Large-scale point cloud data feature extraction method and system based on artificial intelligence

Abstract

The invention provides a large-scale point cloud data feature extraction method and system based on artificial intelligence. The system comprises a feature recognition unit, a three-dimensional scanning unit, a simulation point cloud generation unit, a data fusion unit and an operation platform. The feature recognition unit recognizes a region with obvious features of the to-be-detected piece to obtain a feature region; the three-dimensional scanning unit obtains feature region point cloud data; the simulation point cloud generation unit generates simulation point cloud data according to the to-be-tested piece model; the data fusion unit carries out fusion and reconstruction to generate a three-dimensional point cloud model; and the operation platform displays the three-dimensional point cloud model. According to the method, scanning time and data storage requirements are effectively reduced, the three-dimensional reconstruction efficiency is improved, meanwhile, the problem of data missing in a traditional method is reduced, and the method is particularly suitable for complex samples or samples which are difficult to directly scan; an end-to-end solution is provided to realize rapid and accurate large-scale point cloud data feature extraction, the limitation of a traditional three-dimensional scanning method is overcome, and the method has a wide industrial application prospect.

Description

Technical Field

[0001] The present invention belongs to the technical field of three-dimensional measurement and reconstruction, and specifically, relates to a method and system for extracting features of large-scale point cloud data based on artificial intelligence, in particular, a method for quickly extracting features of large-scale point cloud data based on vision guidance of artificial intelligence and a three-dimensional point cloud reconstruction system based on feature recognition and simulated point cloud fusion. Background Art

[0002] With the development of industrial automation and intelligent manufacturing, 3D measurement technology has been widely used in product design, quality inspection, reverse engineering and other fields. This technological advancement has not only improved production efficiency, but also enhanced product accuracy and quality control. Traditional 3D scanning methods usually rely on comprehensive scanning of the entire object. However, in complex environments, this method often leads to problems such as long scanning time, large data volume, and difficult processing. In addition, due to the combined effects of factors such as the diversity of workpiece material properties, the complexity of surface conditions, and the variability of milling processes, traditional measurement technology is difficult to ensure high accuracy and stability of measurement results, which in turn leads to large deviations in measurement data and cannot meet the stringent requirements of high-precision measurement.

[0003] High-precision measurement of 3D scanning technology requires high-resolution images and complex algorithm processing, which may lead to problems such as large amount of calculation and poor real-time performance. In order to solve these problems, it is necessary to further study and improve related technologies, such as optimizing image acquisition equipment, improving image quality and stability, improving feature extraction algorithms, improving adaptability to complex shapes and textures, and studying new 3D reconstruction algorithms to achieve more efficient and accurate 3D measurement.

[0004] Patent document "Scanning Data Processing Method, Device, 3D Scanning System and Storage Medium" (CN119063632A) discloses filtering processing based on a three-dimensional bounding box, a first tree data structure and a second tree data structure, which achieves the fit between the filtering range and the shape of the workpiece, thereby improving the accuracy of filtering the redundant point cloud data outside the workpiece. However, it still requires a complete scan of the workpiece, which is inefficient.

[0005] Patent document "Three-coordinate measurement system and method based on three-dimensional vision guidance" (CN118936316A) discloses the use of artificial intelligence-based data processing technology to embed and encode the three-dimensional point cloud data of the workpiece, and to associate the embedded and encoded data locally and globally, so as to intelligently obtain the position data recognition result of the position data and posture data of the workpiece to be identified on the workbench based on the multi-scale fusion characteristics between the global and local three-dimensional point cloud data of the workpiece. However, this patent obtains the position data recognition result from the point cloud data to construct the coordinate system of the three-coordinate measurement system, rather than the reconstruction method of the point cloud data.

[0006] Therefore, there is an urgent need for a new method that can accurately reconstruct complete three-dimensional point cloud data while ensuring scanning efficiency, so as to meet the needs of high-precision and high-efficiency measurement in the field of intelligent manufacturing. Summary of the invention

[0007] In view of the defects in the prior art, the purpose of the present invention is to provide a large-scale point cloud data feature extraction method and system based on artificial intelligence.

[0008] A large-scale point cloud data feature extraction system based on artificial intelligence provided by the present invention includes: a feature recognition unit, a three-dimensional scanning unit, a simulated point cloud generation unit, a data fusion unit and an operation platform.

