6D pose estimation method for valve body assembly

Abstract

The invention discloses a 6D pose estimation method for valve body assembly, and relates to the technical field of industrial manufacturing, and the method comprises the following steps: 1, employing an improved YOLOv11 network model to carry out the recognition of a valve body part image, and extracting a local ROI region; 2, performing three-dimensional point cloud mapping and point cloud segmentation on the local ROI to obtain a target point cloud; and step 3, mapping the CAD model of the valve body part into a source point cloud, and performing coarse registration and fine registration on the source point cloud and the target point cloud to obtain an estimated pose. According to the method provided by the invention, quick and accurate recognition and pose estimation of parts with different shapes, weak textures and scattered stacking can be realized, the actual application requirements are met, and the practicability and effectiveness of the method in the field of industrial automation are shown.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial manufacturing, and more particularly to a 6D pose estimation method for valve body assembly. BACKGROUND

[0002] With the rapid development of industrial automation and intelligent manufacturing, part recognition and pose estimation are increasingly widely used in manufacturing, especially in automated assembly, quality detection, robot vision and other fields. Traditional part recognition and pose estimation techniques mainly rely on the texture features of parts. However, in practical applications, many parts lack obvious texture features, or in specific working environments, the texture features of parts are not easy to capture, thus bringing great challenges to the recognition and pose estimation of weak texture parts.

[0003] Currently, with the rapid development of computer vision, machine learning and other technologies, it has become a research hotspot to solve the problem of weak texture part recognition and pose estimation through these advanced technologies. In particular, the application of deep learning technology provides a new solution for weak texture part recognition and pose estimation. However, how to effectively combine these advanced technologies to improve the accuracy and robustness of recognition and estimation is still a problem to be solved.

[0004] Therefore, how to improve the accuracy of part pose estimation is a problem that needs to be solved by those skilled in the art. SUMMARY

[0005] Therefore, the present application provides a 6D pose estimation method for valve body assembly. The part recognition and pose estimation method proposed by the present application can realize fast and accurate recognition and pose estimation of different shapes, weak texture and scattered stacked parts, meet the actual application requirements, and demonstrate its practicability and effectiveness in the field of industrial automation.

[0006] To achieve the above purpose, the present application adopts the following technical solutions:

[0007] A 6D pose estimation method for valve body assembly, comprising the following steps:

[0008] Step 1: using an improved YOLOv11 network model to recognize the valve body part image and extract the local ROI region;

[0009] Step 2: performing three-dimensional point cloud mapping and point cloud segmentation on the local ROI region to obtain the target point cloud;

[0010] Step 3: mapping the valve body part CAD model to the source point cloud, and performing coarse registration and fine registration on the source point cloud and the target point cloud to obtain the estimated pose.

[0011] Preferably, the improved YOLOv11 network model comprises a backbone network, a neck network and a detection head network, based on the YOLOv11 network structure, an improved wavelet transform convolution (ImprovedWTConv) module is integrated in the detection head network, and the detection head network outputs a local ROI region.

[0012] Preferably, the improved wavelet transform convolution module comprises an improved wavelet transform layer, a frequency band feature extraction layer, a frequency band cross attention layer and a feature reconstruction layer.

[0013] The improved wavelet transform layer adaptively outputs four wavelet transform outputs of the frequency bands by using an improved wavelet basis, and is expressed as:

[0014]

[0015] wherein, WT i (X) represents the wavelet transform output of the i-th frequency band, and the dimension is R C×H×W , C represents the number of channels, H represents the height of the feature map, and W represents the width of the feature map; X represents an input feature map; represents the i-th improved wavelet basis, and the dimension is R 1 ×3×3 , which comprises a 3x3 convolution kernel composed of 9 learnable parameters, is automatically optimized through back propagation, and automatically learns the optimal frequency domain decomposition mode; Conv2D represents a two-dimensional convolution operation; i∈{1,2,3,4} is a frequency band index, i=1 represents the LL frequency band, i.e. low frequency, i=2 represents the LH frequency band, i.e. horizontal high frequency, i=3 represents the HL frequency band, i.e. vertical high frequency, and i=4 represents the HH frequency band, i.e. diagonal high frequency;

[0016] The frequency band feature extraction layer independently extracts and reduces the dimension of the wavelet transform output of each frequency band to obtain four frequency band feature outputs, which are expressed as:

[0017] F i = BandConv i (WT i (X))

[0018] wherein, F i represents the feature output of the i-th frequency band, and the dimension is R C / 4×H×W , C / 4 represents the compressed number of channels; BandConv i represents the convolution module of the i-th frequency band, and the mapping function is R C →R C / 4 , which comprises a 1x1 convolution, a depth separable convolution and a 1x1 convolution connected in sequence;

