A method for 3D reconstruction of workpiece based on Gaussian mixture model

Through the three-dimensional reconstruction method of artifacts based on Gaussian hybrid model, using multi-view point cloud data and rotation platform, combined with RANSAC and greedy projection triangulation algorithm, the problems of large errors and high equipment dependence of the existing three-dimensional reconstruction method are solved, achieving high efficiency , accurate three-dimensional model reconstruction.

CN113362463BActive Publication Date: 2024-11-26ZHEJIANG UNIV OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202110535760.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-05-17
Publication Date
2024-11-26
Estimated Expiration
2041-05-17

AI Technical Summary

Technical Problem

Among the existing three-dimensional reconstruction methods, the method based on monocular two-dimensional images has a large error and is susceptible to camera parameters. The method based on deep learning requires a large number of samples and high-performance equipment, and has high equipment dependence and slow reconstruction speed.

Method used

Using the three-dimensional reconstruction method of artifacts based on Gaussian hybrid model, multi-angle point cloud data acquisition, point cloud preprocessing, point cloud registration and surface reconstruction steps are used to obtain multi-angle point cloud data using a converged binocular depth camera and rotating platform. , combined with RANSAC algorithm and Gaussian hybrid model for point cloud registration, and surface reconstruction was carried out using greedy projection triangulation algorithm.

Benefits of technology

It achieves simple data acquisition, fast reconstruction speed, low equipment dependence, high accuracy of the three-dimensional model, clear and concise process, and is suitable for industrial production scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113362463B_ABST
    Figure CN113362463B_ABST
Patent Text Reader

Abstract

A method for three-dimensional reconstruction of workpieces based on a Gaussian mixture model includes: step 1, multi-view point cloud data collection, point cloud preprocessing, point cloud registration, and surface reconstruction; step 2, using a depth camera to collect point cloud data, and using a rotating platform to obtain multi-view information of the workpiece; step 3, then performing point cloud preprocessing such as target extraction on the collected data; step 4, on this basis, registering the workpiece point cloud, establishing a full-view point cloud model and performing surface reconstruction on the model, and reconstructing a three-dimensional model close to the real object. The advantages of the present invention are simple implementation and fast reconstruction speed.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the field of 3D vision technology, and in particular to a workpiece three-dimensional reconstruction method based on a Gaussian mixture model. Background Art

[0002] The application demand of 3D reconstruction technology in data visualization, medical technology, industry and other fields is getting higher and higher. Taking the industrial production scenario as an example, on the spraying production line, the 3D reconstruction model of the workpiece to be processed is used to automatically plan the spraying trajectory and hand it over to the robot for execution. This can reduce manual operation and improve the safety of spraying operations while improving work efficiency and automation.

[0003] Driven by visual sensor technology, various 3D reconstruction methods have developed rapidly. However, the 3D reconstruction method based on monocular 2D images has large errors and is easily affected by camera parameters. The 3D reconstruction method based on deep learning requires a large number of samples and requires the use of high-performance equipment for model training. The 3D reconstruction method of workpieces based on the Gaussian mixture model is to obtain a 3D model by aligning and reconstructing the multi-view point cloud data of the target workpiece based on the point cloud data obtained by the 3D vision sensor. This method has simple data acquisition, low dependence on equipment, and fast reconstruction speed. Summary of the invention

[0004] The present invention aims to overcome the deficiencies of the prior art and provides a method for three-dimensional reconstruction of a workpiece based on a Gaussian mixture model.

[0005] A method for three-dimensional reconstruction of a workpiece based on a Gaussian mixture model comprises the following steps:

[0006] Step 1: Multi-view point cloud data acquisition, collect scene point cloud data containing multi-angle target workpieces. This step uses a fused binocular depth camera to collect point cloud data. The depth camera obtains depth data based on the principle of binocular stereo imaging and infrared structured light ranging. The target workpiece to be scanned is placed on a rotating platform with a controllable rotation angle. The depth camera and the rotating platform form a fixed relative position. The workpiece is placed on the rotating platform at a certain initial position and rotated in increments of fixed angles to obtain multi-angle information of the target workpiece. The depth camera scans and records each angle in the scene and then transmits it back to the computer, and saves the point cloud data in the form of a PCD file in chronological order, named View1, View2, View3,...ViewN and other N-view scene point cloud files.

