic graph neural network (DGNN) to dynamically construct neighborhoods, combining global features and local density analysis to identify and process noise points, edge points, and other outliers. The completion part predicts and restores missing regions based on the Transformer structure and geometric constraints, generating a highly consistent completed point cloud by fusing local block features and global contextual information. The final output point cloud data has advantages such as clear structure, complete information, and low error, providing high-quality input for subsequent model processing.

[0080] The model building module is used to construct the detection and classification model. This module is responsible for building the deep neural network structure for the entire point cloud intelligent analysis, mainly including an input layer, a region proposal layer, and a classification and regression layer. The input layer standardizes the preprocessed point cloud, such as coordinate normalization and uniform point sampling. The region proposal layer generates candidate regions that may contain targets based on seed point extraction and a 3D offset voting mechanism, and enhances their feature representation ability through clustering and voxel convolution. The classification and regression layer inputs these enhanced candidate regions into the classification network and the bounding box regression network, respectively, and outputs class labels and 3D boundary parameters. The overall model has high robustness and adaptability, and is suitable for complex 3D target detection and classification tasks in various scenarios.

[0081] The model training module is used to train the constructed detection and classification model. During the training process, the system inputs the preprocessed point cloud into the model and continuously optimizes the model parameters through mechanisms such as positive and negative sample comparison and error feedback.

[0082] The model prediction module is used to output the final target classification result. This module is responsible for applying the trained model to the analysis task of actual point cloud data. When new point cloud data is input, the module will sequentially perform operations such as feature extraction, region proposal, target classification, and bounding box prediction, and output the target category and 3D localization result for each candidate region.