[0019] The band cross attention layer first performs feature aggregation on the feature outputs of different frequency bands, then calculates channel attention weights, and finally performs weighted fusion on the feature outputs of different frequency bands to obtain a weighted feature map (F attended , is expressed as:

[0020] F combined = Concat([F1, F2, F3, F4]) ∈ R C×H×W

[0021] F global = AdaptiveAvgPool2D(F combined ) ∈ R C×1×1

[0022]

[0023] α = σ(Conv 7×7 [S avg ; S max ])

[0024] F attended = F combined ⊙ α

[0025] Wherein, F combined represents the aggregated feature, and the feature outputs of the four frequency bands are spliced in the channel dimension; F global represents a global feature vector, which is compressed to 1x1 in the spatial dimension by adaptive average pooling; AdaptiveAvgPool2D represents an adaptive average pooling operation; α temp represents an intermediate attention feature, with a dimension of R C / r×1×1 ; r represents a compression ratio (usually r = 16) for reducing computational complexity; Conv represents a convolution operation; α represents the final channel attention weight, with a dimension of R C ×1×1 ; σ represents a Sigmoid activation function, with an output range of [0, 1]; S avg represents the average value of the global feature vector; S max represents the maximum value of the global feature vector; SiLU represents a Swish activation function; F attended is obtained by multiplying F combined by the attention weight α, wherein ⊙ represents element-wise multiplication.

[0026] The feature reconstruction layer reconstructs the weighted feature map (F attended to obtain a reconstructed feature map F reconstructed , which is expressed as:

[0027] F reconstructed = SiLU(BN(Conv 1×1 (Fattended )))

[0028] wherein, Conv 1×1 represents a 1x1 point convolution operation for feature transformation and dimension adjustment; BN represents a batch normalization operation; SiLU represents a Swish activation function, SiLU(x)=x x (x); the reconstructed feature map F reconstructed and the reconstructed weighted feature map (F attended have the same dimension R C×H×W .

[0029] Preferably, the specific process of step 2 comprises:

[0030] Step 21: Collecting a depth image of the valve body part;

[0031] Step 22: Mapping the local ROI region to the depth image to obtain the depth value corresponding to each pixel in the local ROI region;

[0032] Step 23: Converting each pixel in the local ROI region into a 3D space point according to its 2D coordinates and corresponding depth value through the camera intrinsic matrix, to form a target source point cloud;

[0033] Step 24: Segmenting the target source point cloud according to a preset segmentation threshold to obtain a background point cloud and a part point cloud;

[0034] Step 25: Using the Euclidean clustering segmentation method to finely segment the part point cloud to obtain a target point cloud. Through fine segmentation, only the point cloud of the valve body part is obtained.

[0035] Preferably, the process of fine segmentation by the Euclidean clustering segmentation method in step 25 is as follows:

[0036] Step 251: Calculating the Euclidean distance between each point in the part point cloud and all other points;

[0037] Step 252: Comparing the Euclidean distance according to a preset clustering threshold to cluster the points with a Euclidean distance less than the clustering threshold into a target point cloud.

[0038] Preferably, the calculation formula of the Euclidean distance is as follows:

[0039]

[0040] wherein: D(p,q) represents the Euclidean distance between point p and point q, (p x , p y , p z ) and (q x , q y , q z ) are the coordinates of point p and point q in three-dimensional space, respectively.

[0041] Preferably, a sequence registration scheme is adopted in step 3, i.e. coarse registration is performed first and then fine registration is performed, and the specific process is as follows:

[0042] Step 31: the random sphere covering set (RSCS) algorithm is used for the target point cloud and the source point cloud respectively to perform spatial segmentation and feature aggregation processing, to generate a target point cloud superpoint set and a source point cloud superpoint set, and to realize coarse registration; the expression is as follows:

[0043] RSCS(P, r, c)→S

[0044] Wherein, P is the original point cloud, P = {p1, p2, …, pn}; r is the sphere radius parameter; c is the coverage rate or superpoint number parameter; S is the superpoint set, S = {s1, s2, …, sm}, m « n, m represents the number of point clouds in the superpoint set, and n represents the number of original point clouds. n n

[0045] Step 32: the RANSANC algorithm is used to fit the target point cloud superpoint set and the source point cloud superpoint set, to obtain an initial rotation matrix and an initial translation vector.

[0046] Step 33: the ICP algorithm is used to optimize and solve the target point cloud superpoint set and the source point cloud superpoint set according to the initial rotation matrix and the initial translation vector, to obtain an optimal rotation matrix and an optimal translation vector, and to realize fine registration.