[0009] The feature recognition unit recognizes the region with obvious features of the test piece to obtain the feature region.

[0010] The three-dimensional scanning unit scans the feature area and obtains point cloud data of the feature area.

[0011] The simulation point cloud generation unit generates simulation point cloud data according to the model of the test piece.

[0012] The data fusion unit fuses the feature area point cloud data and the simulation point cloud data to reconstruct and generate a three-dimensional point cloud model.

[0013] The operating platform controls the operation of the feature recognition unit, the 3D scanning unit, the simulation point cloud generation unit and the data fusion unit according to the input instructions, and displays the 3D point cloud model.

[0014] Preferably, the feature recognition unit includes an image acquisition module and an image processing module.

[0015] The image acquisition module uses a high-resolution industrial camera to capture and collect the characteristic sample data set of the test piece in all directions to obtain a two-dimensional image of the test piece.

[0016] The omnidirectional shooting is to keep the collection distance the same as the collection distance in the actual application scene, shoot a rated number of two-dimensional images of the same features at different angles under different lighting conditions, and keep the number of two-dimensional images of each feature type the same.

[0017] The image processing module uses image recognition technology to pre-process the two-dimensional image;

[0018] Apply deep learning algorithms to extract features of the test piece from two-dimensional images;

[0019] Based on the features, the feature detection model based on YOLOv5 determines the key points and areas in the image;

[0020] Image annotation tools are used to mark the feature type, location and size range of key points and regions to obtain feature regions.

[0021] The preprocessing includes grayscale, histogram equalization, image scaling, normalization, image enhancement, denoising and contrast enhancement.

[0022] Preferably, the three-dimensional scanning unit comprises a binocular stereo vision non-contact measurement system.

[0023] The binocular stereo vision non-contact measurement system comprises a turntable 1, a baffle 2, a global tracking and positioning system 3, a six-degree-of-freedom industrial robot 4, a computer 5, and a local scanning measurement system 6.

[0024] The baffle 2 surrounds three sides of the binocular stereo vision non-contact measurement system.

[0025] The turntable 1 drives the test piece to rotate.

[0026] The six-degree-of-freedom industrial robot 4 controls the movement and positioning of the local scanning measurement system 6, performs a comprehensive and uniform scanning on the characteristic area of ​​the complex test piece, and obtains local point cloud data.

[0027] The global tracking and positioning system 3 stitches together different local areas through the posture conversion relationship when the turntable 1 rotates, obtains point cloud data of the feature area, and transmits the data fusion module and the operation platform through the computer 5.

[0028] Preferably, the three-dimensional scanning unit uses an automated path planning algorithm to set the motion path and limit of the scanning based on the geometric dimensions and measurement requirements of the object to be measured;

[0029] Scan the feature area, collect point cloud data of the feature area and monitor the scanning quality.

[0030] The simulation point cloud generation unit includes a computer-aided design model matching module.

[0031] According to the geometric information of the computer-aided design model matching module, point cloud data is generated, and the density and accuracy are adjusted to match the point cloud data of the feature area actually scanned to obtain simulated point cloud data.

[0032] The data fusion unit imports feature area point cloud data and simulation point cloud data, registers them using point cloud registration technology, fuses them to generate three-dimensional point cloud data, reconstructs the surface and performs post-processing to obtain a three-dimensional point cloud model.

[0033] Preferably, the simulated point cloud generation unit generates point cloud data using a voxel-based gridding algorithm, and adjusts density and accuracy through denoising, smoothing and normalization preprocessing to match the density and accuracy of feature area point cloud data.

[0034] The data fusion unit comprises:

[0035] Module M4.1: Use iterative closest point algorithm to spatially align feature area point cloud data and simulation point cloud data;

[0036] Module M4.2: Apply filtering algorithms to remove noise and outliers;

[0037] Module M4.3: Apply multi-view stereo matching technology to select the scanned point cloud data for the feature area point cloud data, and fuse them to generate three-dimensional point cloud data;

[0038] Module M4.4: Surface reconstruction of 3D point cloud data to generate smooth and continuous 3D surfaces;

[0039] Module M4.5: Perform redundancy removal, hole filling and smoothing post-processing to obtain a three-dimensional point cloud model.

