Identification control system for smart wearable device assembly

Abstract

This invention discloses an identification and control system for assembling smart wearable devices, relating to the field of control technology, including an identification module, a reconstruction module, and a planning module. The identification module captures and evaluates a set of part images, inputting the evaluated part images into a target detection model to identify the position and category of each part in the image. The reconstruction module assigns part numbers according to the assembly order, detects and matches key points of each part in different part images using a matching algorithm, constructs a projection equation set based on the camera's extrinsic parameters, and uses a nonlinear optimization algorithm to obtain the optimal world coordinates of the key points and fits them to establish a bounding box for each part. The planning module constructs a directed graph based on the occlusion of the bounding box, uses a joint optimization algorithm to iteratively solve for the optimal operation sequence, and combines it with an inverse kinematics algorithm to control the actuator, achieving high-precision part identification and efficient part assembly control during part stacking.

Description

Technical Field

[0001] This invention relates to the field of control technology, and more specifically to an identification and control system for assembling smart wearable devices. Background Technology

[0002] As portable terminals integrating multiple cutting-edge technologies, the efficient and precise assembly of the precision components in smart wearable devices directly impacts product performance and quality. Identification and control systems, as a core technology in the assembly process, enable rapid component positioning, accurate identification, and intelligent assembly planning, thus helping to improve production efficiency, reduce labor costs, and ensure product consistency.

[0003] However, existing identification and control systems for smart wearable devices have significant limitations when faced with scenarios involving stacked parts and complex spatial layouts. Traditional methods struggle to accurately identify parts in a stacked state and cannot effectively plan the grasping and assembly sequence of parts, resulting in inefficient identification and control systems and frequent assembly failures in stacked parts environments. Summary of the Invention

[0004] This invention addresses the shortcomings of existing technologies by proposing an identification and control system for the assembly of smart wearable devices, enabling accurate part identification and actuator path planning control in a stacked state.

[0005] The technical solution to achieve the purpose of this invention is as follows:

[0006] An identification and control system for assembling smart wearable devices includes an identification module, a reconstruction module, and a planning module;

[0007] The identification module captures a collection of parts images. The image set is evaluated using entropy and edge sharpness, and decisions are made by adjusting parameters through an adaptive control algorithm and re-shooting or updating the part image set. Input the target detection model, and preprocess the i-th part image I. i And by extracting and fusing features, generating and filtering candidate boxes, the i-th part image I is obtained. i Parts annotation set in, and χ i,k These are the i-th part drawings I. i The frame center pixel coordinates, frame width, frame height, and part category of the k-th part, where K is the total number of parts;

[0008] The reconstruction module generates assembly numbers for K parts based on their assembly order, detects and matches key points of each part in different part images using a matching algorithm, and uses camera extrinsic parameters. Construct a system of projection equations, use a nonlinear optimization algorithm to obtain the optimal world coordinates of the key points, and fit them to establish the bounding box of each part;

[0009] The planning module performs occlusion detection and occlusion decision on the bounding boxes of K parts, constructs a directed graph of K parts, and uses a joint optimization algorithm to encode the operation sequence q into a chromosome. Constraints are constructed to meet assembly and topology requirements, based on the total length f of the movement path. L (q) and total angle adjustment f θ (q) Construct an objective function f(q), and perform mutation operations and constraint fine-tuning on the chromosome with the goal of minimizing the objective function f(q), iteratively generating the optimal operation sequence q. * And determine the optimal pose sequence b * The joint parameter sequence is solved using an inverse kinematics algorithm. And deliver the actuator.

[0010] Furthermore, the recognition module includes a data acquisition and evaluation unit and a target recognition unit;

[0011] The data acquisition and evaluation unit uses a PLC to synchronously control six industrial cameras to capture images and obtain part atlases. And perform drawing set evaluation, calculate the I value of the i-th part drawing. i The entropy value H i and edge sharpness E i The weighted average is then used to obtain a comprehensive score S, which is then compared with a score threshold S. th Select delivery parts atlas The target recognition unit or an adaptive control algorithm is used to adjust the parameters and re-shoot;

[0012] The target recognition unit will identify the i-th part drawing I. i The input target detection model extracts features through normalization and multi-scale convolution, and then fuses them to generate the i-th part image I. i Enhanced features Enhancement features are processed separately using parallel convolutions. And splice them together to generate the i-th part drawing I. i Predicted tensor O i Based on the predicted tensor O i For each candidate box, the box offset vector Δ is used to calculate the corresponding box center pixel coordinates p. px,c Frame width w px and frame height h px Combined with the prediction tensor O i The confidence α and category score vector S for each candidate box c Calculate the corresponding box score S F Then, select and retain K candidate boxes containing the i-th part to obtain the i-th part image I. i Parts annotation set

[0013] Furthermore, the data collection and evaluation unit collects and evaluates the parts drawing sets. The image set evaluation includes the following steps:

[0014] Based on the probability P of the occurrence of a pixel with a grayscale value of τ in the i-th part image i (τ) Calculate the entropy value H of the i-th part drawing. i ;

[0015] The Sobel operator is used to calculate the i-th part drawing I. i Lateral gradient map in the u direction and longitudinal gradient plot in the v direction And sum the gradient magnitudes to obtain the i-th part drawing I. i Edge sharpness E i The u and v directions are the coordinate axes of the pixel coordinate system;

[0016] The i-th part drawing I i The entropy value H i and edge sharpness E i We take the weighted sum of the minimum and maximum normalized results to obtain the i-th part image I. i Score S i The average score of the six part drawings is used as the comprehensive score S, and compared with the score threshold S. th Compare;

[0017] If greater than or equal to the score threshold S th Then deliver the parts drawing set. If the target recognition unit is less than the score threshold S th An adaptive control algorithm was used to adjust parameters and re-control six industrial cameras for shooting. Parameters included light source intensity I. L and ISO sensitivity G I .

[0018] Furthermore, the adaptive control algorithm includes the following steps:

[0019] The average entropy value of the 6 part drawings With the preset entropy value H pre The difference and entropy correction coefficient β H The product of the light source intensity during this shooting. The light source intensity I is obtained by superimposing the results for the next shot. L ;

[0020] The average image noise level of the 6 part drawings Compared with the preset image noise level δ pre The difference and the noise level correction factor β δThe product of the two is the ISO sensitivity of this shot. By superimposing the values, we can obtain the ISO sensitivity (G) for the next shot. I The noise level of each part image is proportional to the proportion of the high-frequency part after the Fourier transform of the part image.