[0047] Step 34: the estimated pose is calculated according to the optimal rotation matrix and the optimal translation vector.

[0048] Preferably, the process of step 32 for fitting is as follows:

[0049] Step 321: determine the model type M(x; θ) of the valve body part target model for fitting; select a point-to-point rigid transformation fitting type.

[0050] Step 322: randomly extract k point clouds from the source point cloud superpoint set as sample data x i , and select the point clouds closest to the sample data from the target point cloud superpoint set as target data y i , to construct a sample data set D = {x i , y i}; k should be sufficient for fitting the target model;

[0051] Step 323: use the sample data x i to fit the valve body part target model, to obtain model parameters θ opt , the model parameters including a rotation matrix and a translation vector, and the valve body part target model is denoted as M(x, θ opt ).​​

[0052] Step 324: filtering out the point cloud that meets the valve body part target model M(x, θ i ) from all target data y opt , to form an inner point set D in ; the expression of the filtering judgment is:

[0053] d(y i , M(x, θ opt )) < ε;

[0054] y i ≥ thr;

[0055] wherein d is a loss function, and ε is a tolerance parameter set in advance; the filtered points are inner points, and the inner point set D in formed thereby is called a consensus set; if the estimated inner points exceed a certain threshold thr, it indicates that the model parameters are relatively accurate, and the estimated inner points are also accurate to a high degree;

[0056] Step 325: re-fitting the valve body part target model using the inner point set D in , re-estimating the model parameters, and obtaining an optimized target model M(x, θ' opt ); the fitting effect m is used to measure the fitting effect of the model on the data;

[0057] Step 326: returning to step 321, iterating N times, updating the fitting effect each time, and obtaining the optimal fitting effect m best ; the model parameters corresponding to the optimal fitting effect are taken as the final output result, and the initial rotation matrix and the initial translation vector are obtained.

[0058] Preferably, the specific process of step 33 is as follows:

[0059] Step 331: taking the initial rotation matrix and the initial translation vector as initial estimates, finding the nearest points in the target point cloud superpoint set B corresponding to the points in the source point cloud superpoint set A, forming corresponding point pairs by pairing them two by two, and obtaining a corresponding point set;

[0060] Step 332: calculating the accurate rotation matrix and the accurate translation vector according to the corresponding point set;

[0061] Step 333: constructing a target function, minimizing the target function according to the accurate rotation matrix and the accurate translation vector, and performing k times of iterative calculations until the preset number of iterations is reached, to obtain the optimal rotation matrix and the optimal translation vector; the target function is represented as:

[0062]

[0063] wherein R * is the optimal rotation matrix, t * is the optimal translation vector, and C is the number of corresponding point pairs, represents the coordinates of the i-th point in the source point cloud superpoint set A at the k-th iteration, represents the coordinates of the i-th point in the target point cloud superpoint set B at the k-th iteration. The convergence condition is that the variation amplitude of the objective function value of two adjacent iterations is less than a pre-defined threshold.

[0064] Preferably, a particle swarm optimization algorithm is introduced in step 333 to globally optimize the objective function and find a more global optimal solution. For two sets of point clouds A and B, wherein A is the source point cloud superpoint set and B is the target point cloud superpoint set, in the process of each iteration, for each point a i in the source point cloud superpoint set A, find the nearest point b i in the target point cloud superpoint set B, calculate the weighted distance between them, and then use a global optimization function to optimize the following objective function:

[0065]

[0066] wherein R and T are the rotation matrix and the translation matrix respectively, and C represents the number of corresponding point pairs. By iteratively optimizing the above objective function until the termination condition is met or the maximum number of iterations is reached, the final optimal rotation matrix and optimal translation vector are obtained, and improved point cloud precise registration is achieved.

[0067] According to the technical solution described above, compared with the prior art, the present disclosure provides a 6D pose estimation method for valve assembly. First, the improved YOLOv11 algorithm is used to quickly identify and locate the ROI region of the part, and then the ROI region containing the part is aligned with the depth map to convert it into the corresponding three-dimensional point cloud. In the point cloud processing stage, the depth threshold segmentation and Euclidean clustering segmentation methods are used to separate the target part from the background and other interfering objects. In the pose estimation stage, the Random Sphere Cover Set (RSCS) algorithm + RANSAC algorithm are used for coarse registration, and then the improved ICP algorithm with adaptive weight mechanism and global optimization strategy is used for precise registration to obtain the final 6D pose of the part. This improved strategy significantly optimizes the point pair selection and registration process, enhancing the robustness and accuracy of the algorithm. The present disclosure has wide application prospects in the fields of industrial automation and intelligent manufacturing. Specific benefits include:

[0068] (1) Valve part identification: By improving the wavelet pooling module (Wavelet Pool) and replacing the traditional convolution operation in the YOLOv11 network, the identification of valve parts of different sizes is improved.