[0040] According to a large-scale point cloud data feature extraction method based on artificial intelligence provided by the present invention, the large-scale point cloud data feature extraction system based on artificial intelligence is adopted, comprising:

[0041] Step S1: Identify the area with obvious features of the test piece and obtain the feature area;

[0042] Step S2: Scan the feature area to obtain point cloud data of the feature area;

[0043] Step S3: Generate simulated point cloud data according to the model of the test piece;

[0044] Step S4: Fuse the feature area point cloud data and the simulation point cloud data to reconstruct and generate a three-dimensional point cloud model.

[0045] Preferably, the step S1 comprises:

[0046] Step S1.1: Using a high-resolution industrial camera, capture a feature sample data set of the test piece in all directions to obtain a two-dimensional image of the test piece;

[0047] Step S1.2: Using artificial intelligence recognition technology, identify the region with obvious features in the two-dimensional image of the test piece to obtain the feature region.

[0048] The omnidirectional shooting is to keep the collection distance the same as the collection distance in the actual application scene, shoot a rated number of two-dimensional images of the same features at different angles under different lighting conditions, and keep the number of two-dimensional images of each feature type the same.

[0049] Preferably, the step S1.2 comprises:

[0050] Image preprocessing step: Use image recognition technology to preprocess the two-dimensional image.

[0051] Feature recognition step: Apply deep learning algorithms to extract and recognize features of the processed two-dimensional image to determine the feature area of ​​the test piece.

[0052] The preprocessing includes grayscale, histogram equalization, image scaling, normalization, image enhancement, denoising and contrast enhancement.

[0053] The feature extraction and recognition includes:

[0054] Step S1.2.1: Use a deep learning network to extract features of the test piece from the two-dimensional image;

[0055] Step S1.2.2: According to the features, the key points and regions in the image are determined based on the feature detection model of YOLOv5;

[0056] Step S1.2.3: Use image annotation tools to mark the feature type, position and size range of key points and regions to obtain feature regions.

[0057] Preferably, step S2 comprises:

[0058] Scanning path planning steps: Use the automated path planning algorithm to set the scanning motion path and limits based on the geometric dimensions and measurement requirements of the object to be measured.

[0059] Data collection steps: Use a structured light measurement device to scan the feature area, collect point cloud data of the feature area and monitor the scanning quality.

[0060] In step S3, point cloud data is generated according to the geometric information of the model of the workpiece to be tested, and the density and accuracy are adjusted to match the point cloud data of the feature area actually scanned to obtain simulated point cloud data.

[0061] In step S4, the feature area point cloud data and the simulation point cloud data are imported, registered using the point cloud registration technology, fused to generate three-dimensional point cloud data, and the surface is reconstructed and post-processed to obtain a three-dimensional point cloud model.

[0062] Preferably, the structured light measurement device is a binocular stereoscopic vision three-dimensional measurement system based on fringe projection, which controls movement and positioning through a six-degree-of-freedom industrial robot, performs a comprehensive and uniform scan of the feature area, collects scanning data to generate feature area point cloud data, and visualizes it through an operating platform.

[0063] In step S3, a voxel-based gridding algorithm is used to generate point cloud data, and the density and accuracy are adjusted through denoising, smoothing and normalization preprocessing to match the density and accuracy of the feature area point cloud data.

[0064] The step S4 includes:

[0065] Step S4.1: spatially aligning the feature area point cloud data and the simulation point cloud data using an iterative closest point algorithm;

[0066] Step S4.2: Apply filtering algorithm to remove noise and outliers;

[0067] Step S4.3: Apply multi-view stereo matching technology, select the scanned point cloud data for the feature area point cloud data, and fuse them to generate three-dimensional point cloud data;

[0068] Step S4.4: Surface reconstruction is performed on the three-dimensional point cloud data to generate a smooth and continuous three-dimensional surface;

[0069] Step S4.5: Perform post-processing such as removing redundancy, filling holes and smoothing to obtain a three-dimensional point cloud model.

[0070] Compared with the prior art, the present invention has the following beneficial effects:

[0071] 1. The present invention reduces the scanning time and data storage requirements by scanning only the feature areas, thereby improving the efficiency of three-dimensional reconstruction and being particularly suitable for samples that are complex or difficult to scan directly.

[0072] 2. The present invention combines the latest artificial intelligence technology and visual guidance technology to provide an end-to-end solution to achieve fast and accurate large-scale point cloud data feature extraction, overcome the limitations of traditional 3D scanning methods, and has broad industrial application prospects.