[0021] Furthermore, the object detection model includes a backbone network, a neck network, branch detection heads, a bounding box decoder, and a score filter;

[0022] Backbone network normalized i-th component diagram I i The corresponding feature map F is generated through convolution and max pooling. i The first-level feature map is obtained through three consecutive downsampling operations. Secondary feature map and third-level feature maps

[0023] The neck network will have three levels of feature maps The upsampling results and secondary feature maps Concatenation, using point-by-point convolution to generate intermediate features intermediate features The upsampling results and the first-level feature map F i (1) The concatenation process, employing pointwise convolution and self-attention mechanisms, is used to fuse and enhance features, generating augmented features.

[0024] Branch detection head in enhancing features At each feature point, three candidate boxes of different sizes are generated, and convolution is performed within each candidate box. The convolution results within each candidate box are then subjected to three different non-linear mappings, outputting the box offset vector Δ, confidence α, and class score vector S for each candidate box. c And stitch them together into a predicted vector, and use the i-th part drawing I i The predicted vectors of each candidate box are concatenated to form the i-th part image I. i Predicted tensor O i , where the category score vector S c Record the category score of the candidate boxes for each part category;

[0025] The bounding box decoder starts from the i-th part drawing I. i Predicted tensor O i The offset vector Δ of each candidate box is extracted sequentially, and the center of the candidate box is corrected based on the offset vector Δ to enhance the features. pixel coordinates (u) c ,v c ), preset width w px ,pre and preset height hpx,pre Generate the pixel coordinates p of the center of each candidate box. px,c Frame width w px and frame height h px ;

[0026] The scoring filter extracts the S from the category score vector of each candidate box. c Category score of the c-th part category The confidence level α is normalized using the Sigmoid function and then multiplied to obtain the single-class bounding box score for each candidate box with respect to the c-th part category. Compare the maximum single-class bounding box score with the bounding box score threshold S. F,th If it is less than the box score threshold S F,th If the candidate box is greater than or equal to the box score threshold S, then discard the candidate box. F,th If the candidate box is selected, it is retained and labeled according to the part category corresponding to the highest single-class box score, and the i-th part image I is generated. i Parts annotation set

[0027] Furthermore, the reconstruction module includes a retrieval numbering unit and a three-dimensional reconstruction unit;

[0028] The retrieval number unit is based on the i-th part drawing I. i The part category χ of the k-th part i,k Match the assembly number n in the parts assembly manual and replace the i-th part drawing I. i Given the original number k of the k-th part, construct the i-th part diagram I. i Assembly parts annotation set in, and These are the assembly diagrams for part number n in the i-th part drawing. i The pixel coordinates of the center of the frame, the width of the frame, and the height of the frame;

[0029] The 3D reconstruction unit uses a matching algorithm to detect key points of the assembly part with assembly number n in six part images and generates key point descriptors. A FLANN matcher is used to match key points in different part images, and J matching point groups are selected based on a random sampling consensus algorithm. These matching points are then combined with the i-th part image I. i Corresponding camera extrinsic parameter T i Construct the world coordinates of the key points reflecting J key points and the i-th part drawing I. i The i-th projection equation is used to map and associate the pixel coordinates of keypoints. This equation is then integrated to construct a system of projection equations. A nonlinear optimization algorithm is employed to solve this system, minimizing the reprojection error as the optimization objective, to obtain the world coordinate set of the keypoints. Then, a bounding box Sq(n) is fitted to assemble the parts with assembly number n, and a bounding box set is constructed. The j-th matching point group includes key points that have matching relationships and originate from the six part drawings. The world coordinates of the j-th key point of the assembly part numbered n.

[0030] Specifically, the 3D reconstruction unit selects J matching point groups for the part with assembly number n, and in the J matching point groups, the part in the i-th image I... i The set of keypoint pixel coordinates in the keypoint is The corresponding key point world coordinate set is Based on the i-th part drawing I i Corresponding camera extrinsic parameter T i Construct the i-th projection equation, i.e., the world coordinate set of key points. Equal to camera extrinsic parameter T i Distortion parameter R i and the set of key pixel coordinates are The product of these equations is used to construct a system of projection equations by combining the projection equations of the six part drawings, using the world coordinate set of the key points. Using the distortion parameters of six part images as optimization variables, and minimizing the reprojection error as the optimization objective, a nonlinear optimization algorithm is employed. The optimization variables are initialized and the reprojection error is calculated. Gradient descent is then used to update the optimization variables, and the reprojection error is calculated again until the reprojection error converges, yielding the optimal world coordinate set of key points for part number n. Among them, the reprojection error is the world coordinate set of the key points. The average error between the key point prediction pixel coordinate set of the six part drawings obtained by inverse calculation using the projection equation system and the corresponding key point pixel coordinate set.

[0031] Furthermore, the OBB method utilizes the world coordinate set of key points. The mean of the world coordinates of the J key points is used as the centroid world coordinates of the bounding box Sq(n). Calculate the world coordinate set of key points using principal component analysis. The covariance matrix is ​​obtained and the three principal axis directions of the bounding box Sq(n) are determined based on the eigenvectors corresponding to the three largest eigenvalues. The length l(n), width w(n), and height h(n) of the bounding box Sq(n) are calculated based on the projection point range of the world coordinates of the J key points in the principal axis directions, and the bounding box Sq(n) of the assembly part with assembly number n is generated.

[0032] Furthermore, the planning module includes a topology building unit, a path planning unit, and a parameter transformation unit;

[0033] Topology building elements calculate the projection overlap rate of the bounding boxes of two parts onto the XOY plane of the world coordinate system and compare it with the overlap rate threshold η. thIn comparison, when the projection overlap rate is greater than the overlap rate threshold η th If the assembly number of the higher-ranking part is greater than that of the lower-ranking part, then a directed edge is created between the higher-ranking part and the lower-ranking part. If the number is less than that, it is ignored. After traversing K parts, according to the chain propagation principle, when three parts are connected by directed edges in the same direction, a directed edge in the same direction is created between the first part and the last part to construct a directed graph of K parts.