[0069] (2) Point cloud rough matching: the RANSAC algorithm is combined with the random sphere covering set (RSCS) algorithm to reduce the number of point pairs in the RANSAC registration and ICP optimization process, significantly improve the calculation efficiency, retain the local geometric features, and help improve the accuracy of feature matching;

[0070] (3) Point cloud fine matching: the global optimization algorithm particles swarm optimization (PSO) algorithm is introduced in the ICP iteration process to avoid the algorithm falling into a local optimal solution and find a more global optimal solution. BRIEF DESCRIPTION OF DRAWINGS

[0071] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced below. Obviously, the drawings in the following description are only embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of the provided drawings.

[0072] Figure 1 The 6D pose estimation method for valve body assembly provided by the present application is shown in the overall flowchart.

[0073] Figure 2 The improved YOLOv11 network model for identifying valve body parts provided by the present application is shown in the schematic diagram.

[0074] Figure 3 The mask diagram of the local ROI region provided by the present application is shown in the schematic diagram.

[0075] Figure 4 The original point cloud diagram cut out according to the mask diagram provided by the present application is shown in the schematic diagram.

[0076] Figure 5 The point cloud processing effect provided by the present application is shown in the schematic diagram.

[0077] Figure 6 The pose estimation effect based on point cloud registration provided by the present application is shown in the schematic diagram.

[0078] Figure 7 The improved YOLOv11 network model structure provided by the present application is shown in the schematic diagram.

[0079] Figure 8 The ImprovedWTConv module structure provided by the present application is shown in the schematic diagram. DETAILED DESCRIPTION

[0080] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of the present application.

[0081] The embodiments of the present application disclose a 6D pose estimation method for valve body assembly, as shown in the formula (I), comprising the following steps: Figure 1

[0082] S1: An improved YOLOv11 network model is used to identify the valve body part image and extract the local ROI region;

[0083] S2: The local ROI region is mapped and segmented by three-dimensional point cloud to obtain the target point cloud;

[0084] S21: The depth image of the valve body part is collected;

[0085] S22: The local ROI region is mapped to the depth image to obtain the depth value corresponding to each pixel in the local ROI region;

[0086] S23: The 2D coordinates and corresponding depth values of each pixel in the local ROI region are converted into 3D space points through the camera intrinsic matrix to form the target source point cloud; the target source point cloud contains the information of the valve body part itself, the information of the environment background and the information of other interfering objects;

[0087] S24: The target source point cloud is segmented by background point cloud according to the preset segmentation threshold to obtain the background point cloud and the part point cloud;

[0088] S25: The part point cloud is finely segmented by using the Euclidean clustering segmentation method to obtain the target point cloud;

[0089] S251: The Euclidean distance between each point in the part point cloud and all other points is calculated;

[0090] S252: The Euclidean distance is compared and judged according to the preset clustering threshold, the points with the Euclidean distance less than the clustering threshold are clustered as the same object point cloud, and the target point cloud is obtained;

[0091] ​S3: Map the valve body part CAD model to the source point cloud, perform coarse registration and fine registration on the source point cloud and the target point cloud to obtain the estimated pose; adopt a sequential registration scheme, that is, perform coarse registration first and then fine registration, so as to achieve point cloud registration to find the best correspondence between the two sets of point clouds and determine the pose of one object relative to the other. The purpose of coarse registration is to quickly obtain the approximate correspondence between the two sets of point clouds, and fine registration is to further refine the correspondence on the basis of coarse registration to achieve higher registration accuracy; Figure 6 As shown in the figure, during the coarse registration, the source point cloud super point set and the target point cloud super point set are screened out to form the initial point cloud. The initial point cloud is fitted using the RANSANC algorithm, and then the ICP algorithm is used for optimization and solution. The specific process is as follows:

[0092] S31: The target point cloud and the source point cloud are respectively subjected to spatial segmentation and feature aggregation processing using the random sphere covering set (RSCS) algorithm to generate the target point cloud super point set and the source point cloud super point set; the expression is:

[0093] RSCS(P,r,c)→S

[0094] Among them, P is the original valve body part point cloud, P={p1,p2,…,p n}; r is the sphere radius parameter; c is the coverage or superpoint number parameter; S is the superpoint set, S={s1,s2,…,s n},m<<n, m represents the number of points in the point cloud in the superpoint set, and n represents the number of points in the original point cloud; the random sphere covering set algorithm samples the point cloud through random spherical areas, converting the dense point cloud data in three-dimensional space into a sparse superpoint (SuperPoints) representation while retaining the key geometric features of the point cloud, thus realizing feature representation and dimensionality reduction processing of the point cloud;