[0073] 3. The present invention can effectively improve the integrity and accuracy of three-dimensional reconstruction by combining actual scanning data with simulation data generated by the design model, and reduce the data missing problem in traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments made with reference to the following drawings:

[0075] Figure 1 This is a flowchart of a large-scale point cloud data feature extraction method based on artificial intelligence;

[0076] Figure 2 This is a main view schematic diagram of a large-scale point cloud data feature extraction system based on artificial intelligence;

[0077] Figure 3 A bird's-eye view of a large-scale point cloud data feature extraction system based on artificial intelligence.

[0078] The figure shows:

[0079] 1. Turntable; 2. Baffle; 3. Global tracking and positioning system; 4. Six-degree-of-freedom industrial robot; 5. Computer; 6. Local scanning measurement system. DETAILED DESCRIPTION

[0080] The present invention is described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0081] According to the large-scale point cloud data feature extraction method based on artificial intelligence provided by the present invention, by combining the two-dimensional image feature recognition technology based on artificial intelligence with the structured light measurement technology, effectively identifying the feature area and using a three-dimensional scanning device to scan and generate point cloud data, and integrating it with the point cloud simulation generated based on the sample design model to generate complete point cloud data, and efficiently and accurately reconstruct the three-dimensional of the sample to be tested. Figure 1 For example, it includes:

[0082] Step S1: for a given sample, use artificial intelligence technology to identify the area with obvious features of the sample to be tested;

[0083] Specifically, a high-resolution industrial camera is used to take all-round photos of the sample to be tested to obtain a complete and effective feature sample data set and a two-dimensional image of the sample; artificial intelligence recognition technology is used to identify areas with obvious features of the sample.

[0084] In order to fully consider the diversity of the actual application environment of the model in the future, ensure the diversity and balance of the feature sample data set, and enable the model to adapt to the detection needs under various working conditions, the following sample collection requirements are formulated:

[0085] (1) The sample image acquisition distance must be basically the same as the acquisition distance in the actual application scenario

[0086] (2) For the same feature, multiple shots should be taken, and the images taken each time should be different.

[0087] (3) For different types of milling features, the number of samples needs to be kept roughly the same.

[0088] (4) Considering the special optical properties of the sample material to be tested, samples must be collected under different lighting conditions.

[0089] The artificial intelligence-based technology includes image processing and feature recognition algorithms to determine the feature areas of the sample. Using image recognition technology, the image of the sample is preprocessed, including but not limited to grayscale, histogram equalization, image enhancement, denoising, contrast enhancement, etc., to enhance the recognizability of the feature area and improve the accuracy of feature recognition.

[0090] The feature recognition algorithm is to apply a deep learning algorithm, such as a convolutional neural network (CNN), to extract and recognize features of the processed image and determine the key feature areas of the sample, including the following steps:

[0091] Step S1.1: Use a deep learning network to extract features from the captured image;

[0092] Step S1.2: Determine key points and regions in the image based on the extracted features;

[0093] Step S1.3: Mark the key points and regions of the feature area for subsequent three-dimensional scanning.

[0094] The feature detection model based on YOLOv5 can quickly identify various feature areas on the surface of the sample to be tested. Feature area positioning, with the help of professional image annotation tools, can accurately annotate the milling features in the collected sample drawings, including key information such as feature type, location, size range, etc.

[0095] In more preferred examples, a one-stage target detection strategy is adopted to directly predict the category and location information of the target from the input image. The input end of YOLOv5 is mainly responsible for preprocessing the input image, including image scaling, normalization, data enhancement and other operations. The backbone network (Backbone) adopts the CSPNet structure and uses multiple convolution kernels of different scales for feature extraction. It can capture feature information at different levels in the image, from low-level edge and texture features to high-level semantic features, providing rich feature representation for subsequent target detection. The neck network (Neck) adopts the structure of feature pyramid network (FPN) and path aggregation network (PAN), which is mainly used to further fuse and optimize the features extracted by the backbone network. The prediction end (Prediction) performs target detection prediction based on the fused feature map. The target category probability is calculated by the softmax function:

[0096]

[0097] A softmax calculation is performed to obtain the confidence score of each category within the bounding box. Finally, multiple overlapping bounding boxes are screened using the non-maximum suppression (NMS) algorithm, and the bounding box with the highest confidence and no overlap is retained as the final acceptance result.

