A pose solving method for disordered stacked parts with weak textures based on deep learning

Abstract

The present invention relates to a method for solving the pose of disordered stacked parts with weak textures based on deep learning. First, computer technology is used to automatically generate labeled data for network training, which greatly reduces the cost of manual labeling. Secondly, the calculation of the attention matrix is ​​added to the network design to improve the effect of point cloud instance segmentation. The added feature point prediction module predicts the point cloud of a single part instance and solves the predefined feature point position. Then, the network is enabled to solve the pose of the part through point cloud registration. Compared with the traditional point cloud registration method, the proposed feature point registration method greatly reduces the time loss and avoids the disadvantage of poor algorithm performance due to different sensor sampling distances. Finally, in the inference stage, the Hough voting method is used to solve the accurate position of the feature point, optimize the error introduced by instance segmentation, and further improve the accuracy of the pose prediction of disordered stacked parts with weak textures.

Description

Technical Field

[0001] The present invention relates to the technical field related to intelligent manufacturing, and in particular to a method for solving the pose of disordered stacked parts with weak textures based on deep learning. Background Art

[0002] Currently, automated industrial sorting production lines, such as those for automotive parts production, load a large number of parts onto the material frame, often in a scattered, disordered state. Robots automatically place the parts one by one in designated locations before welding and assembly processes proceed. The mainstream approach in the market uses RGB or RGBD cameras to calculate the homogeneous transformation matrix of the workpiece relative to the robot base using image and point cloud algorithms. For disorderly stacked, weakly textured workpieces, the design of algorithms to determine their pose is challenging due to the lack of surface features and significant occlusion between them. Meanwhile, deep learning is rapidly developing. Although deep learning methods can extract higher-dimensional features of objects, the high cost of training data annotation, high computing power requirements, and poor stability and reliability have prevented their widespread application in industrial scenarios. Summary of the Invention

[0003] The purpose of the present invention is to solve at least one of the deficiencies of the prior art and to provide a method for solving the pose of disordered stacked parts with weak textures based on deep learning.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions:

[0005] Specifically, a pose solving method for disordered stacked parts with weak textures based on deep learning is proposed, which includes the following:

[0006] Part actual point cloud acquisition: Scan the part in the plane to obtain the scanned point cloud, and perform segmentation and preprocessing operations on the scanned point cloud to obtain point cloud files of the front and back sides of several parts;

[0007] CAD model point cloud generation: pre-generate a batch of homogeneous transformation matrices of parts relative to the world coordinate system when stacking parts, and generate training data by reading the point cloud files and transforming them to the corresponding stacking positions based on the homogeneous transformation matrix;

[0008] Deep learning model and training: Train the improved FPCC network based on the training data and save the weight file of the trained network.

[0009] Specifically, the improved FPCC network adds a feature point prediction module on the basis of the original FPCC network, and adds an attention matrix to the instance segmentation loss calculation of the original FPCC network;

[0010] Calculation of the pose of actual stacked parts: Obtain the scanned point cloud of the target part, perform point cloud preprocessing on the scanned point cloud of the target part, and input it into the weight file for prediction. The output center score, point cloud instance label and feature point coordinates are obtained, and Hough voting is performed to obtain the feature point position of the target part.

[0011] The homogeneous transformation matrix corresponding to the target part is obtained by performing ICP point cloud registration with the feature point position and the feature point template.

[0012] Based on the homogeneous transformation matrix corresponding to the target part, the workpiece with a registration error lower than a preset threshold is found and its pose is output.

[0013] Further, specifically, scanning a part in a plane to obtain a scanned point cloud, segmenting and preprocessing the scanned point cloud to obtain point cloud files of the front and back sides of several parts, including:

[0014] The scanning point cloud of the part in the plane is collected by a 3D camera, and the point cloud belonging to the part is segmented from the scanning point cloud to obtain the part point cloud. A total of K part point clouds are collected from both the front and back sides. At the same time, the CAD model of the part is sampled into a point cloud, and the FPS farthest point sampling algorithm is used to sample K part feature points. K point cloud files are constructed based on the part point cloud and the part feature points.

[0015] Furthermore, specifically, a homogeneous transformation matrix of parts relative to the world coordinate system when a batch of parts are stacked is pre-generated, including:

[0016] A rectangular material frame is placed in the simulation environment. A batch of parts with random heights and random positions are generated above the material frame, and the parts are allowed to fall freely into the material frame. The number of parts is gradually increased from 3 to 50, and this is repeated 20 times. The homogeneous transformation matrix of each part relative to the world coordinate system is recorded when it is stationary.