[0095] S32: Use the RANSANC algorithm to fit the target point cloud super point set and the source point cloud super point set to obtain the initial rotation matrix and initial translation vector;

[0096] S321: Determine the model type M(x;θ) of the target model to be fitted; select the point-to-point rigid transformation fitting type;

[0097] S322: Randomly extract k point clouds from the source point cloud superpoint set as sample data x i , select the point cloud closest to the sample data from the target point cloud super point set as the target data y i , construct a sample data set D = {x i ,y i}; k should be sufficient to fit the target model;

[0098] S323: Using the extracted sample data xi fitting target model to obtain model parameters θ opt , the target model is denoted as M(x, θ opt ); the model parameters include a rotation matrix and a translation vector;

[0099] S324: judging all data y i that meet the target model to form an inlier set; the expression for judging is:

[0100] d(y i , M(x, θ opt )) < ε;

[0101] y i ≥ thr;

[0102] wherein d is a loss function, ε is a tolerance parameter set in advance; the screened points are inliers, and the inlier set is called a consensus set, i.e., a consistent set; if the estimated inliers exceed a certain threshold thr, it indicates that the model parameters are relatively accurate, and the estimated inliers are also accurate;

[0103] S325: using the inlier set D in to re-fit the target model of the valve body part, re-estimate the model parameters, and obtain an optimized target model M(x, θ' opt ), which is the required target model excluding the interference of outliers; using the optimized target model to estimate the fitting effect m, which is a measurement index for measuring the fitting effect of the model on the data;

[0104] S326: returning to S321, iterating N times, updating the fitting effect each time to obtain the optimal fitting effect m best , the model parameters corresponding to the optimal fitting effect are the final output results, and the initial rotation matrix and the initial translation vector are obtained;

[0105] S33: using the ICP algorithm to optimize and solve the target point cloud superpoint set and the source point cloud superpoint set according to the initial rotation matrix and the initial translation vector, to obtain the optimal rotation matrix and the optimal translation vector;

[0106] S331: taking the initial rotation matrix and the initial translation vector as initial estimates, finding the nearest points in the target point cloud superpoint set B corresponding to the points in the source point cloud superpoint set A, and combining the corresponding points two by two to form a corresponding point set;

[0107] S332: calculating the accurate rotation matrix and the accurate translation vector according to the corresponding point set;

[0108] S333: Construct a target function, minimize the target function according to the accurate rotation matrix and the accurate translation vector, perform k times of iteration calculation until the convergence condition is met to obtain a preset number of iterations, obtain an optimal rotation matrix and an optimal translation vector; the target function is represented as:

[0109]

[0110] wherein R * is the optimal rotation matrix, t * is the optimal translation vector, C is the number of corresponding point pairs, represents the coordinates of the i-th point in the source point cloud superpoint set A at the k-th iteration, is the coordinates of the i-th point in the target point cloud superpoint set B at the k-th iteration. The convergence condition is that the variation amplitude of the target function value of two adjacent iterations is less than a pre-given threshold value;

[0111] S34: Calculate the estimated pose according to the optimal rotation matrix and the optimal translation vector.

[0112] Further, the calculation formula of the Euclidean distance is:

[0113]

[0114] wherein D(p, q) represents the Euclidean distance between points p and q, (p x , p y , p z ) and (q x , q y , q z ) are the coordinates of points p and q in the three-dimensional space.

[0115] Further, the original ICP algorithm realizes registration by finding the nearest point pairs between two sets of point clouds through iteration and minimizing the distance between the point pairs, however, this process is often susceptible to abnormal points and highly sensitive to the initial pose estimation, therefore, the particle swarm optimization algorithm is introduced in step 333 for global optimization to avoid the ICP algorithm falling into a local optimal solution and find a more global optimal solution; for two sets of point clouds A and B, wherein A is a source point cloud superpoint set and B is a target point cloud superpoint set, in the process of each iteration, for each point a i in the source point cloud superpoint set A, find the nearest point b i in the target point cloud superpoint set B, calculate the weighted distance sum between them, and then optimize the following target function by using a global optimization function:

[0116]

[0117] Wherein: R and T are rotation matrix and translation matrix respectively; C represents the number of corresponding point pairs; through iterative optimization of the above objective function until the termination condition is met or the maximum iteration number is reached, the final optimal rotation matrix and optimal translation vector are obtained, and improved point cloud precise registration is realized.