[0098] The loss function of YOLOv5 consists of three parts, which are used to measure the bounding box position prediction error, the bounding box size prediction error, and the target category prediction error.

[0099] Bounding box position prediction error loss function

[0100] Among them, λ coord is a weight coefficient used to balance the importance of position prediction error loss; S 2 represents the total number of grid cells; B represents the number of bounding boxes predicted for each grid cell; Indicates whether the jth bounding box in the jth grid cell is responsible for predicting the target (1 if so, 0 otherwise); (x i ,y i )and Represent the center coordinates of the predicted bounding box and the true bounding box respectively.

[0101] Bounding box size prediction error loss function

[0102] The meanings of the parameters here are the same as those in the position prediction error loss function, w i and h i and Represents the width and height of the predicted and true bounding boxes respectively.

[0103] Target category prediction error loss function:

[0104]

[0105] Among them, λ noobj is the weight coefficient; Indicates whether the grid cell contains the target (1 if it does, 0 otherwise); p i (c) and They represent the probability that the predicted bounding box belongs to the category and the probability that the true bounding box belongs to the category; C i and They represent the confidence of the predicted bounding box and the confidence of the true bounding box respectively (1 if there is an object, otherwise 0).

[0106] The total loss function is:

[0107] By minimizing the total loss function, the YOLOv5 model can continuously adjust the network parameters during the training process to improve the recognition accuracy and positioning accuracy of the characteristic areas of complex samples. In practical applications, the trained model can quickly and accurately detect various milling features on the surface of thin-walled parts, providing important basic data for subsequent scanning path planning and quality control.

[0108] Step S2: Scanning the identified feature area using a three-dimensional measurement system to obtain point cloud data of the feature area;

[0109] Specifically, a high-precision three-dimensional measurement system is used to control the movement and positioning of the scanner through a six-degree-of-freedom industrial robot to perform comprehensive, uniform and accurate scanning of the identified feature areas; scanning data is collected to generate point cloud data of the feature areas; a visual interface is used for operators to input instructions and view point cloud data.

[0110] In more preferred embodiments, step S2 includes:

[0111] Scanning path planning based on feature recognition area: During the project implementation, the motion path and limit of the robot arm were finely set based on the specific geometric dimensions and measurement requirements of the workpiece, ensuring that under multi-axis motion control, the structured light 3D measurement system can achieve multi-angle and all-round scanning coverage on the workpiece surface. Using the automated path planning algorithm, the robot arm can intelligently adjust its posture during the measurement process to avoid missed measurements or re-measurements caused by too large or too small a range of motion, ensuring that the measurement device can accurately cover all marked feature areas.

[0112] The three-dimensional measurement system is a binocular stereoscopic vision three-dimensional measurement system based on fringe projection, and the three-dimensional measurement system is driven by a six-degree-of-freedom industrial robot to scan and measure the workpiece to be measured.

[0113] The robot arm has at least six degrees of freedom, ensuring that the workpiece to be measured can be measured in all directions from various angles and positions.

[0114] Data collection of feature areas based on structured light measurement: Start the structured light 3D measurement system to scan the identified feature areas, collect point cloud data, and monitor the scanning quality in real time to ensure the integrity and accuracy of the data.

[0115] Step S3: for the unscanned area, simulating and generating point cloud data for the unscanned area according to the design model of the sample to be tested;

[0116] Based on the geometric information of the CAD model, an algorithm (such as a voxel-based meshing method) is used to generate simulated point cloud data. The density and accuracy of the simulated point cloud data are adjusted to match the actual scan data.

[0117] Specifically, for the unscanned areas, point cloud data is generated by sampling and simulating from the CAD model according to the design model of the sample. That is, the computer-aided design (CAD) model is matched with the actual scan data to generate simulated point cloud data of the unscanned areas. The simulated point cloud data is preprocessed, including denoising, smoothing and normalization, to match the density and accuracy of the actual scan data.

[0118] Step S4: Fusing the feature area point cloud data obtained in step 2 with the two parts of simulated point cloud data obtained in step 3 to reconstruct and generate complete three-dimensional point cloud data.

[0119] Specifically, the point cloud data obtained by scanning and the point cloud data generated by simulation are imported into the point cloud data processing software.

[0120] The iterative closest point algorithm (ICP) or other point cloud registration techniques are used to spatially align the feature area point cloud data and the simulation point cloud data, i.e., register them, to ensure data consistency.