[0034] The path planning unit uses a joint optimization algorithm to encode the operation sequence q into a chromosome, constructing assembly order constraints, topological sorting constraints, operation integrity constraints, and buffer constraints, based on the total length f of the movement path. L (q) and total angle adjustment f θ The objective function f(q) is constructed using the mean of (q). A genetic algorithm is then used to perform mutation operations on the chromosomes with the goal of minimizing the objective function f(q), iteratively generating the optimal operation sequence q. * ;

[0035] The parameter transformation unit is based on the m-th optimal operation. Determine the corresponding optimal pose The optimal pose sequence b is obtained based on the inverse kinematics algorithm. * The m-th optimal pose With the (m+1)th optimal pose The change is converted into the parameter of the m-th joint. Generate joint parameter sequence And deliver the actuator, where the 0th joint parameter From the initial pose b0 of the actuator to the first optimal pose Joint parameters.

[0036] Specifically, the joint optimization algorithm constructs assembly order constraints, topology sorting constraints, operational integrity constraints, and cache constraints, including the following steps:

[0037] Establish assembly sequence constraints, ensuring that the assembly operations for parts with earlier assembly numbers are also performed first;

[0038] Construct a topological sorting constraint: for two parts in a directed graph with a directed edge, the grab operation of the part to which the edge is pointed is after the grab operation of the other part.

[0039] To establish operational integrity constraints, the direct assembly operation chain requires that the part grab operation must be followed immediately by the assembly operation. The cached assembly operation chain requires that the part grab operation must be followed immediately by the cache operation, the cache retrieval operation must be followed immediately by the assembly operation, and the cache operation must precede the cache retrieval operation but does not necessarily have to be adjacent.

[0040] The cache constraints require that cache area numbers for identical parts be the same and that a cache area can only store a single part at any given time.

[0041] Specifically, based on the m-th operation q m Determine the corresponding pose b m Define the total length of the movement path f. L (q) represents the sum of the world coordinates between adjacent operations in the pose of M operations, and the total angle adjustment f is defined as follows. θ Let f(q) be the sum of the deviations between the execution angle vectors of adjacent operations in the poses of M operations, and define the objective function f(q) as the total length of the movement path f. L (q) and total angle adjustment f θ The mean of (q).

[0042] Specifically, the genetic algorithm arranges the assembly operations of K parts according to the assembly order constraint. Before the assembly operation of part number 1, a grasping operation with the same assembly number is inserted. Based on the topological sorting constraint, target parts are selected sequentially according to the assembly number order. An occlusion operation that obstructs the target part is retrieved using a directed graph. Before the grasping operation of the target part, grasping operations and caching operations for occlusion operations that do not yet exist are inserted, until all K parts have been traversed. Based on the operation integrity constraint, grasping and caching operations with the same assembly number and adjacent operations, as well as grasping and assembly operations, are bound separately. A cache retrieval operation with the same assembly number is inserted and bound before a single assembly operation. R sets of cache zone number vectors are randomly generated and sequentially assigned to all cache operations and their corresponding cache retrieval operations. R operation sequences are initialized and generated, based on the r-th operation sequence q. r Determine the corresponding pose sequence b r And calculate the r-th objective function value f(q) r The operation sequence corresponding to the minimum objective function value is determined as the optimal operation sequence q. * Keeping the first half of the operation sequences with smaller objective function values ​​unchanged, and fixing the assembly operations and their associated fetching or cache fetching operations in the second half of the operation sequences with larger objective function values, the bound fetching and cache fetching operations are randomly swapped and fine-tuned according to topological sorting constraints. The cache area number vector is then randomly reassigned, and the objective function values ​​of the R operation sequences are recalculated to redetermine the optimal operation sequence q. * Repeat the optimal operation sequence q for R consecutive rounds. * Keeping the sequence constant, output the optimal operation sequence q. * .

[0043] Furthermore, based on the m-th operation q m Determine the corresponding pose b m ,include:

[0044] When the m-th operation q m When capturing Gr, the pose b m Execution world coordinates p m The execution direction is the world coordinate of the center of the upper surface of the bounding box of the grabbed part and the main axis direction parallel to its height, respectively.

[0045] When the m-th operation q m When caching Bf or retrieving Gb and the cache area number is l, the position b m Execution world coordinates p m The execution direction is the sum of the world coordinates of the center of the l-th buffer, the height of the bounding box of the grabbed part, and the Z-axis direction of the world coordinate system, respectively.

[0046] When the m-th operation q m For assembly In, pose b m Execution world coordinates p m The execution direction and the direction of execution are the world coordinates of the assembly center of the assembled part and the direction of the normal vector of the assembly plane, respectively.

[0047] Compared with existing technologies, this invention improves the image quality of part image sets by capturing them through an identification module and evaluating the sets. It employs a target detection model to preprocess, extract and fuse features, and generate and filter candidate boxes in each part image set to obtain the candidate box location information, size information, and part category of each part. The reconstruction module numbers each part according to its assembly order and uses a matching algorithm to detect and match key points of each part in different part images. Based on the camera's extrinsic parameters, a projection equation system is constructed, and a nonlinear optimization algorithm is used to obtain the optimal world coordinates of the key points. The system fits and establishes a bounding box for each part. The planning module performs occlusion detection and occlusion decision on the bounding box of each part to construct a directed graph. The directed graph reflects the stacking situation between different parts. A joint optimization algorithm is used to encode the operation sequence into chromosomes and construct constraints to meet assembly and topology requirements. The chromosomes are mutated and the constraints are fine-tuned with the goal of minimizing the total length of the movement path and the total amount of angle adjustment. The optimal operation sequence is generated iteratively and the optimal pose sequence is determined. The inverse kinematics algorithm is used to solve the joint parameter sequence and deliver it to the actuator, realizing high-precision part recognition and efficient actuator planning and control in the state of part stacking. Attached Figure Description

[0048] Figure 1 A schematic diagram of an identification control system used in the assembly of smart wearable devices;

[0049] Figure 2 This is a flowchart of the image set evaluation process;

[0050] Figure 3 Here is a flowchart of the OBB method;

[0051] Figure 4 This is a schematic diagram of a directed graph. Detailed Implementation

[0052] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0053] like Figure 1 As shown, a specific embodiment of the present invention discloses an identification control system for assembling smart wearable devices, including an identification module, a reconstruction module, and a planning module;