[0118] On the other hand, in one specific embodiment, as shown in Figure 7 The improved YOLOv11 network model includes a backbone network, a neck network and a detection head network, and based on the YOLOv11 network structure, the ImprovedWTConv module is integrated in the detection head network, and the detection head network outputs the local ROI region.

[0119] On the other hand, in one specific embodiment, as shown in Figure 8 The learnable wavelet transform (ImprovedWTConv) module includes a learnable wavelet transform layer, a frequency band feature extraction layer, a frequency band cross attention layer and a feature reconstruction layer.

[0120] The traditional discrete wavelet transform (DWT) decomposes the signal into four subbands:

[0121]

[0122] Wherein, h[n] and g[n] are low-pass and high-pass filter coefficients respectively; LL(x,y) is the low-frequency-low-frequency subband, which retains the main structural information of the image; LH(x,y) is the low-frequency-high-frequency subband, which captures the horizontal edge and details; HL:(x,y) is the high-frequency-low-frequency subband, which captures the vertical edge and details; HH(x,y) is the high-frequency-high-frequency subband, which captures the diagonal edge and texture; x,y are the coordinate positions of the output subband; m,n are the relative position indexes within the filter window; is the normalization factor to ensure energy conservation; parameter initialization: LL band (low frequency): initialized as the mean filter LH band (horizontal high frequency): initialized as the horizontal edge detector HL band (vertical high frequency): initialized as the vertical edge detector HH band (diagonal high frequency): initialized as the Laplacian operator [0,-1,0;-1,4,-1;0,-1,0]; frequency domain characteristic analysis: the LL band captures low-frequency information, retaining the overall shape and contour of the part; the LH / HL band extracts horizontal / vertical edge features, highlighting the part boundary; the HH band detects corner points and texture details for precise positioning.

[0123] The learnable wavelet transform layer adopts a learnable wavelet basis to adaptively output four frequency band wavelet transform outputs, represented as:

[0124]

[0125] where WT i (X) denotes the wavelet transform output of the i-th frequency band, with dimension R C×H×W , C denotes the number of channels, H denotes the height of feature map, and W denotes the width of feature map; X denotes the input feature map; denotes the i-th learnable wavelet basis, with dimension R 1 ×3×3 , which contains 9 learnable parameters to form a 3x3 convolution kernel, is automatically optimized and learns the optimal frequency domain decomposition mode through backpropagation; Conv2D denotes a two-dimensional convolution operation; i ∈ {1, 2, 3, 4} is the frequency band index, i = 1 denotes the LL frequency band, i.e., the low frequency, i = 2 denotes the LH frequency band, i.e., the horizontal high frequency, i = 3 denotes the HL frequency band, i.e., the vertical high frequency, and i = 4 denotes the HH frequency band, i.e., the diagonal high frequency;

[0126] The frequency band feature extraction layer independently extracts features and reduces the dimension of the wavelet transform output of each frequency band, obtaining the feature output of the four frequency bands, denoted as:

[0127] F i = BandConv i (WT i (X))

[0128] where F i denotes the feature output of the i-th frequency band, with dimension R C / 4×H×W , and C / 4 denotes the compressed number of channels, reducing memory consumption; BandConv i denotes the convolution module of the i-th frequency band, which is a mapping function R C → R C / 4 , and the convolution module includes a 1x1 convolution, a depth separable convolution, and a 1x1 convolution connected in sequence. The channel compression ratio of this convolution module is 4:1, which can reduce the computational complexity;

[0129] The frequency band cross-attention layer adaptively learns the importance weights of different frequency bands based on the frequency band cross-attention mechanism, captures the dependency between channels through the "compression-activation" mechanism, highlights important feature channels, and suppresses redundant channels. First, the feature outputs of different frequency bands are aggregated, then the channel attention weights are calculated, and finally the feature outputs of different frequency bands are weighted and fused to obtain the weighted feature map F attended , denoted as:

[0130] F combined = Concat([F1,F2,F3,F4]) ∈ R C×H×W

[0131] F global=AdaptiveAvgPool2D(F combined )∈R C×1×1

[0132]

[0133] α=σ(Conv 7×7 [S avg ;S max ])

[0134] F attended =F combined ⊙α

[0135] Among them, F combined Represents the aggregated features, which concatenates the feature outputs of the four frequency bands in the channel dimension; F global Represents the global feature vector, and the spatial dimension is compressed to 1×1 through adaptive average pooling; AdaptiveAvgPool2D represents the adaptive average pooling operation; α temp Represents the intermediate attention feature, with dimension R C / r×1×1 ; r represents the compression ratio (usually r = 16), which is used to reduce computational complexity; Conv represents the convolution operation; α represents the final channel attention weight, with a dimension of R C ×1×1 ;σ represents the Sigmoid activation function, the output range is [0,1]; S avg represents the average value of the global eigenvector; S max Represents the maximum value of the global eigenvector; SiLU represents the Swish activation function; F attended is by putting F combined It is obtained by multiplying the attention weight α, where ⊙ represents element-by-element multiplication; S avg and S max It is the statistical information extracted from the global feature vector and used to calculate the attention weight, helping the model to better focus on important feature channels and suppress redundant channels, thereby improving the quality of feature representation;