[0121] Apply filtering algorithms to remove noise and outliers to improve data quality. Apply multi-view stereo matching technology to fuse the registered point cloud data. During the fusion process, select the scanned point cloud data for the identified feature area point cloud data to generate a complete 3D point cloud model; reconstruct the surface of the sample point cloud model to generate a smooth and continuous 3D surface. Post-process the fused point cloud data, including redundancy removal, hole filling and smoothing, to improve the reconstruction quality.

[0122] According to the present invention, a large-scale point cloud data feature extraction system based on artificial intelligence includes: an operating platform, a feature recognition unit, a three-dimensional scanning unit, a simulated point cloud generation unit and a data fusion unit.

[0123] The feature recognition unit is used to process the sample image and recognize and determine the feature area of ​​the sample, including image acquisition and image processing modules.

[0124] The image acquisition module uses a high-resolution industrial camera to capture and collect the characteristic sample data set of the test piece in all directions to obtain a two-dimensional image of the test piece;

[0125] The omnidirectional shooting is to keep the acquisition distance the same as the acquisition distance in the actual application scenario, shoot a rated number of two-dimensional images of the same features at different angles under different lighting conditions, and keep the number of two-dimensional images of each feature type the same;

[0126] The image processing module uses image recognition technology to pre-process the two-dimensional image;

[0127] Apply deep learning algorithms to extract features of the test piece from two-dimensional images;

[0128] Based on the features, the feature detection model based on YOLOv5 determines the key points and areas in the image;

[0129] Use image annotation tools to mark the feature type, location and size range of key points and regions to obtain feature regions;

[0130] The preprocessing includes grayscale, histogram equalization, image scaling, normalization, image enhancement, denoising and contrast enhancement.

[0131] The 3D scanning unit is used to drive the 3D measurement system to scan the feature area of ​​the workpiece and obtain point cloud data, including Figure 2 , Figure 3 Take binocular stereo vision non-contact measurement system and six-degree-of-freedom industrial robot as examples.

[0132] The binocular stereo vision non-contact measurement system comprises: a turntable 1, a baffle 2, a global tracking and positioning system 3, a six-degree-of-freedom robot 4, a computer 5, and a local scanning measurement system 6, wherein:

[0133] The turntable 1 is used to drive the measured object to rotate so that different areas of the object face the local measurement system, which is convenient for the local measurement system to measure large and complex components;

[0134] The global tracking and positioning system 3 is used to realize global system tracking and positioning. Different local areas are spliced ​​together through the posture transformation relationship when the turntable rotates, so as to realize the rough splicing of point clouds from different perspectives and obtain the complete three-dimensional point cloud data of the measured object;

[0135] The local scanning and measuring system 6 is used to scan and measure the area to be scanned of a large and complex component to obtain local point cloud data.

[0136] The 3D scanning unit uses an automated path planning algorithm to set the scanning motion path and limit positions based on the geometric dimensions and measurement requirements of the object to be measured.

[0137] Scan the feature area, collect point cloud data of the feature area and monitor the scanning quality.

[0138] The simulation point cloud generation unit is used to extract and generate simulation point cloud data of the unscanned area based on the sample design model, including a computer-aided design (CAD) model matching module.

[0139] According to the geometric information of the computer-aided design model matching module, point cloud data is generated by using a voxel-based gridding algorithm. The density and accuracy are adjusted through denoising, smoothing and normalization preprocessing to match the point cloud data of the feature area actually scanned to obtain simulated point cloud data.

[0140] The data fusion unit includes a computer-aided design (CAD) model matching module, which is used to register and fuse the scanned point cloud data with the simulated point cloud data using point cloud registration technology to reconstruct complete three-dimensional point cloud data.

[0141] The data fusion unit comprises:

[0142] Module M4.1: Use iterative closest point algorithm to spatially align feature area point cloud data and simulation point cloud data;

[0143] Module M4.2: Apply filtering algorithms to remove noise and outliers;

[0144] Module M4.3: Apply multi-view stereo matching technology to select the scanned point cloud data for the feature area point cloud data, and fuse them to generate three-dimensional point cloud data;

[0145] Module M4.4: Surface reconstruction of 3D point cloud data to generate smooth and continuous 3D surfaces;

[0146] Module M4.5: Perform redundancy removal, hole filling and smoothing post-processing to obtain a three-dimensional point cloud model.