[0017] Further, specifically, the process of generating training data includes,

[0018] Read the saved homogeneous transformation matrix of the parts relative to the world coordinate system one by one, randomly select the part point cloud and its corresponding part feature points in the collected point cloud file and transform them to the physical simulation position to obtain the transformed part point cloud and part feature points, remove all the blocked points in the transformed part point cloud and part feature points, assume that a single part point cloud has N points, let the i-th (i≤N) point of the workpiece point cloud be p i (x i ,y i ,z i ), x i ,y i ,z iare the coordinate values ​​of the i-th point in the point cloud, and the center point of the workpiece point cloud is c, p i The Euclidean distance from the point to the center point c is d i , calculate p respectively i The center score s i , point to feature point p k OffsetOf ki ,

[0019]

[0020] Of Ki =p i -p k

[0021] All points in the scene are used as training data according to the point coordinates, normalized point coordinates, instance labels, center scores, and feature point offsets.

[0022] Further, specifically, the improved FPCC network includes,

[0023] A feature point prediction module is added to the FPCC network, so that each point can estimate the coordinates of K feature points and then obtain the redundant feature point coordinates. After obtaining the redundant feature point coordinates, the Hough voting method is used to obtain the position of the feature point with the highest vote. Finally, the ICP point cloud registration method of the feature points is used to solve the pose of the workpiece. At the same time, the attention matrix is ​​added to the instance segmentation loss calculation of the FPCC network. According to the distance matrix D between the two points, an N×N same-instance attention matrix A is constructed. same and an N×N non-identical instance attention matrix A diff ,in,

[0024]

[0025] A sameij Indicates A same The element in the i-th row and j-th column of the matrix, A diffij Indicates A diff The element in the i-th row and j-th column of the matrix.

[0026] The present invention also proposes a device for solving the pose of disordered stacked parts with weak textures based on deep learning, comprising the following:

[0027] A point cloud file production module is used to scan parts in a plane to obtain scanned point clouds, and to segment and pre-process the scanned point clouds to obtain point cloud files of the front and back sides of several parts;

[0028] a training data generation module, configured to pre-generate a homogeneous transformation matrix of a batch of parts relative to a world coordinate system when the parts are stacked, and to generate training data by reading the point cloud file and transforming it to a corresponding stacking position based on the homogeneous transformation matrix;

[0029] The network model training module is used to train the improved FPCC network based on the training data and save the weight file of the trained network.

[0030] Specifically, the improved FPCC network adds a feature point prediction module on the basis of the original FPCC network, and adds an attention matrix to the instance segmentation loss calculation of the original FPCC network;

[0031] A feature point position determination module is used to obtain a scanned point cloud of a target part, perform point cloud preprocessing on the scanned point cloud of the target part, and then input the result into the weight file for prediction. The module then obtains the output center score, point cloud instance label, and feature point coordinates, and performs Hough voting to obtain the feature point positions of the target part.

[0032] The homogeneous transformation matrix calculation module is used to obtain the homogeneous transformation matrix corresponding to the target part through ICP point cloud registration of feature point positions and feature point templates.

[0033] The pose solving module is used to find the workpiece with a registration error lower than a preset threshold and output its pose based on the homogeneous transformation matrix corresponding to the target part.

[0034] The beneficial effects of the present invention are:

[0035] The present invention proposes a method for solving the pose of disordered stacked parts with weak textures based on deep learning. The method automatically generates labeled data for network training through computer technology, which greatly reduces the cost of manual labeling. In terms of network design, the effect of point cloud instance segmentation has been improved due to the addition of attention matrix calculation, and the added feature point prediction module enables the network to have the ability to solve the pose of parts. In the inference stage, the accurate position of the feature points is solved by the Hough voting method, which optimizes the error introduced by instance segmentation to a certain extent and increases the accuracy of the prediction. At the same time, the feature point registration method, compared with the direct point cloud registration method, not only ensures a certain degree of registration accuracy but also greatly reduces the time loss, and avoids the difference in algorithm performance due to different sensor sampling distances. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The above and other features of the present disclosure will become more apparent through a detailed description of the embodiments shown in conjunction with the accompanying drawings. The same reference numerals in the drawings of the present disclosure represent the same or similar elements. Obviously, the drawings described below are only some embodiments of the present disclosure. It is possible for a person skilled in the art to derive other drawings based on these drawings without inventive effort. In the drawings:

[0037] Figure 1 FIG2 is a schematic diagram of a part point cloud collected in one embodiment of the present invention;

[0038] Figure 2 FIG. 1 is a schematic diagram of part feature points collected in one embodiment of the present invention;

[0039] Figure 3 Schematic diagram showing a homogeneous transformation matrix of a part when collecting training data in one embodiment of the present invention;

[0040] Figure 4 FIG2 is a schematic diagram of a part point cloud in which all occluded points are removed by Open3D in one embodiment of the present invention;

[0041] Figure 5 Shown is a structural principle diagram of the improved FPCC network in the present invention;

[0042] Figure 6 FIG2 is a schematic diagram of a scanning point cloud of a target part in one embodiment of the present invention;

[0043] Figure 7 The figure shows the segmentation effect after segmenting the scanned point cloud of the target part in one embodiment of the present invention;

[0044] Figure 8 The figure shows the final effect after the scanning point cloud of the target part is processed in one embodiment of the present invention;

[0045] Figure 9 Shown is a flow chart of the method for solving the pose of disordered stacked parts with weak textures based on deep learning of the present invention;

[0046] Figure 10 Shown is a structural block diagram of the posture solving device for disordered stacked parts with weak textures based on deep learning of the present invention. DETAILED DESCRIPTION

[0047] The following will be combined with the embodiments and drawings to clearly and completely describe the concept, specific structure and technical effects of the present invention so as to fully understand the purpose, scheme and effect of the present invention. It should be noted that the embodiments and features in the embodiments of this application can be combined with each other unless there is a conflict. The same reference numerals used throughout the drawings indicate the same or similar parts.

[0048] Example 1, with reference to Figure 9 The present invention proposes a pose solving method for disordered stacked parts with weak textures based on deep learning, which includes the following:

[0049] Step 110: Scan the part in the plane to obtain a scan point cloud, and perform segmentation and preprocessing operations on the scan point cloud to obtain point cloud files of the front and back sides of the parts;

[0050] Step 120: pre-generate a batch of homogeneous transformation matrices of parts relative to the world coordinate system when the parts are stacked, and generate training data by reading the point cloud files and transforming them to corresponding stacking positions based on the homogeneous transformation matrices;

[0051] Step 130: train the improved FPCC network based on the training data, and save the weight file of the trained network.

[0052] Specifically, the improved FPCC network adds a feature point prediction module on the basis of the original FPCC network, and adds an attention matrix to the instance segmentation loss calculation of the original FPCC network;

[0053] Step 140: Obtain a scanned point cloud of the target part, perform point cloud preprocessing on the scanned point cloud of the target part, and then input it into the weight file for prediction to obtain output center scores, point cloud instance labels, and feature point coordinates, and perform Hough voting to obtain the feature point positions of the target part;

[0054] Step 150: Perform ICP point cloud registration with the feature point position and the feature point template to obtain the homogeneous transformation matrix corresponding to the target part.

[0055] Step 160 : Based on the homogeneous transformation matrix corresponding to the target part, find the workpiece whose registration error is lower than a preset threshold and output its pose.

[0056] In this embodiment 1, the problem of high cost of training data annotation is solved by automatically generating training data for the deep learning network through the CAD model of the workpiece. At the same time, a network structure is proposed that can effectively solve the pose problem of disordered stacked parts with weak textures.

[0057] As a preferred embodiment of the present invention, specifically, scanning a part in a plane to obtain a scanned point cloud, segmenting and preprocessing the scanned point cloud to obtain point cloud files of the front and back sides of several parts, including:

[0058] The scanning point cloud of the part in the plane is collected by a 3D camera, and the point cloud belonging to the part is segmented from the scanning point cloud to obtain the part point cloud. A total of K part point clouds are collected from both the front and back sides. At the same time, the CAD model of the part is sampled into a point cloud, and the FPS farthest point sampling algorithm is used to sample K part feature points. K point cloud files are constructed based on the part point cloud and the part feature points.