[0136] Feature reconstruction layer, weighted feature map (F attended Reconstruct and obtain the reconstructed feature map F reconstructed , expressed as:

[0137] F reconstructed =SiLU(BN(Conv 1×1 (F attended )))

[0138] Among them, Conv 1×1denotes a 1x1 point convolution operation for feature transformation and dimension adjustment; BN denotes a Batch Normalization operation to realize distribution standardization and stabilize the training process; SiLU denotes a Swish activation function, SiLU(x) = x x σ(x), which provides a smooth nonlinear transformation; the reconstructed feature map F reconstructed and the reconstructed weighted feature map (F attended The dimensions of F C×H×W The feature reconstruction layer extracts redundant information and retains key features, thereby further optimizing the attention-weighted features.

[0139] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other. For the device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method part.

[0140] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to the embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A 6D pose estimation method for valve body assembly, characterized in that: The following steps are involved: Step 1: Use the improved YOLOv11 network model to identify the valve body part image and extract the local ROI area; Step 2: Perform 3D point cloud mapping and point cloud segmentation on the local ROI area to obtain the target point cloud; Step 3: Map the valve body part CAD model to a source point cloud, perform coarse and fine registration on the source point cloud and the target point cloud to obtain the estimated pose.

2. A 6D pose estimation method for valve body assembly according to claim 1, characterized in that: The improved YOLOv11 network model includes a backbone network, a neck network, and a detection head network. Based on the YOLOv11 network structure, a learnable wavelet transform module is integrated into the detection head network, and the detection head network outputs a local ROI area.

3. A 6D pose estimation method for valve body assembly according to claim 2, characterized in that: The learnable wavelet transform module includes a learnable wavelet transform layer, a frequency band feature extraction layer, a frequency band cross attention layer, and a feature reconstruction layer; The learnable wavelet transform layer uses a learnable wavelet basis to adaptively output the wavelet transform output of 4 frequency bands, which is expressed as: Among them, WT i (X) represents the wavelet transform output of the i-th frequency band; X represents the input feature map; Represents the i-th learnable wavelet basis; Conv2D represents a two-dimensional convolution operation; i∈{1,2,3,4} is a frequency band index, i=1 represents the LL band, i=2 represents the LH band, i=3 represents the HL band, and i=4 represents the HH band; The frequency band feature extraction layer performs independent feature extraction and dimensionality reduction on the wavelet transform output of each frequency band to obtain the feature output of four frequency bands, which can be expressed as: F i =BandConv i (WT i (X)) Among them, F i Represents the feature output of the i-th frequency band; BandConv i represents the convolution module of the i-th frequency band, which includes a 1×1 convolution, a depth-wise separable convolution, and a 1×1 convolution connected in sequence; The frequency band cross attention layer performs weighted fusion on the feature outputs of different frequency bands to obtain a weighted feature map (F attended , expressed as: F combined =Concat([F1,F2,F3,F4])∈R C×H×W F global =AdaptiveAvgPool2D(F combined )∈R C×1×1 α=σ(Conv 7×7 [S avg ;S max ]) F attended =F combined ⊙α Among them, F combined represents the aggregate feature, F global Represents the global feature vector; AdaptiveAvgPool2D represents the adaptive average pooling operation; α temp represents the intermediate attention feature; r represents the compression ratio; Conv represents the convolution operation; α represents the final channel attention weight; σ represents the Sigmoid activation function; S avg represents the average value of the global eigenvector; S max Represents the maximum value of the global eigenvector; SiLU represents the Swish activation function; R C×H×W Indicates dimension, C indicates the number of channels, H indicates the height of the feature map, and W indicates the width of the feature map; Feature reconstruction layer, weighted feature map (F attended Reconstruct and obtain the reconstructed feature map F reconstructed , expressed as: F reconstructed =SiLU(BN(Conv 1×1 (F attended ))) Among them, Conv 1×1 Represents a 1×1 point convolution operation; BN represents a batch normalization operation; SiLU represents a Swish activation function, SiLU(x)=x×σ(x).