[0007] Step 2: Point cloud preprocessing: remove irrelevant data from the multi-view scene point clouds View1, View2, View3, ...ViewN collected in step 1, and extract the target workpiece point cloud data. The specific steps are as follows:

[0008] Step 2-1: Set the ROI parameters according to the relative position of the depth camera and the rotating platform in step 1, and perform ROI region segmentation and screening on the spatial points in the complex scene point cloud View1, View2, View3, ...ViewN, that is, select the three-dimensional points whose x, y, z (the three-dimensional spatial coordinates of each point) are within the ROI region. Preliminarily segment the small scene point cloud containing only the ground, the rotating platform, and the workpiece.

[0009] Step 2-2: The point cloud data set in the small scene point cloud obtained in the previous step is expressed as:

[0010] A{a1,a2,a3,...a n}

[0011] The random sampling consensus (RANSAC) algorithm is used to fit the plane in the point cloud set A. The RANSAC algorithm fits the plane and its plane parameters in the small scene, and divides the points in the A set into two categories: plane points and out-of-plane points. The subscript index of the plane point and the subscript index of the out-of-plane point are recorded, and the plane is removed according to the subscript to eliminate the plane points. After the RANSAC algorithm obtains the plane parameters, the position of the plane in the small scene point cloud can be determined. By measuring the height of the rotating platform as H, all points at a height of H above the plane can be eliminated, thereby eliminating the point cloud of the rotating platform part and obtaining preliminary target workpiece data.

[0012] Step 2-3: The preliminary target workpiece data obtained in step 2-2 will have outliers caused by insufficient algorithm efficiency, as well as surface burr noise and edge noise left when the 3D vision sensor collects data. Therefore, the StatisticalOutlierRemovel filter in the PCL point cloud library is used to filter the above results as input to remove outliers and surface outlier noise. Finally, multi-view target workpiece point cloud data Obj1, Obj2, Obj3, ... ObjN are extracted from View1, View2, View3, ... ViewN.

[0013] Step 3: Point cloud registration, point cloud registration based on Gaussian mixture model. This process is to register Obj1, Obj2, Obj3, ... ObjN obtained in steps 2-3 in pairs, obtain the transformation matrix, perform global splicing, and obtain a complete point cloud model.

[0014] Step 3-1: Establish a Gaussian mixture model for the point clouds of two adjacent viewpoints. Select the two adjacent viewpoint point clouds to be registered from Obj1, Obj2, Obj3, ... ObjN, set the target point cloud Scene, and the point cloud to be registered Model. It is known that the Gaussian continuous probability density distribution function is:

[0015]

[0016] Among them, μ is the mean vector, ∑ is the covariance matrix, and d is the dimension of the data.

[0017] The Gaussian mixture model is built according to the following guidelines:

[0018] S1. The number of Gaussian components in the Gaussian mixture model is equal to the number of point clouds in each point cloud dataset.

[0019] S2. For each Gaussian component in the Gaussian mixture model, its mean vector is set according to the spatial position of the point.

[0020] S3. All Gaussian components in the Gaussian mixture model share the same covariance matrix.

[0021] Finally, all Gaussian components as described above are added with the same weight, and we get:

[0022]

[0023] where w i Weight coefficient of Gaussian mixture model. According to the above rules, Gaussian mixture models gmm(S) and gmm(M) are established for Scene and Model. gmm represents the functional relationship in (2) above. Input S represents point cloud Scene and M represents point cloud Model.

[0024] Step 3-2: Establish a transformation matrix between two point clouds with parameter θ. The Model point cloud after parameter transformation is expressed as Transform(M,θ), and its Gaussian mixture model can be expressed as gmm(Transform(M,θ)), where Transform represents a function that performs corresponding rigid body transformation according to the transformation matrix with parameter θ;

[0025] Step 3-3: Perform square integration of the difference between the two Gaussian mixture models to establish a differentiated objective function:

[0026] ∫(gmm(S)-gmm(Transform(M,θ))) 2 dx(3) uses the fixed rotation parameter used in step 1 as the initial value of parameter θ, uses the Gauss-Newton algorithm to perform iterative optimization operations to minimize the above objective function, and records the value of parameter θ. The transformation matrix T is calculated based on the parameter value.