[0054] The identification module captures a collection of parts images. The image atlas is evaluated using entropy and edge sharpness to determine whether to adjust the light source intensity I using an adaptive control algorithm. L and ISO sensitivity G I And reshoot the parts catalog that passed the evaluation. Input the target detection model, and preprocess the i-th part image I. i The i-th part image I is obtained sequentially through feature extraction, feature fusion, candidate box generation, and score filtering. i Parts annotation set in, and χ i,k These are the i-th part drawings I. i The frame center pixel coordinates, frame width, frame height, and part category of the k-th part, where K is the total number of parts;

[0055] The reconstruction module reorders K parts according to their assembly order and generates assembly numbers. It then uses a matching algorithm to detect key points of the part with assembly number n in different part images and performs correlation matching, based on the camera's extrinsic parameters. A system of projection equations for world coordinates and pixel coordinates is constructed. With minimizing the reprojection error as the optimization objective, a nonlinear optimization algorithm is used to obtain the optimal world coordinates of key points and fit them to establish a bounding box Sq(n) for the assembly of part number n. This completes the construction of the bounding box set.

[0056] The planning module performs occlusion detection and occlusion decision on the bounding boxes of K parts, constructs a directed graph of K parts, and uses a joint optimization algorithm to encode the operation sequence q into a chromosome, constructing assembly order constraints, topological sorting constraints, operation integrity constraints, and cache constraints, based on the total movement path length f of the actuator. L (q) and total angle adjustment f θ (q) Construct an objective function f(q), and perform mutation operations and constraint fine-tuning on the chromosome with the goal of minimizing the objective function f(q), iteratively generating the optimal operation sequence q. *Determine the corresponding optimal pose sequence b * The inverse kinematics algorithm is then used to convert the sequence into a joint parameter sequence. And deliver the actuator.

[0057] Furthermore, the recognition module includes a data acquisition and evaluation unit and a target recognition unit;

[0058] The data acquisition and evaluation unit uses a PLC to synchronously control six industrial cameras arranged in a circular array to capture images and obtain part atlases. And perform drawing set evaluation, calculate the I value of the i-th part drawing. i The entropy value H i and edge sharpness E i And obtain the part drawing set according to the preset weights. The overall score S is based on the overall score S and the score threshold S. th The comparison results will determine the choice of the delivery parts drawing set. The light source intensity I is adjusted by the target recognition unit or by using an adaptive control algorithm. L and ISO sensitivity G I And then take control of the six industrial cameras to take pictures again;

[0059] The target recognition unit uses a target detection model to identify the i-th part drawing I. i Normalize to generate the i-th standard part drawing. Feature extraction is performed using multi-scale convolution, and feature fusion is used to generate the i-th part image I. i Enhanced features Enhancement features are processed separately using parallel convolutions. The convolution results are then stitched together to generate the i-th part image I. i Predicted tensor O i Based on the predicted tensor O i For each candidate box, the box offset vector Δ is used to calculate the corresponding box center pixel coordinates p. px,c Frame width w px and frame height h px Combined with the prediction tensor O i The confidence α and category score vector S for each candidate box c Calculate the corresponding box score S F Then, a score-based filtering process is performed, retaining K candidate boxes containing the part to obtain the i-th part image I. i Parts annotation set

[0060] like Figure 2 As shown, furthermore, the data acquisition and evaluation unit collects data from the parts drawing set. The image set evaluation includes the following steps:

[0061] Calculate the entropy value of the i-th part drawing. Among them, P i (τ) represents the i-th part drawing I i The probability of a pixel with a grayscale value of τ appearing, and the entropy value H. i Used to measure the richness of image information in the i-th part drawing;

[0062] The Sobel operator is used to calculate the i-th part drawing I. i Lateral gradient map in the u direction and longitudinal gradient plot in the v direction For the i-th part drawing I i lateral gradient plot With longitudinal gradient plot The gradient magnitudes are summed to obtain the i-th part drawing I. i Edge sharpness in, and These are the horizontal gradient plots. With longitudinal gradient plot At pixel coordinate p px The horizontal and vertical gradients at the location, with the u and v directions being the coordinate axes of the two-dimensional pixel coordinate system;

[0063] Confirm the minimum entropy value H in the 6 part drawings min Maximum entropy H max Minimum edge sharpness E min and maximum edge sharpness E max The i-th part drawing I i The entropy value H i and edge sharpness E i We take the weighted sum of the minimum and maximum normalized results to obtain the i-th part image I. i Score S i The details are as follows:

[0064]

[0065] Where, ω H and ω E The part image set is obtained by averaging the scores of the six part images, using preset entropy and clarity weights. Overall score and with the scoring threshold S th Compare;

[0066] If the overall score S is greater than or equal to the score threshold S th Then the parts drawing set The image set passed the evaluation and was delivered to the target recognition unit;

[0067] If the overall score S is less than the score threshold Sth The light source intensity I is adjusted using an adaptive control algorithm. L and ISO sensitivity G I And then take control of the six industrial cameras to take pictures again.

[0068] Furthermore, the adaptive control algorithm includes the following steps:

[0069] Due to the light source intensity I L The light source intensity I directly affects the distribution of light and shadow and the rendering of details in the part drawing, therefore, the entropy value is used to correct the light source intensity I. L Calculate the average entropy value of the 6 part drawings. average entropy With the preset entropy value H pre The difference and entropy correction coefficient β H The product of the light source intensity captured in this photograph is added together. Obtain the light source intensity for the next shot.

[0070] Due to image noise level and ISO sensitivity G I There is a positive correlation, therefore ISO sensitivity G is corrected by the image noise level. I Calculate the average image noise level of 6 part drawings. Specifically as follows:

[0071]

[0072] Where F[] denotes Fourier transform, Represents the i-th part drawing I i The high-frequency portion of the signal after Fourier transform, I i (p px ) represents the i-th part drawing I i Mid-pixel coordinates p px The pixel value at that location will average the image noise level. Compared with the preset image noise level δ pre The difference and the noise level correction factor β δ The product of these values ​​is the ISO sensitivity of this shot. Get the ISO sensitivity for the next shot.