[0147] Artificial intelligence technology and the 3D measurement system are integrated into a unified operating platform, allowing operators to complete all operations of taking pictures, identifying, scanning and fusion through a single interface.

[0148] In the description of the present application, it should be understood that the terms "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present application.

[0149] The above describes the specific embodiments of the present invention. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which does not affect the essence of the present invention. In the absence of conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. A large-scale point cloud data feature extraction system based on artificial intelligence, characterized in that: include: Feature recognition unit, 3D scanning unit, simulation point cloud generation unit, data fusion unit and operation platform; The feature recognition unit recognizes the region with obvious features of the test piece and obtains the feature region; The three-dimensional scanning unit scans the feature area to obtain point cloud data of the feature area; The simulation point cloud generation unit generates simulation point cloud data according to the model of the test piece; The data fusion unit fuses the feature area point cloud data and the simulation point cloud data to reconstruct and generate a three-dimensional point cloud model; The operating platform controls the operation of the feature recognition unit, the 3D scanning unit, the simulation point cloud generation unit and the data fusion unit according to the input instructions, and displays the 3D point cloud model.

2. The large-scale point cloud data feature extraction system based on artificial intelligence according to claim 1 is characterized in that: The feature recognition unit includes an image acquisition module and an image processing module; The image acquisition module uses a high-resolution industrial camera to capture and collect the characteristic sample data set of the test piece in all directions to obtain a two-dimensional image of the test piece; The omnidirectional shooting is to keep the acquisition distance the same as the acquisition distance in the actual application scenario, shoot a rated number of two-dimensional images of the same features at different angles under different lighting conditions, and keep the number of two-dimensional images of each feature type the same; The image processing module uses image recognition technology to pre-process the two-dimensional image; Apply deep learning algorithms to extract features of the test piece from two-dimensional images; Based on the features, the feature detection model based on YOLOv5 determines the key points and areas in the image; Use image annotation tools to mark the feature type, location and size range of key points and regions to obtain feature regions; The preprocessing includes grayscale, histogram equalization, image scaling, normalization, image enhancement, denoising and contrast enhancement.

3. The large-scale point cloud data feature extraction system based on artificial intelligence according to claim 1 is characterized in that: The three-dimensional scanning unit includes a binocular stereo vision non-contact measurement system; The binocular stereo vision non-contact measurement system comprises a turntable (1), a baffle (2), a global tracking and positioning system (3), a six-degree-of-freedom industrial robot (4), a computer (5), and a local scanning measurement system (6); The baffle (2) surrounds three sides of the binocular stereo vision non-contact measurement system; The turntable (1) drives the test piece to rotate; The six-degree-of-freedom industrial robot (4) controls the movement and positioning of the local scanning measurement system (6), performs a comprehensive and uniform scanning of the characteristic area of ​​the complex test piece, and obtains local point cloud data; The global tracking and positioning system (3) stitches together different local areas through the position conversion relationship when the turntable (1) rotates, obtains point cloud data of the feature area, and transmits the data fusion module and the operation platform through the computer (5).

4. The large-scale point cloud data feature extraction system based on artificial intelligence according to claim 1 is characterized in that: The three-dimensional scanning unit uses an automated path planning algorithm to set the motion path and limit of the scan based on the geometric dimensions and measurement requirements of the object to be measured; Scan the feature area, collect point cloud data of the feature area and monitor the scanning quality; The simulation point cloud generation unit includes a computer-aided design model matching module; Generate point cloud data according to the geometric information of the computer-aided design model matching module, adjust the density and accuracy to match the point cloud data of the feature area actually scanned, and obtain simulated point cloud data; The data fusion unit imports feature area point cloud data and simulation point cloud data, registers them using point cloud registration technology, fuses them to generate three-dimensional point cloud data, reconstructs the surface and performs post-processing to obtain a three-dimensional point cloud model.

5. The large-scale point cloud data feature extraction system based on artificial intelligence according to claim 4 is characterized in that: The simulated point cloud generation unit generates point cloud data using a voxel-based gridding algorithm, and adjusts density and accuracy through denoising, smoothing and normalization preprocessing to match the density and accuracy of the feature area point cloud data; The data fusion unit comprises: Module M4.1: Use iterative closest point algorithm to spatially align feature area point cloud data and simulation point cloud data; Module M4.2: Apply filtering algorithms to remove noise and outliers; Module M4.3: Apply multi-view stereo matching technology to select the scanned point cloud data for the feature area point cloud data, and fuse them to generate three-dimensional point cloud data; Module M4.4: Surface reconstruction of 3D point cloud data to generate smooth and continuous 3D surfaces; Module M4.5: Perform redundancy removal, hole filling and smoothing post-processing to obtain a three-dimensional point cloud model.