[0027] Step 3-4: Take Obj1 as the reference for registration, and the coordinate system where Obj1 is located as the reference coordinate system. According to step 3-3, Obj1 and Obj2 (Obj1 is regarded as Scene, Obj2 is regarded as Model) are registered to obtain the transformation matrix T 12 , Obj2 can be obtained through T 12 The matrix is ​​transformed to the reference coordinate system. Obj2 and Obj3 are processed by Gaussian mixture model point cloud registration to obtain the transformation matrix T 23 , Obj3 can be obtained through T 12 *T 23 Matrix transformation to the reference coordinate system. Perform pairwise registration calculations and transform to the reference coordinate system according to the transformation matrix, and splice the multi-view point clouds to obtain a full-view point cloud model.

[0028] Step 4: Surface reconstruction, using the greedy projection triangulation algorithm to reconstruct the surface of the model's scattered point cloud.

[0029] Step 4-1: Establish a kd-tree spatial structure index for the full-view point cloud model to speed up the point cloud query. And use the kd-tree structure to find the K neighborhood of the target point.

[0030] Step 4-2: Select an initial triangle in the point cloud collection, obtain the K neighborhood of the midpoint of the growing edge of the triangle, and project the points in the neighborhood to the two-dimensional plane.

[0031] Step 4-3: Select the projection point with the smallest cosine value with the growing edge in the two-dimensional plane as the optimal expansion point.

[0032] Step 4-4: Map the optimal expansion points back to the three-dimensional space to form new triangles in the triangular mesh.

[0033] Step 4-5: Repeat steps 4-2 to 4-5 to continuously generate new triangles until all points on the surface of the point cloud model form a complete triangular mesh surface.

[0034] The advantages of the present invention are:

[0035] 1. The data collection system used is simple to set up and obtains comprehensive information.

[0036] 2. The algorithm process is clear and concise. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a flowchart of the method of the present invention.

[0038] Figure 2-a It is the small scene point cloud of the ROI preliminary segmentation of the present invention.

[0039] Figure 2-b It is the target extraction result point cloud data of the present invention.

[0040] Figure 3 This is the working process of the point cloud registration step of the present invention. DETAILED DESCRIPTION

[0041] The present invention is further described in detail below with reference to the accompanying drawings.

[0042] A method for 3D reconstruction of workpieces based on Gaussian mixture model, the workflow diagram is as follows: Figure 1 As shown in the figure, this method consists of four steps: multi-view point cloud data acquisition, point cloud preprocessing, point cloud registration, and surface reconstruction. Figure 2-a The display is the point cloud data of the small scene after ROI segmentation. Figure 2-b Displays the extracted point cloud data of the target workpiece after removing the plane. Figure 3 In point cloud registration, there are three perspective registration processes.

[0043] Combination Figure 1-3 Taking the front wheel fender of an electric vehicle as an example, the specific implementation method of the present invention is as follows:

[0044] Step 1: Multi-view point cloud data collection. Use the Intel Realsense D435 depth camera to collect the point cloud data of the work scene. The depth camera is placed on a 1.8m high tripod and connected to the USB3.0 port of the computer using an extended USB3.0 transmission data cable. The workpiece is placed on a rotating platform (the initial position can be facing the depth camera), the ROI parameters between the rotating platform and the depth camera are measured and recorded, and the rotation angle of the rotating platform is set (taking 45° as an example). The rotating platform controls the rotation angle through the motor. Every 45° rotation, the depth camera will collect the data and send it to the computer until the workpiece rotates 360°. The scene point cloud data of one angle is collected.

[0045] The depth camera and the computer transmit images and point cloud data via a USB cable. The computer uses Ubuntu 16.04 operating system + ROS Kinetic + PCL as the software platform for data collection, processing and presentation.

[0046] The realsense-ros package supported by the ROS community publishes point cloud information transmitted from the realsense depth camera to the computer. You can subscribe to the above point cloud information by checking the topic name corresponding to the point cloud information through rostopic and use the fromROSMsg conversion function contained in the pcl_conversions in the PCL function library to process the published sensor_msgs / PointCloud2 format data to generate the corresponding PCL format point cloud data, and export it to PCD files through IO operations and store them in the hard disk. The files are View1, View2, View3,...ViewN.

[0047] Step 2: Point cloud preprocessing to extract the target workpiece point cloud data.

[0048] Step 2-1: Set the ROI parameters according to the relative positions of the depth camera and the rotating platform in step 1.