[0073] Furthermore, the object detection model includes a backbone network, a neck network, branch detection heads, a bounding box decoder, and a score filter;

[0074] The backbone network will display the i-th part diagram I. i Normalize to generate the i-th standard part drawing. The corresponding feature map F is generated by dimensionality reduction using a 7×7 convolution kernel and max pooling.i The first-level feature map is obtained by downsampling three times consecutively. Secondary feature map and third-level feature maps Each downsampling step involves dimensionality reduction of the previous downsampling result through convolution. The output dimension of each downsampling step is half the input dimension. The input for the first downsampling step is the feature map F. i ;

[0075] The neck network will have three levels of feature maps The upsampling results and secondary feature maps Concatenation, using point-by-point convolution to generate intermediate features Upsample intermediate features again and with the first-level feature map The concatenation process involves fusing the concatenated results through pointwise convolutions and then enhancing them using a self-attention mechanism to generate enhanced features.

[0076] Branch detection head in enhancing features At each feature point, three candidate boxes of different sizes are generated, and convolution is performed within each candidate box. The convolution results within each candidate box are input into the first, second, and third nonlinear branches, respectively, and the box offset vector Δ, confidence α, and class score vector S of each candidate box are output. c And concatenate the predicted vectors of each candidate box column by column, and combine the i-th part drawing I i The prediction vectors of each candidate box are concatenated along the channel dimension to form the i-th part image I. i Predicted tensor O i The nonlinear branches include a linear layer, a normalized layer, and a ReLU function. The output dimensions of the first, second, and third nonlinear branches are 1×4, 1×1, and 1×C, respectively, where C is the total number of part categories and S is the category score vector. c Record the category score of the candidate boxes for each part category;

[0077] The bounding box decoder starts from the i-th part drawing I. i Predicted tensor O i The box offset vector Δ = [Δu, Δv, Δw] of each candidate box is extracted sequentially. px ,Δh px ], where Δu, Δv, Δw px and Δh px The offsets of each candidate box in the u-direction, v-direction, width offset, and height offset are respectively given. The offsets Δu in the u-direction and Δv in the v-direction are normalized using the Sigmoid function and then compared with the candidate box center in the enhanced feature calculation. pixel coordinates (u) c ,v c The corresponding directional coordinates of the superposition are then multiplied by the enhanced feature. step size Generate the pixel coordinates p of the center of each candidate box. px,c The details are as follows:

[0078]

[0079] Among them, enhanced features step size Equivalent to enhanced features With the i-th part drawing I i The dimensionality ratio, due to enhanced features The dimension is equal to that of the first-level feature map. The dimension, and the first-level feature map For the i-th part drawing I i The enhanced features were obtained through two downsampling operations. step size Calculate the offset Δw of the box width for each candidate box. px And frame height offset Δh px The width correction ratio exp(Δw) for each candidate box is obtained by using the exponent raised to the power of the natural constant. px ) and high correction ratio exp(Δh px ), and the preset width w of each candidate box px,pre and preset height h px,pre Multiply them separately to generate the width w of each candidate box. px =w px,pre ×exp(Δw px ) and frame height h px =h px,pre ×exp(Δh px );

[0080] The scoring filter extracts the S from the category score vector of each candidate box. c Category score of the c-th part category The confidence level α is normalized using the Sigmoid function and then multiplied to obtain the single-class bounding box score for each candidate box with respect to the c-th part category. Compare the maximum single-class bounding box score with the bounding box score threshold S. F,th If the score of the largest single-class bounding box is less than the bounding box score threshold S F,th If the score of the largest single-class box is greater than or equal to the box score threshold S, then discard the candidate box. F,th If the candidate box is retained, the part category corresponding to the highest single-class box score is used as the part label in the candidate box, and the i-th part image I is obtained by integrating them.i Part annotation set of K parts

[0081] Furthermore, the reconstruction module includes a retrieval numbering unit and a three-dimensional reconstruction unit;

[0082] Retrieve part assembly manual by number, based on part drawing I. i The part category χ of the k-th part i,k Match the assembly number n in the parts assembly manual and replace the i-th part drawing I. i Given the original number k of the k-th part, construct the i-th part diagram I. i Assembly parts annotation set in, and These are the assembly diagrams for part number n in the i-th part drawing. i The pixel coordinates of the center of the frame, the width of the frame, and the height of the frame;

[0083] The 3D reconstruction unit uses a matching algorithm to detect key points of the assembly part with assembly number n in six part images and generates key point descriptors. A FLANN matcher is used, combined with the similarity of the key point descriptors, to match key points in different part images. Anomalies in the matching point groups are excluded based on a random sampling consensus algorithm, and J matching point groups are selected. These are then combined with the i-th part image I. i Corresponding camera extrinsic parameter T i Construct the world coordinates of the key points reflecting J key points and the i-th part drawing I. i The i-th projection equation is used to map and associate the pixel coordinates of J keypoints. This is then used to construct a system of projection equations. With minimizing the reprojection error as the optimization objective, a nonlinear optimization algorithm is employed to obtain the world coordinate set of the keypoints. Then, a bounding box Sq(n) is fitted to assemble the parts with assembly number n, and a bounding box set is constructed. The matching algorithms include SIFT, SURF, and ORB algorithms. The FLANN matcher and random sample consensus algorithm are existing algorithms and will not be discussed further. The j-th matching point group includes key points that have matching relationships and originate from the six part drawings. The world coordinates of the j-th key point of the assembly part numbered n.

[0084] Specifically, the 3D reconstruction unit selects J matching point groups for the part with assembly number n, and in the J matching point groups, the part in the i-th image I... i The set of keypoint pixel coordinates in the keypoint is The corresponding key point world coordinate set is Based on Zhang Zhengyou's calibration method, the i-th part drawing I is taken in advance. i External parameters of industrial cameras Ti The i-th projection equation is constructed as follows:

[0085]

[0086] Among them, R i To photograph the i-th part image I i The distortion parameters of the industrial camera are used to construct a projection equation system by combining the projection equations of the matching point groups corresponding to 6 part images, using the world coordinate set of key points. Using the distortion parameters of six part images as optimization variables, and minimizing the reprojection error as the optimization objective, a nonlinear optimization algorithm is employed. The initial reprojection error is calculated based on the initialized optimization variables. Gradient descent is then performed using the partial derivatives of the reprojection error with respect to each optimization variable to update the optimization variables. The error calculation is repeated again based on the updated optimization variables, and this process is repeated until the reprojection error converges, yielding the optimal world coordinate set of key points for part number n. Among them, Zhang Zhengyou's calibration method is an existing algorithm and will not be described in detail. The reprojection error is the world coordinate set of the key points. The average error between the key point prediction pixel coordinate set and the corresponding key point pixel coordinate set of the six part drawings obtained by reverse derivation of the projection equation system.