[0059] In this preferred embodiment, the part is placed on a plane, and the point cloud of the part is collected by a 3D camera to segment the point cloud belonging to the part. A total of 10 part point clouds are collected from both the front and back sides. Figure 1 As shown. Read the CAD model of the object, sample the CAD model into a point cloud, and use the FPS farthest point sampling algorithm to sample K (10) part feature points, as shown Figure 2 shown.

[0060] As a preferred embodiment of the present invention, specifically, pre-generating a homogeneous transformation matrix of parts relative to the world coordinate system when a batch of parts are stacked includes:

[0061] Place a rectangular material frame in the simulation environment, generate a batch of parts with random heights and random positions within the range above the material frame, and make the batch of parts fall freely into the material frame. The number of parts gradually increases from 3 to 50, and repeats 20 times. Figure 3 As shown in Figure 2. Record the homogeneous transformation matrix of each part relative to the world coordinate system when it is stationary.

[0062] As a preferred embodiment of the present invention, specifically, the process of generating training data includes:

[0063] Read the saved homogeneous transformation matrix of the parts relative to the world coordinate system one by one, randomly select the part point cloud and its corresponding part feature points in the collected point cloud file and transform them to the physical simulation position to obtain the transformed part point cloud and part feature points, and remove all blocked points in the transformed part point cloud and part feature points, such as Figure 4 As shown, assuming that the single part point cloud after removing the blocked points has I points, let the i-th point of the workpiece point cloud be p i (x i ,y i ,z i ), i≤N, x i ,y i ,z i are the coordinate values ​​of the i-th point in the point cloud, and the center point of the workpiece point cloud is c, p iThe Euclidean distance from the point to the center point c is d i , calculate p respectively i The center score s i , point to feature point p k OffsetOf ki ,

[0064]

[0065] Of Ki =p i -p k

[0066] All points in the scene are written one by one as training data according to the point coordinates (3 channels), the normalized coordinates (3 channels), the instance labels (1 channel), the center scores (1 channel), and the feature point offsets (K×3 channels).

[0067] As a preferred embodiment of the present invention, specifically, the improved FPCC network includes:

[0068] Designed as Figure 5 The network structure shown in the figure adds a feature point prediction module on the basis of the FPCC network, so that each point can estimate the coordinates of K feature points and then obtain the redundant feature point coordinates. After obtaining the redundant feature point coordinates, the Hough voting method is used to obtain the position of the feature point with the highest vote. Finally, the ICP point cloud registration method of the feature points is used to solve the accurate pose of the workpiece. At the same time, the attention matrix is ​​added to the instance segmentation loss calculation of the FPCC network, and an N×N same-instance attention matrix A is constructed according to the distance matrix D between the two points. same and an N×N non-identical instance attention matrix A diff ,in,

[0069]

[0070]

[0071] A sameij Indicates A same The element in the i-th row and j-th column of the matrix, A diffij Indicates A diff The element in the i-th row and j-th column of the matrix.

[0072] In this preferred embodiment, by constructing the network structure in this manner, the distance between the features of points in the same instance and the features of its center point in the feature space can be made smaller, while the distance between the features of points that are close but belong to different instances in the feature space can be made greater, effectively improving the instance segmentation performance of the network.

[0073] For the subsequent acquisition of the scanned point cloud of the target part, the scanned point cloud of the target part is pre-processed and then input into the weight file for prediction, and the output center score, point cloud instance label and feature point coordinates are obtained, and Hough voting is performed to obtain the feature point position of the target part;

[0074] The homogeneous transformation matrix corresponding to the target part is obtained by performing ICP point cloud registration with the feature point position and the feature point template.

[0075] The process of finding the workpiece with a registration error lower than a preset threshold and outputting its position and posture based on the homogeneous transformation matrix corresponding to the target part.

[0076] That is, the inference stage, the input is Figure 6 The point cloud data shown in the figure is used to obtain the center point score, feature vector and estimated feature point coordinates of each point through network inference. The point i with the highest center point score is selected, and then the points within the range of d (d is the maximum radius of the workpiece) around point i are removed from the candidate point set. The point with the highest score in the candidate point set is repeatedly selected until the candidate point set is empty. The distance between the selected feature point and the features of all other points in the feature space is calculated, and the points within the distance of 1 are selected as the instance point cloud of the center point. The point cloud instance segmentation effect is shown in the figure. Figure 7 As shown. Traverse the points belonging to the same instance and obtain the predicted feature point position of each point. For a single feature point, use the knn algorithm to solve the point with the most neighbors as the accurate position of the feature point. Finally, perform point cloud registration with the feature point template to complete the pose estimation of the part. Select the workpiece with higher confidence through the root mean square error of the point cloud. Transform the CAD model to the output homogeneous transformation matrix position so that the model point is black. The final effect is as shown below. Figure 8 shown.