[0049] Perform ROI region segmentation and screening on the spatial points in the complex scene point cloud View1, View2, View3, ...ViewN, that is, select the three-dimensional points whose x, y, z (the three-dimensional spatial coordinates of each point) are located in the ROI region. A small scene point cloud containing only the ground, rotating platform, and workpiece is initially segmented. The segmentation effect is shown below: Figure 2-a shown.

[0050] Step 2-2: Perform plane removal processing based on 2-1, and further remove the plane and rotating platform. The point cloud data set in the small scene point cloud obtained in 2-1 is expressed as:

[0051] A{a1,a2,a3,...an}

[0052] The random sampling consensus (RANSAC) algorithm is used to fit the plane in the point cloud set A. The RANSAC algorithm fits the plane and its plane parameters in the small scene, and divides the points in the A set into two categories: points on the plane and points that do not belong to the plane. The subscripts of the data points that belong to the plane and the subscripts of the data points that do not belong to the plane are recorded, and the plane is removed according to the subscripts to eliminate the points that belong to the plane. After the RANSAC algorithm obtains the plane parameters, the position of the plane in the small scene point cloud can be determined. By measuring the height of the rotating platform as H, all points at a height of H above the plane can be eliminated, thereby eliminating the data of the rotating platform part and obtaining preliminary target workpiece data.

[0053] Step 2-3: The preliminary target workpiece data obtained in step 2-2 will have outliers caused by insufficient algorithm efficiency, as well as surface burr noise and edge noise left when the 3D vision sensor collects data, which will cause errors in subsequent steps. Therefore, the StatisticalOutlierRemovel filter in the PCL library is used to filter the above results as input to remove outliers and surface noise. Finally, multi-view target workpiece point cloud data Obj1, Obj2, Obj3, ... ObjN are extracted from View1, View2, View3, ... ViewN. The extracted target workpiece data is as follows: Figure 2-b shown.

[0054] Step 3: Point cloud registration. This process is to register Obj1, Obj2, Obj3, ... ObjN obtained in steps 2-3 in pairs, perform global stitching, and obtain a complete model.

[0055] Step 3-1: Perform registration processing for the point clouds of two adjacent frames Obj1 and Obj, Obj1 is the target point cloud Scene, and Obj2 is the point cloud Model to be registered. The Gaussian continuous probability density distribution function is known to be:

[0056]

[0057] Among them, μ is the mean vector, ∑ is the covariance matrix, and d is the dimension of the data.

[0058] The Gaussian mixture model of Obj1 and Obj2 is established according to the following criteria:

[0059] 1) The number of Gaussian components in the Gaussian mixture model is equal to the number of point clouds in each point cloud dataset.

[0060] 2) For each Gaussian component in the Gaussian mixture model, its mean value vector is set according to the spatial position of the point.

[0061] 3) All Gaussian components in the Gaussian mixture model share the same covariance matrix.

[0062] Finally, all Gaussian components as described above are added with the same weight, and we get:

[0063]

[0064] where w i is the weight coefficient of each Gaussian component in the Gaussian mixture model. According to the above rules, the Gaussian mixture models gmm(S) and gmm(M) are established for Scene and Model. gmm represents the functional relationship in (2) above. The input S represents the point cloud Scene and M represents the point cloud Model.

[0065] Step 3-2: Establish a transformation matrix between two point clouds with parameter θ. The Model point cloud after parameter transformation is expressed as Transform(M,θ), and its Gaussian mixture model can be expressed as gmm(Transform(M,θ)), where Transform represents a function that performs corresponding rigid body transformation according to the transformation matrix with parameter θ;

[0066] Step 3-3: Perform square integration of the difference between the two Gaussian mixture models to establish a differentiated objective function:

[0067] ∫(gmm(S)-gmm(Transform(M,θ))) 2 dx(3) takes the fixed rotation parameter used in step 1 as the initial value of parameter θ (taking 45° as an example), and uses the Gauss-Newton algorithm to perform iterative optimization operations to obtain the parameter value that minimizes the above objective function. The transformation matrix T is calculated based on the parameter value.