[0087] like Figure 3 As shown, further, the OBB method is used based on the world coordinate set of key points of the assembly part with assembly number n. The bounding box Sq(n) of the assembly numbered part n is generated by fitting the bounding box. The OBB method utilizes the world coordinate set of key points. The mean of the world coordinates of the J key points is used as the centroid world coordinates of the bounding box Sq(n). Calculate the world coordinate set of key points using principal component analysis. The covariance matrix is ​​obtained and the eigenvectors corresponding to the three largest eigenvalues ​​are obtained. The three eigenvectors are used as the three principal axes of the bounding box Sq(n). Based on the projection point range of the world coordinates of J key points in the three principal axes, the length l(n), width w(n), and height h(n) of the bounding box Sq(n) are calculated to obtain the bounding box Sq(n) of the assembly part with assembly number n.

[0088] Furthermore, the planning module includes a topology building unit, a path planning unit, and a parameter transformation unit;

[0089] like Figure 4 As shown, the topology building unit calculates the projection overlap rate of the bounding boxes of parts A and B in the XOY plane of the world coordinate system and compares it with the overlap rate threshold η. th To perform occlusion detection, the overlap rate of the projected bounding boxes of part A and part B is greater than the overlap rate threshold η.th When a suspected occlusion relationship is determined between part A and part B, it is determined whether the assembly number of the part with the higher centroid world coordinate in the bounding box of the suspected occlusion relationship is greater than that of the other part. If it is greater, the part with the higher centroid world coordinate is regarded as the occluding part, and a directed edge from the occluding part to the occluded part is established. If it is less, it is ignored. After traversing K parts, according to the chain propagation principle, when part A has a directed edge pointing to part B and part B has a directed edge pointing to part C, if part A does not have a directed edge pointing to part C, a directed edge from part A to part C is added, and a directed graph of K parts is constructed.

[0090] The path planning unit uses a joint optimization algorithm to encode the operation sequence q into a chromosome. It constructs assembly order constraints, topological sorting constraints, operation integrity constraints, and buffer constraints to ensure that the actuator operation order conforms to the assembly order and topological relationship of the K parts, and that the actuator operations meet the requirements of real-world conditions. The total length f of the actuator's movement path is then determined. L (q) and total angle adjustment f θ The objective function f(q) = 0.5 is constructed using the mean of f(q). L (q)+f θ [q], with the objective function f(q) as the optimization objective, a genetic algorithm is used to perform mutation operations on the chromosomes and iteratively generate the optimal operation sequence. in, Let M be the m-th optimal operation, and M be the total number of operations.

[0091] The parameter transformation unit is based on the m-th optimal operation. Determine the corresponding optimal pose The optimal pose sequence The m-th optimal pose With the (m+1)th optimal pose The changes are transformed into the parameters of the m-th joint using the existing inverse kinematics algorithm. Generate joint parameter sequence And deliver the actuator, where the 0th joint parameter From the initial pose b0 of the actuator to the first optimal pose Joint parameters.

[0092] Specifically, the joint optimization algorithm constructs assembly order constraints, topology sorting constraints, operational integrity constraints, and cache constraints, including the following steps:

[0093] Construct assembly sequence constraints. For two parts with assembly numbers n1 and n2 that satisfy 1 ≤ n1 < n2 ≤ K, assume that the corresponding assembly operations are the m1-th operations. and the m2th operation The condition m1 < m2 is satisfied;

[0094] Construct a topological sorting constraint: for two parts in a directed graph with assembly numbers n1 and n2, and a directed edge from n1 to n2, then the part with assembly number n1 occludes the part with assembly number n2. Assume the corresponding grabbing operations are the m3rd operations. and the m4th operation The condition m3 < m4 is satisfied;

[0095] To construct operational integrity constraints, for assembling part number n, there are two valid operation chains. The direct assembly operation chain requires that the grab operation Gr(n,0) and the assembly operation In(n,0) of assembling part number n must be adjacent, and the grab operation Gr(n,0) must come first. The cache assembly operation chain requires that the grab operation Gr(n,0) and the cache operation Bf(n,l) of assembling part number n must be adjacent, and the grab operation Gr(n,0) must come first. The cache retrieval operation Gb(n,l) and the assembly operation In(n,0) must be adjacent, and the cache retrieval operation Gb(n,l) must come first. In addition, the cache operation Bf(n,l) must come before the cache retrieval operation Gb(n,l). However, there can be operations on other parts with different assembly numbers between the cache operation Bf(n,l) and the cache retrieval operation Gb(n,l).

[0096] Construct cache constraints, including cache consistency constraints and cache capacity constraints. The cache consistency constraint requires that the cache area number l of the cache operation Bf(n,l) and the cache fetch operation Gb(n,l) for assembling part number n must be the same. Assume that the cache operation Bf(n,l) and the cache fetch operation Gb(n,l) for assembling part number n are the m5th operations respectively. and the m6th operation The cache capacity constraint requires starting from the m5th operation. Up to the m6th operation The buffer zone with buffer zone number l contains only the part with assembly number n.

[0097] Furthermore, constructing the objective function f(q) includes the following steps:

[0098] Based on the m-th operation q m Determine the corresponding pose b m =[p m ,θ m ], where p m and θ m The executor executes the m-th operation q respectively. m The required execution world coordinates and execution angle vector, execution angle vector θ m Record the execution of the m-th operation q by the executor mThe deviation angle of the execution direction relative to the X-axis, Y-axis and Z-axis of the world coordinate system;

[0099] Define the total length f of the movement path. L (q) represents the sum of the execution world coordinates between adjacent operations in the M operations, i.e.

[0100] Define the angle adjustment total f θ (q) represents the sum of the deviations between the execution angle vectors of adjacent operations in the M operations, i.e.

[0101] Define the objective function f(q) as the total length of the movement path f L (q) and total angle adjustment f θ The mean of (q).