[0077] Reference Figure 10 The present invention also proposes a device for solving the pose of disordered stacked parts with weak textures based on deep learning, comprising the following:

[0078] The point cloud file production module 100 is used to scan a part in a plane to obtain a scanned point cloud, and to segment and pre-process the scanned point cloud to obtain point cloud files of the front and back sides of several parts;

[0079] A training data generation module 200 is configured to pre-generate a homogeneous transformation matrix of a batch of parts relative to a world coordinate system when the parts are stacked, and to generate training data by reading the point cloud file and transforming it to a corresponding stacking position based on the homogeneous transformation matrix;

[0080] The network model training module 300 is used to train the improved FPCC network based on the training data and save the weight file of the trained network.

[0081] Specifically, the improved FPCC network adds a feature point prediction module on the basis of the original FPCC network, and adds an attention matrix to the instance segmentation loss calculation of the original FPCC network;

[0082] The feature point position determination module 400 is used to obtain a scanned point cloud of the target part, perform point cloud preprocessing on the scanned point cloud of the target part, and then input the result into the weight file for prediction. The module then obtains the output center score, point cloud instance label, and feature point coordinates, and performs Hough voting to obtain the feature point positions of the target part.

[0083] The homogeneous transformation matrix calculation module 500 is used to obtain the homogeneous transformation matrix corresponding to the target part by performing ICP point cloud registration with the feature point position and the feature point template.

[0084] The pose solving module 600 is used to find a workpiece with a registration error lower than a preset threshold based on the homogeneous transformation matrix corresponding to the target part and output its pose.

[0085] In addition, the functional modules in various embodiments of the present invention may be integrated into a single processing module, or each module may exist physically separately, or two or more modules may be integrated into a single module. The aforementioned integrated modules may be implemented in the form of hardware or software functional modules.

[0086] If the integrated module is implemented in the form of a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present invention implements all or part of the process in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or system that can carry the computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electrical carrier signal, telecommunication signal and software distribution medium, etc.

[0087] Although the present invention has been described in considerable detail and with particularity with respect to several described embodiments, it is not intended to be limited to any of these details or embodiments or any particular embodiment, but rather should be construed as providing a broad possible interpretation of these claims in view of the prior art by reference to the appended claims, thereby effectively encompassing the intended scope of the invention. In addition, the invention has been described above in terms of embodiments foreseen by the inventors for the purpose of providing a useful description, and those insubstantial modifications of the invention that are not currently foreseen may still represent equivalent modifications of the invention.

[0088] The above description is merely a preferred embodiment of the present invention. The present invention is not limited to the above-described embodiments. As long as the technical effects of the present invention are achieved by the same means, they shall fall within the scope of protection of the present invention. Within the scope of protection of the present invention, various modifications and variations of the technical solutions and / or implementation methods may be made.

Claims

1. A pose solving method for disordered stacked parts with weak textures based on deep learning, characterized by: These include: Scanning a part in a plane to obtain a scanned point cloud, and performing segmentation and preprocessing operations on the scanned point cloud to obtain point cloud files of the front and back sides of several parts; Pre-generating a batch of homogeneous transformation matrices of parts relative to a world coordinate system when the parts are stacked, and generating training data by reading the point cloud files and transforming them to corresponding stacking positions based on the homogeneous transformation matrices; Train the improved FPCC network based on the training data and save the weight file of the trained network. Specifically, the improved FPCC network adds a feature point prediction module on the basis of the original FPCC network, and adds an attention matrix to the instance segmentation loss calculation of the original FPCC network; Obtain a scanned point cloud of the target part, perform point cloud preprocessing on the scanned point cloud of the target part, and then input it into the weight file for prediction, obtain the output center score, point cloud instance label and feature point coordinates, and perform Hough voting to obtain the feature point position of the target part; The homogeneous transformation matrix corresponding to the target part is obtained by performing ICP point cloud registration with the feature point position and the feature point template. Based on the homogeneous transformation matrix corresponding to the target part, find the workpiece with a registration error lower than the preset threshold and output its pose; Specifically, the improved FPCC network includes, A feature point prediction module is added to the FPCC network, so that each point can estimate the coordinates of K feature points and then obtain the redundant feature point coordinates. After obtaining the redundant feature point coordinates, the Hough voting method is used to obtain the position of the feature point with the highest vote. Finally, the ICP point cloud registration method of the feature points is used to solve the accurate pose of the workpiece. At the same time, the attention matrix is ​​added to the instance segmentation loss calculation of the FPCC network, and an N×N instance attention matrix is ​​constructed according to the distance matrix D between the two points. and an N×N non-identical instance attention matrix ,in, ； ； ； ； ； ； express The element in the i-th row and j-th column of the matrix, express The element in the i-th row and j-th column of the matrix.