[0068] Step 3-4: Take Obj1 as the reference for registration, and regard the coordinate system where Obj1 is located as the reference coordinate system. The transformation matrix T is obtained by registration 12 (Transformation matrix from Obj2 to Obj2), Obj2 can be transformed by T 12 The matrix is ​​transformed to the reference coordinate system. Obj2 and Obj3 are processed by Gaussian mixture model point cloud registration to obtain the transformation matrix T 23 , Obj3 can be obtained through T 12 *T 23 Matrix transformation to the reference coordinate system. Perform pairwise registration calculations and transform to the reference coordinate system according to the transformation matrix, and splice the multi-view point clouds to obtain a full-view point cloud model.

[0069] Step 4: Surface reconstruction: Use the greedy projection triangulation algorithm to reconstruct the surface of the model's scattered point cloud.

[0070] Step 4-1: Establish a kd-tree spatial structure index for the full-view point cloud model, and use the kd-tree structure to find the K neighborhood of the target point. The value of K can be adjusted according to the actual situation and reconstruction effect.

[0071] Step 4-2: Select an initial triangle in the point cloud collection, obtain the K neighborhood of the midpoint of the growing edge of the triangle, and project the points in the neighborhood to the two-dimensional plane.

[0072] Step 4-3: Select the projection point with the smallest cosine value with the growing edge in the two-dimensional plane as the optimal expansion point.

[0073] Step 4-4: Map the optimal expansion points back to the three-dimensional space to form new triangles in the triangular mesh.

[0074] Step 4-5: Repeat steps 4-2 to 4-5 to continuously generate new triangles until all points on the surface of the point cloud model form a complete triangular mesh surface.

[0075] It should be emphasized that the implementation cases described in the present invention are illustrative rather than restrictive. Therefore, the present invention includes but is not limited to the embodiments described in the specific implementation plans. Any other similar implementation plans derived by those skilled in the art based on the technical solution of the present invention also fall within the scope of protection of the present invention.