[0102] Specifically, the genetic algorithm generates assembly operations for K parts based on assembly order constraints, constructing an initial operation sequence q. o In the initial operation sequence q o Insert a grab operation for the part with assembly number 1 at the beginning;

[0103] Based on topological sorting constraints, for a part with assembly number n, the assembly numbers of all occluding parts that occlude the part with assembly number n are retrieved according to the directed graph. It is then determined whether there are grabbing and caching operations corresponding to the assembly numbers of all occluding parts in the assembly operation In(n,0) of the part with assembly number n.

[0104] If it does not exist, insert the grab operation and cache operation corresponding to the assembly number of the missing occluded part; if it already exists, ignore it and continue to search based on the part with assembly number n+1.

[0105] Continue until K parts have been traversed, generating a transition operation sequence q. tr Based on operational integrity constraints, the transition operation sequence q is... tr Bind adjacent fetch and cache operations with the same assembly number, as well as adjacent fetch and assembly operations with the same assembly number. For a single assembly operation, insert a cache fetch operation with the same assembly number into the single assembly operation and bind it. The order of the two successfully bound operations remains unchanged, and they can only be shifted as a whole.

[0106] Count the number of parts with cache operations and cache retrieval operations, and randomly generate R sets of cache zone number vectors. Each set of cache zone number vectors records the cache zone number randomly assigned to the parts with cache operations. Initialize and generate R operation sequences, where R is an even number.

[0107] Based on the r-th operation sequence q rDetermine the corresponding r-th pose sequence b r And calculate the r-th objective function value f(q) r The operation sequence corresponding to the minimum objective function value is taken as the optimal operation sequence q. * Keep the first half of the operation sequences with smaller objective function values ​​unchanged, perform mutation operations on the second half of the operation sequences with larger objective function values, fix the assembly operations of K parts and the bound grab operations or cache operations unchanged, randomly swap the bound grab operations and cache operations and check and fine-tune according to the topology sorting constraints, and re-randomly allocate the cache area number vector.

[0108] Recalculate the objective function values ​​of the R operation sequences after the mutation operation to update the optimal operation sequence q. * Repeat the mutation operation and update the optimal operation sequence q. * The optimal operation sequence q for R consecutive rounds. * Keeping the sequence constant, output the optimal operation sequence q. * .

[0109] Furthermore, based on the m-th operation q m Determine the corresponding pose b m ,include:

[0110] When the m-th operation q m When =Gr(n,0), the operation type is grasping Gr, and the pose is b. m Execution world coordinates p m The execution direction is the world coordinate of the center of the upper surface of the bounding box Sq(n) of the assembly part with assembly number n and the principal axis direction parallel to its height, respectively.

[0111] When the m-th operation q m =Bf(n,l) or q m When =Gb(n,l), the operation type is cache Bf or fetch cache Gb with cache area number l, and the position b is... m Execution world coordinates p m The execution direction is the center world coordinate of the l-th buffer, plus the height h(n) of the bounding box Sq(n) of the assembly numbered part n, and the Z-axis direction of the world coordinate system, respectively.

[0112] When the m-th operation q m When =In(n,0), the operation type is assembly In, and the pose is b. m Execution world coordinates p m The execution direction and the assembly center world coordinates and the assembly plane normal vector direction of the assembly part with assembly number n are respectively.

[0113] This invention discloses a recognition and control system for assembling smart wearable devices, comprising a recognition module, a reconstruction module, and a planning module. The recognition module captures a set of part images and performs image evaluation to improve image quality. It uses a target detection model to preprocess the evaluated part images, extract and fuse features, and generate and filter candidate boxes to obtain the position information, size information, and part category of candidate boxes for each part in the image set. The reconstruction module numbers each part according to the assembly order, detects and matches key points of each part in different part images using a matching algorithm, constructs a projection equation system based on camera extrinsic parameters, and employs a nonlinear optimization algorithm. The optimal world coordinates of key points are obtained and bounding boxes for each part are fitted. The planning module performs occlusion detection and occlusion decision on the bounding boxes of each part to construct a directed graph. The directed graph reflects the stacking situation between different parts. The operation sequence is encoded into chromosomes using a joint optimization algorithm. Constraints are constructed to meet assembly and topology requirements. The chromosomes are mutated and the constraints are fine-tuned with the goal of minimizing the total length of the movement path and the total amount of angle adjustment. The optimal operation sequence is generated iteratively and the optimal pose sequence is determined. The joint parameter sequence is solved using an inverse kinematics algorithm and delivered to the actuator, realizing high-precision part recognition and efficient actuator planning and control in the stacked state of parts.