[0083] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for analyzing three-dimensional point cloud data based on artificial intelligence, characterized by, The method comprises the following steps: Data acquisition: collect three-dimensional point cloud raw data by laser radar scanning, and divide the three-dimensional point cloud raw data into data for training and data for prediction according to a proportion; Data preprocessing: preprocessing the data for training, including data denoising and data completion steps; The data denoising: the data for training is denoised to obtain denoised point cloud data; wherein the denoising part adopts a dynamic graph neural network DGNN to dynamically construct a neighborhood, combines global features and local density analysis, and identifies and processes abnormal points; The data completion: the denoised point cloud data is completed to obtain the data completed point cloud data; wherein the completion part predicts and recovers the missing area based on the Transformer structure and geometric constraints, generates the completed point cloud by fusing the local block features and the global context information; Model construction: for constructing a detection and classification model, including an input layer, a region proposal layer and a classification and regression layer; The input layer: receiving the point cloud data after data preprocessing, and unifying the format of the point cloud; The region proposal layer: for generating candidate regions that may contain objects; the specific implementation includes seed point extraction and feature encoding, voting vector prediction, voting aggregation and candidate center generation, and candidate region feature enhancement; Seed point extraction and feature encoding: a sampling algorithm is used to extract a plurality of seed points from the point cloud, local feature representation is constructed for the neighborhood region of each seed point, and a multilayer perceptron is used to extract local spatial geometric features; Voting vector prediction: based on the extracted seed point features, a neural network is used to predict a three-dimensional offset vector, the seed point coordinates and the offset vector are added, and the voting point coordinates, i.e. the candidate center point, are obtained; Voting aggregation and candidate center generation: a density clustering algorithm is used to aggregate multiple voting points into one or more candidate target centers, and a weighted average processing is performed on the aggregation result, and the local features of each voting point are combined as the feature representation of the center point; Candidate region feature enhancement: a three-dimensional neighborhood voxel region is established for each generated candidate target center point, features are extracted from the convolution neighborhood points, the center features obtained by voting are fused with the neighborhood features, and a target region feature vector with strong expression ability is formed; The classification and regression layer: the enhanced candidate region features are detected, and the classification and bounding box refinement are performed on each candidate region; The classification branch is used to extract the features in the candidate region, and the class probability is output by a fully connected layer; the regression branch is used to predict the bounding box parameters; Model training: input the point cloud data for training after data preprocessing into the constructed model for training; Model prediction: input the point cloud data for prediction into the trained model to obtain the final predicted target classification result. 2.The AI-based three-dimensional point cloud data analysis method of claim 1, wherein, The intelligent denoising processing in the data preprocessing operation to obtain the denoised point cloud data is as follows: First, a dynamic graph neural network DGNN is used to construct a neighborhood topology, and the point cloud nodes are initialized, taking each point as the center to calculate the initial neighborhood; a dynamic graph neural network model is used to input the point cloud coordinates to learn the dynamic edge weight between nodes, and output the updated neighborhood topology; Then global level denoising and local level denoising are performed; The double paths of the global level and the local level are fused, and the least square method is used to optimize the point cloud surface smoothness, and the denoised point cloud data is output. 3.The AI-based three-dimensional point cloud data analysis method of claim 2, wherein, The update function of the dynamic edge weight is: wherein represents the point coordinates of i and j, represents the curvature of i and j, is the weight of the control distance. 4.The AI-based three-dimensional point cloud data analysis method of claim 2, wherein, The global level denoising includes global feature extraction, global rough segmentation and multi-filter scheduling; Global feature extraction: based on the DGNN neighborhood, the local plane is fitted using PCA, the normal is calculated, the curvature is calculated using the local point cloud height difference; then the geometric moments are calculated, including the zeroth moment and the second moment; Global rough segmentation: set threshold according to normal and curvature features, perform point cloud segmentation, set adaptive different filtering strategies according to threshold segmentation; Multi-filter scheduling: anisotropic diffusion filtering is performed on non-flat areas, the gradient is preserved along the normal direction, the noise is smoothed along the tangent direction, and the point coordinates are iteratively updated; For flat areas, statistical filtering is used for processing. 5.The AI-based three-dimensional point cloud data analysis method of claim 2, wherein, The local level denoising includes local feature clustering, local fine classification and adaptive repair; Local feature clustering: calculate the local density of each point, then calculate the minimum distance of each point to higher density points; Local fine classification: clustering is performed according to the two calculation results, outputting three types of labels: outlier points, edge points and smooth points; Adaptive repair: for outlier points, learn the normal distribution of the neighborhood using a generative adversarial network for completion and repair; for edge points, use anisotropic diffusion filtering to preserve details along the normal and denoise along the tangent direction; for smooth points, use statistical filtering for processing. 6.The AI-based three-dimensional point cloud data analysis method of claim 3, wherein, The implementation steps for data completion of the denoised point cloud data are as follows: First, read the denoised point cloud, downsample to a uniform density through voxel grid, mark the missing area, if there is an RGB image, align it with the point cloud to extract the color and texture features of the missing area neighborhood; construct a bounding box for the missing area and divide it into multiple sub-blocks; For each point cloud within each sub-block, a dynamic graph is used to construct neighborhoods, and self-attention weights are calculated for each point: where d is the feature dimension, and MLP is a multi-layer perceptron; an output local block feature vector is obtained All local block features are input into a global Transformer layer to calculate inter-block dependencies and obtain global features Global and local features are fused to obtain local block fusion features: ; for each missing point of the missing region , collect the known neighborhood points N within its range, then compute neighborhood geometric features including normal n, curvature c and neighborhood barycenter , ; Generating a geometry constraint vector and the position and normal consistency of the completion point The local block fusion features are then concatenated and a geometric constraint vector The coordinates of the completed points are generated by a multi-layer MLP : For each missing sub-block, M completed points are generated, where M is 1.5 times the number of missing points; Use the voxel grid to remove duplicate points in the generated completion points, and retain points with reasonable density; sort the points according to the geometric constraint error, and select points with an error less than the set constraint error threshold to form a preliminary completion point cloud; Merge the preliminary completion point cloud with the denoised point cloud and input it into the global Transformer to calculate the cross-region structure loss: wherein K represents the number of point cloud data, is a global Transformer network, is a denoised point cloud, is a preliminary completed point cloud; to the completion point compute its laplacian smoothing loss with the known points in its neighborhood: ； Construct the joint optimization objective function: wherein is the weight of both Iteratively adjust the coordinates of the completion point cloud using the gradient descent method to minimize the total loss, and stop iteration when the set number of iterations is reached to output the completed point cloud. 7.The artificial intelligence-based three-dimensional point cloud data analysis method of claim 1, wherein, The regression branch of the classification and regression layer in the model construction is specifically implemented as follows: Decouple the bounding box parameters into multiple branches to predict, including center coordinates, size and direction angle, and establish different feature sub-channels respectively; Use a multi-layer perceptron in parallel to model different parameter types; Output the original regression results of each sub-channel through the respective model; In the output of the original regression results, the relative offset between the center point coordinates and the seed point coordinates is calculated and normalized for output, and the calculation method is: wherein is the center point coordinates of the real target, is the corresponding seed point coordinates, a represents x, y, z three coordinates, and Y is a normalization factor; for the size parameter, a logarithmic encoding method is used for output, and the calculation method is: wherein, is the position of the real target, is the reference size value, b represents the length, width and height of the three positions; for the direction angle parameter, the regression result is directly retained.

8. An artificial intelligence-based three-dimensional point cloud data analysis system, characterized by, The system is applicable to the method of any one of claims 1-7, comprising a data acquisition module, a data preprocessing module, a model construction module, a model training module and a model prediction module; The data acquisition module is used to acquire point cloud data; The data preprocessing module is used to preprocess the acquired data to obtain preprocessed data; The model construction module is used to construct a detection and classification model; The model training module is configured to perform training operation on the constructed detection and classification model. The model prediction module is configured to output a final target classification result.

Images