Claims

1. A method for three-dimensional reconstruction of a workpiece based on a Gaussian mixture model, comprising the following steps: Step 1: Multi-view point cloud data collection, collecting scene point cloud data containing multi-angle target workpieces; The point cloud data is collected by using a fusion binocular depth camera. The depth camera obtains depth data based on the binocular stereo imaging principle and the infrared structured light ranging principle. The target workpiece to be scanned is placed on a rotating platform with a controllable rotation angle. The depth camera and the rotating platform form a fixed relative position. The workpiece is placed on the rotating platform at a certain initial position and rotated step by step with a fixed angle as an increment to obtain multi-angle information of the target workpiece. The depth camera scans and records each angle in the scene and transmits it back to the computer, and saves the point cloud data in the form of a PCD file in chronological order, named as N perspective scene point cloud files of View1, View2, View3, ...ViewN; Step 2: Point cloud preprocessing, removing irrelevant data from the multi-view scene point clouds View1, View2, View3, ...ViewN collected in step 1, and extracting the target artifact; The specific steps are as follows: Step 2-1: Set ROI parameters according to the relative positions of the depth camera and the rotating platform in step 1, and perform ROI region segmentation and screening on the spatial points in the complex scene point cloud View1, View2, View3, ...ViewN, that is, select the 3D points whose 3D spatial coordinates x, y, z are within the ROI region for each point; preliminarily segment a small scene point cloud containing only the ground, the rotating platform, and the workpiece; Step 2-2: The point cloud data set in the small scene point cloud obtained in the previous step is expressed as: <h2 style=";text-align:left;direction:ltr">A{a1,a2,a3,...a<h2 style=";text-align:left;direction:ltr"> n <h2 style=";text-align:left;direction:ltr">} The random sampling consensus (RANSAC) algorithm is used to fit the plane in the point cloud set A. The RANSAC algorithm fits the plane and its plane parameters in the small scene, and divides the points in the A set into two categories: plane points and out-of-plane points. The subscript index of the plane points and the subscript index of the out-of-plane points are recorded, and the plane is removed according to the subscript to eliminate the plane points. After the RANSAC algorithm obtains the plane parameters, the position of the plane in the small scene point cloud can be determined. By measuring the height of the rotating platform as H, all points at a height of H above the plane can be eliminated, thereby eliminating the point cloud of the rotating platform part and obtaining preliminary point cloud data of the target workpiece. Step 2-3: The preliminary target workpiece point cloud data obtained in step 2-2 will have outliers caused by insufficient algorithm efficiency, as well as surface burr noise and edge noise left when the 3D vision sensor collects data; therefore, the StatisticalOutlierRemovel filter in the PCL point cloud library is used to filter the preliminary target workpiece point cloud data as input to remove outliers and surface noise; finally, multi-view target workpiece point cloud data Obj1, Obj2, Obj3, ... ObjN are extracted from View1, View2, View3, ... ViewN; Step 3: Point cloud registration, based on the Gaussian mixture model, point cloud registration is performed; this process is to register Obj1, Obj2, Obj3, ... ObjN obtained in steps 2-3 in pairs, obtain the transformation matrix, perform global splicing, and obtain a complete point cloud model; Step 3-1: Establish a Gaussian mixture model for the point clouds of two adjacent viewpoints; select the two adjacent viewpoint point clouds to be registered from Obj1, Obj2, Obj3, ... ObjN, set the target point cloud as Scene, and the point cloud to be registered as Model; the Gaussian continuous probability density distribution function is known to be: Among them, μ is the mean vector, ∑ is the covariance matrix, and d is the dimension of the data; The Gaussian mixture model is built according to the following guidelines: S1. The number of Gaussian components in the Gaussian mixture model is equal to the number of point clouds in each point cloud dataset; S2. For each Gaussian component in the Gaussian mixture model, its mean value vector is set according to the spatial position of the point; S3. All Gaussian components in the Gaussian mixture model share the same covariance matrix; Finally, all Gaussian components as described above are added with the same weight, and we get: where w i Represents the weight coefficient of the Gaussian mixture model; for Scene and Model, establish Gaussian mixture models gmm(S), gmm(M) according to the above rules, where gmm represents the functional relationship in (2) above, input S represents the point cloud Scene, and M represents the point cloud Model; Step 3-2: Establish a transformation matrix between two point clouds with parameter θ. The Model point cloud after parameter transformation is expressed as Transform(M,θ), and its Gaussian mixture model can be expressed as gmm(Transform(M,θ)), where Transform represents a function that performs corresponding rigid body transformation according to the transformation matrix with parameter θ; Step 3-3: Perform square integration of the difference between the two Gaussian mixture models to establish a differentiated objective function: ∫(gmm(S)-gmm(Transform(M,θ))) 2 dx(3) The fixed rotation parameter used in step 1 is used as the initial value of the parameter θ, and the Gauss-Newton algorithm is used to perform iterative optimization operations to minimize the above objective function and record the value of the parameter θ; the transformation matrix T is calculated based on the parameter value; Step 3-4: Take Obj1 as the reference for registration, and regard the coordinate system where Obj1 is located as the reference coordinate system; according to step 3-3, register Obj1 and Obj2 to obtain the transformation matrix T 12 , where Obj1 is considered as Scene, Obj2 is considered as Model, and Obj2 can be accessed through T 12 Matrix transformation to the reference coordinate system; Gaussian mixture model point cloud registration processing is performed on Obj2 and Obj3 to obtain the transformation matrix T 23 , Obj3 can be obtained through T 12 *T 23 Matrix transformation to the reference coordinate system; perform registration calculations for each view in turn and transform to the reference coordinate system according to the transformation matrix, and splice the multi-view point clouds to obtain a full-view point cloud model; Step 4: Surface reconstruction, using the greedy projection triangulation algorithm; reconstructing the surface of the model's scattered point cloud; Step 4-1: Establish a kd-tree spatial structure index for the full-view point cloud model to speed up the point cloud query; and use the kd-tree structure to find the K neighborhood of the target point; Step 4-2: Select an initial triangle in the point cloud set, obtain the K neighborhood of the midpoint of the growing edge of the triangle, and project the points in the neighborhood onto the two-dimensional plane; Step 4-3: Select the projection point with the smallest cosine value with the growing edge in the two-dimensional plane as the optimal expansion point; Step 4-4: Map the optimal expansion point back to the three-dimensional space to form a new triangle in the triangular mesh; Step 4-5: Repeat steps 4-2 to 4-5 to continuously generate new triangles until all points on the surface of the point cloud model form a complete triangular mesh surface.

2. A method for three-dimensional reconstruction of a workpiece based on a Gaussian mixture model as claimed in claim 1, characterized in that : The multi-view point cloud data acquisition method described in step 1 only uses one depth camera to obtain point cloud data, and obtains multi-view information of the workpiece with the help of a rotating platform.