6. A method for extracting features of large-scale point cloud data based on artificial intelligence, using the large-scale point cloud data feature extraction system based on artificial intelligence as claimed in any one of claims 1 to 5, characterized in that: include: Step S1: Identify the area with obvious features of the test piece and obtain the feature area; Step S2: Scan the feature area to obtain point cloud data of the feature area; Step S3: Generate simulated point cloud data according to the model of the test piece; Step S4: Fuse the feature area point cloud data and the simulation point cloud data to reconstruct and generate a three-dimensional point cloud model.

7. The method for extracting large-scale point cloud data features based on artificial intelligence according to claim 6, characterized in that: The step S1 comprises: Step S1.1: Using a high-resolution industrial camera, capture a feature sample data set of the test piece in all directions to obtain a two-dimensional image of the test piece; Step S1.2: Using artificial intelligence recognition technology, identify the region with obvious features in the two-dimensional image of the test piece to obtain the feature region; The omnidirectional shooting is to keep the collection distance the same as the collection distance in the actual application scene, shoot a rated number of two-dimensional images of the same features at different angles under different lighting conditions, and keep the number of two-dimensional images of each feature type the same.

8. The method for extracting large-scale point cloud data features based on artificial intelligence according to claim 7, characterized in that: The step S1.2 comprises: Image preprocessing step: Use image recognition technology to preprocess the two-dimensional image; Feature recognition step: Apply deep learning algorithms to extract and recognize features of the processed two-dimensional image to determine the feature area of ​​the test piece; The preprocessing includes grayscale, histogram equalization, image scaling, normalization, image enhancement, denoising and contrast enhancement; The feature extraction and recognition includes: Step S1.2.1: Use a deep learning network to extract features of the test piece from the two-dimensional image; Step S1.2.2: According to the features, the key points and regions in the image are determined based on the feature detection model of YOLOv5; Step S1.2.3: Use image annotation tools to mark the feature type, position and size range of key points and regions to obtain feature regions.

9. The method for extracting large-scale point cloud data features based on artificial intelligence according to claim 6, characterized in that: The step S2 comprises: Scanning path planning steps: Use the automated path planning algorithm to set the scanning motion path and limit based on the geometric dimensions and measurement requirements of the object to be measured; Data collection steps: Use a structured light measurement device to scan the feature area, collect point cloud data of the feature area and monitor the scanning quality; In the step S3, point cloud data is generated according to the geometric information of the model of the test piece, and the density and accuracy are adjusted to match the point cloud data of the characteristic area actually scanned to obtain simulated point cloud data; In step S4, the feature area point cloud data and the simulation point cloud data are imported, registered using the point cloud registration technology, fused to generate three-dimensional point cloud data, and the surface is reconstructed and post-processed to obtain a three-dimensional point cloud model.

10. The method for extracting large-scale point cloud data features based on artificial intelligence according to claim 9, characterized in that: The structured light measurement device is a binocular stereoscopic vision three-dimensional measurement system based on fringe projection. It controls movement and positioning through a six-degree-of-freedom industrial robot, performs a comprehensive and uniform scan of the feature area, collects scanning data to generate feature area point cloud data, and visualizes it through an operating platform; In step S3, a voxel-based gridding algorithm is used to generate point cloud data, and density and precision are adjusted through denoising, smoothing and normalization preprocessing to match the density and precision of the point cloud data in the feature area; The step S4 includes: Step S4.1: spatially aligning the feature area point cloud data and the simulation point cloud data using an iterative closest point algorithm; Step S4.2: Apply filtering algorithm to remove noise and outliers; Step S4.3: Apply multi-view stereo matching technology, select the scanned point cloud data for the feature area point cloud data, and fuse them to generate three-dimensional point cloud data; Step S4.4: Surface reconstruction is performed on the three-dimensional point cloud data to generate a smooth and continuous three-dimensional surface; Step S4.5: Perform redundancy removal, hole filling and smoothing post-processing to obtain a three-dimensional point cloud model.

Images