4. The 6D pose estimation method for valve body assembly according to claim 1, characterized in that: The specific process of step 2 includes: Step 21: Acquire a depth image of the valve body part; Step 22: Map the local ROI area to the depth image and obtain the depth value corresponding to each pixel in the local ROI area; Step 23: The 2D coordinates and corresponding depth values ​​of each pixel in the local ROI area are converted into 3D space points through the camera intrinsic parameter matrix to form the target source point cloud; Step 24: Segment the target source point cloud into a background point cloud according to a preset segmentation threshold to obtain a background point cloud and a part point cloud; Step 25: Use the Euclidean clustering segmentation method to finely segment the part point cloud to obtain the target point cloud.

5. The 6D pose estimation method for valve body assembly according to claim 4, characterized in that: The process of fine segmentation using the Euclidean clustering segmentation method in step 25 is as follows: Step 251: Calculate the Euclidean distance between each point in the part point cloud and all other points; Step 252: Compare and judge the Euclidean distance according to a preset clustering threshold, and cluster the points with Euclidean distance less than the clustering threshold into the target point cloud.

6. The 6D pose estimation method for valve body assembly according to claim 1, characterized in that: In step 3, a sequential registration scheme is used, first performing coarse registration and then fine registration. The specific process is as follows: Step 31: Use the random sphere covering set algorithm to perform spatial segmentation and feature aggregation on the target point cloud and the source point cloud, respectively, to generate the target point cloud super point set and the source point cloud super point set; the expression is: RSCS(P,r,c)→S Among them, P is the original valve body part point cloud, P={p1,p2,…,p n }; r is the sphere radius parameter; c is the coverage or superpoint number parameter; S is the superpoint set, S={s1,s2,…,s n },m<<n, m represents the number of points in the point cloud in the superpoint set, and n represents the number of points in the original point cloud; Step 32: Use the RANSANC algorithm to fit the target point cloud super point set and the source point cloud super point set to obtain the initial rotation matrix and initial translation vector; Step 33: Using the ICP algorithm to optimize the target point cloud super point set and the source point cloud super point set according to the initial rotation matrix and the initial translation vector, and obtain the optimal rotation matrix and the optimal translation vector; Step 34: Calculate the estimated pose based on the optimal rotation matrix and the optimal translation vector.

7. The 6D pose estimation method for valve body assembly according to claim 6, characterized in that: The fitting process in step 32 is as follows: Step 321: Determine the model type M(x;θ) of the fitted valve body part target model; Step 322: Randomly extract k point clouds from the source point cloud superpoint set as sample data x i , select the point cloud closest to the sample data from the target point cloud super point set as the target data y i , construct a sample data set D = {x i ,y i };; Step 323: Using sample data x i Fit the target model of the valve body parts to obtain the model parameters θ opt , the model parameters include the rotation matrix and translation vector, and the target model of the valve body part is recorded as M(x,θ opt ); Step 324: From all target data y i Select the target model M(x,θ) that meets the requirements of valve body parts opt ) point cloud, forming an internal point set; the judgment expression is: d(y i ,M(x,θ opt ))<e; and i ≥thr; Among them, d is the loss function, ε is the tolerance parameter set in advance; thr is the preset point cloud number threshold; Step 325: Use the inner point set to refit the valve body part target model, re-estimate the model parameters, and obtain the optimized target model M(x,θ′ opt ), use the optimized target model to estimate the fitting effect m; Step 326: Return to step 321 and iterate N times, updating the fitting effect each time to obtain the optimal fitting effect m best , the model parameters corresponding to the optimal fitting effect are taken as the final output results to obtain the initial rotation matrix and initial translation vector.

8. The 6D pose estimation method for valve body assembly according to claim 6, characterized in that: The specific process of step 33 is: Step 331: Using the initial rotation matrix and the initial translation vector as initial estimates, find the closest points in the target point cloud superpoint set that correspond to the points in the source point cloud superpoint set, and combine them into corresponding point pairs to obtain a corresponding point set. Step 332: Calculate the precise rotation matrix and precise translation vector based on the corresponding point set; Step 333: Construct an objective function and minimize the objective function based on the precise rotation matrix and the precise translation vector until the convergence condition is met and the preset number of iterations is reached to obtain the optimal rotation matrix, optimal translation vector and matching point pair; the objective function is expressed as: Among them, R * is the optimal rotation matrix, t * is the optimal translation vector, C is the number of corresponding point pairs, Represents the coordinates of the i-th point in the source point cloud superpoint set at the k-th iteration, is the coordinate of the i-th point in the target point cloud superpoint set at the k-th iteration.

9. The 6D pose estimation method for valve body assembly according to claim 8, characterized in that: In step 333, a particle swarm optimization algorithm is introduced to perform global optimization on the objective function.

Images