2. The method for solving the pose of disordered stacked parts with weak textures based on deep learning according to claim 1, characterized in that: Specifically, a part in a scanning plane is scanned to obtain a scanning point cloud, and the scanning point cloud is segmented and preprocessed to obtain point cloud files of the front and back sides of several parts, including: A 3D camera is used to capture a scanned point cloud of a part in a plane, and the point cloud belonging to the part is segmented from the scanned point cloud to obtain a part point cloud. A total of X part point clouds are collected from both the front and back sides. At the same time, a CAD model of the part is sampled into a point cloud, and K part feature points are sampled using the FPS farthest point sampling algorithm. X point cloud files are constructed based on the part point cloud and the part feature points.

3. The method for solving the pose of disordered stacked parts with weak textures based on deep learning according to claim 2, characterized in that: Specifically, a batch of parts are pre-generated to obtain the homogeneous transformation matrix of the parts relative to the world coordinate system when they are stacked, including: A rectangular material frame is placed in the simulation environment. A batch of parts with random heights and random positions are generated above the material frame, and the parts are allowed to fall freely into the material frame. The number of parts is gradually increased from 3 to 50, and this is repeated 20 times. The homogeneous transformation matrix of each part relative to the world coordinate system is recorded when it is stationary.

4. The method for solving the pose of disordered stacked parts with weak textures based on deep learning according to claim 3, characterized in that: Specifically, the process of generating training data includes: Read the homogeneous transformation matrix of the saved parts relative to the world coordinate system one by one, randomly select the part point cloud and its corresponding part feature points in the collected point cloud file and transform them to the physical simulation position to obtain the transformed part point cloud and part feature points, remove all the blocked points in the transformed part point cloud and part feature points, assume that a single part point cloud has N points, let the workpiece point cloud be The points are , , are the coordinate values ​​of the i-th point in the point cloud, and the center point of the workpiece point cloud is , Point to center The Euclidean distance is , respectively calculated The center score , point to feature point Offset , ； ； ； ； All points in the scene are used as training data according to the point coordinates, normalized point coordinates, instance labels, center scores, and feature point offsets.

5. A device for solving the pose of disordered stacked parts with weak textures based on deep learning, characterized in that: The method for solving the pose of disordered stacked parts with weak textures based on deep learning according to claim 1 is applied, and the device comprises the following: A point cloud file production module is used to scan parts in a plane to obtain scanned point clouds, and to segment and pre-process the scanned point clouds to obtain point cloud files of the front and back sides of several parts; a training data generation module, configured to pre-generate a homogeneous transformation matrix of a batch of parts relative to a world coordinate system when the parts are stacked, and to generate training data by reading the point cloud file and transforming it to a corresponding stacking position based on the homogeneous transformation matrix; The network model training module is used to train the improved FPCC network based on the training data and save the weight file of the trained network. Specifically, the improved FPCC network adds a feature point prediction module on the basis of the original FPCC network, and adds an attention matrix to the instance segmentation loss calculation of the original FPCC network; A feature point position determination module is used to obtain a scanned point cloud of a target part, perform point cloud preprocessing on the scanned point cloud of the target part, and then input the result into the weight file for prediction. The module then obtains the output center score, point cloud instance label, and feature point coordinates, and performs Hough voting to obtain the feature point positions of the target part. The homogeneous transformation matrix calculation module is used to obtain the homogeneous transformation matrix corresponding to the target part through ICP point cloud registration of feature point positions and feature point templates. The pose solving module is used to find the workpiece with a registration error lower than a preset threshold and output its pose based on the homogeneous transformation matrix corresponding to the target part.

Images