[0114] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. An identification control system for smart wear device assembly, characterized by, The recognition module, the reconstruction module and the planning module are included. The recognition module photographs a part atlas and performs atlas evaluation, adjusts parameters through an adaptive control algorithm, re-photographs or inputs the part atlas into a target detection model, obtains a part label set of each part picture in the part atlas through preprocessing, feature extraction and fusion, candidate box generation and screening, and the part label set includes related parameters of the candidate box of the part existing in the part picture. The reconstruction module generates an assembly number for each part according to a part assembly sequence, detects and matches key points of each part in different part pictures through a matching algorithm, constructs a projection equation set according to a camera external parameter set, and obtains optimal key point world coordinates of the key points through a nonlinear optimization algorithm to fit and establish a bounding box of each part. The planning module performs occlusion detection and occlusion judgment on the bounding box of each part to construct a directed graph, encodes an operation sequence into a chromosome through a joint optimization algorithm, constructs constraints to meet assembly requirements and topological requirements, performs mutation operation and constraint fine-tuning on the chromosome to minimize the total length of the movement path and the total amount of angle adjustment, iteratively generates an optimal operation sequence and determines an optimal pose sequence, and solves a joint parameter sequence through an inverse kinematics algorithm and delivers an executor. 2.The identification control system for smart wear device assembly of claim 1, wherein, The planning module includes a topological establishment unit and a path planning unit. The topological establishment unit calculates a projection overlap rate of the bounding boxes of two parts and compares the projection overlap rate with an overlap rate threshold value, when the projection overlap rate is greater than the overlap rate threshold value, judges whether the assembly number of a high part is greater than that of a low part, if yes, establishes a directed edge from the high part to the low part, traverses all parts, establishes a directed edge in the same direction between the starting part and the ending part of three parts connected by the directed edge in the same direction according to the chain transmission principle, and constructs a directed graph. The path planning unit encodes an operation sequence into a chromosome through a joint optimization algorithm, constructs assembly sequence constraints, topological sorting constraints, operation integrity constraints and buffer constraints, constructs an objective function based on the average of the total length of the movement path and the total amount of angle adjustment, and performs mutation operation on the chromosome through a genetic algorithm to minimize the objective function, and iteratively generates an optimal operation sequence. 3.The identification control system for smart wear assembly of claim 2, wherein, The genetic algorithm initializes R operation sequences, determines a corresponding pose sequence according to the rth operation sequence and calculates the rth objective function value, determines the operation sequence corresponding to the minimum objective function value as the optimal operation sequence, maintains the first half of the operation sequences with smaller objective function values in the R operation sequences, fixes the assembly operations and the bound operations in the second half of the operation sequences with larger objective function values, randomly exchanges the bound grabbing operations and the buffer operations, and adjusts and fine-tunes them according to the topological sorting constraints, reallocates the buffer area number vector and calculates the objective function values of the R operation sequences to redefine the optimal operation sequence, and repeats the execution until the optimal operation sequence remains unchanged for consecutive R rounds, and outputs the optimal operation sequence. 4.The identification control system for smart wear assembly of claim 3, wherein, The genetic algorithm initializes R operation sequences, including the following steps: Arrange the assembly operations of the parts according to the assembly sequence constraints, and insert a grabbing operation before the assembly operation of the part with the assembly number 1. Based on the topological sorting constraint, the target part is selected in turn according to the assembly number sequence, the shielding part shielding the target part is searched by combining the directed graph, and the shielding operation and the caching operation of the shielding part which does not exist yet are inserted before the grabbing operation of the target part; Until all the parts are traversed, the caching operation is taken before the separate assembly operation according to the operation integrity constraint, and the grabbing operation and the caching operation and the grabbing operation and the assembly operation of the same assembly number and adjacent are bound respectively; R groups of caching area number vectors are randomly generated and sequentially allocated to all the caching operations and the corresponding caching operation taking, and R operation sequences are initialized. 5.The identification control system for smart wear assembly of claim 2, wherein, The joint optimization algorithm constructs the assembly sequence constraint, the topological sorting constraint, the operation integrity constraint and the caching constraint, including the following steps: The assembly sequence constraint is constructed, and the assembly operation sequence of the part with the assembly number in the front is also in the front; The topological sorting constraint is constructed, and for two parts with a directed edge in the directed graph, the grabbing operation of the pointed part is after the grabbing operation of the other part; The operation integrity constraint is constructed, and the direct assembly operation chain requires that the grabbing operation of the part is followed by the assembly operation, the caching assembly operation chain requires that the grabbing operation of the part is followed by the caching operation, the caching operation is before the taking operation but not necessarily adjacent, and the caching operation is before the taking operation but not necessarily adjacent; The caching constraint requires that the caching area numbers of the same part are the same, and the caching area at the same time can only store a single part. 6.The identification control system for smart wear assembly of claim 1, wherein, The identification module includes a collection and evaluation unit and a target identification unit; The collection and evaluation unit synchronously controls the industrial camera to shoot, acquires the part atlas and performs atlas evaluation, calculates the entropy value and edge definition of each part image and obtains a comprehensive score by weighting, compares the comprehensive score with a score threshold to select the delivery part atlas to the target identification unit or adjust the parameters by using an adaptive control algorithm and shoot again; The target identification unit inputs each part image into a target detection model, extracts features by normalization and multi-scale convolution and fuses to generate enhanced features of each part image, processes the enhanced features by parallel convolution to generate a prediction tensor of each part image, calculates the corresponding box center pixel coordinates, box width and box height according to the box offset vector of each candidate box in the prediction tensor, calculates the corresponding box score by combining the confidence score and the class score vector of each candidate box in the prediction tensor, and selects and retains the candidate box with the part to obtain the part label set of each part image. 7.The identification control system for smart wear assembly of claim 6, wherein, The target detection model includes a backbone network, a neck network and a branch detection head; The backbone network normalizes each part image and generates a corresponding feature map by convolution and maximum pooling, and obtains a first-level feature map, a second-level feature map and a third-level feature map by three consecutive downsampling; The neck network splices the up-sampling result of the third-level feature map with the second-level feature map, fuses to generate an intermediate feature by point-by-point convolution, splices the up-sampling result of the intermediate feature with the first-level feature map, and fuses and enhances by point-by-point convolution and self-attention mechanism to generate enhanced features; The branch detection head generates three different sizes of candidate boxes at each feature point in the enhanced features and performs convolution in each candidate box, performs three different nonlinear mappings on the convolution results in each candidate box, outputs a box offset vector, a confidence score, and a class score vector for each candidate box, and splices them into a prediction vector, and splices the prediction vector of each candidate box in each part image into a prediction tensor of each part image, wherein the class score vector records the class score of the candidate box for each part class. 8.The identification control system for smart wear assembly of claim 1, wherein, The reconstruction module comprises a three-dimensional reconstruction unit; The three-dimensional reconstruction unit detects key points of parts with different assembly numbers in each part image and generates key point descriptors using a matching algorithm, matches the key points in different part images using a FLANN matcher, screens out J matching point groups based on a random sample consensus algorithm, constructs a projection equation set reflecting the mapping relationship between the key point world coordinates of the J key points and the key point pixel coordinates in each part image in combination with the camera extrinsic parameters corresponding to each part image, and solves to obtain the key point world coordinate set and fits to establish the bounding box of the parts with different assembly numbers using a nonlinear optimization algorithm with the minimum reprojection error as the optimization objective, wherein the assembly number is determined according to the part installation manual, and each matching point group comprises key points that have a matching relationship and are respectively derived from all part images. 9.The identification control system for smart wear assembly of claim 8, wherein, The three-dimensional reconstruction unit selects the J matching point groups of the part with the assembly number n, constructs the i-th projection equation reflecting the mapping relationship between the key point pixel coordinate set of the J matching point groups in the i-th part image and the corresponding key point world coordinate set according to the camera extrinsic parameters corresponding to the i-th part image, that is, the key point world coordinate set is equal to the product of the camera extrinsic parameters, the distortion parameters and the key point pixel coordinate set, combines the projection equations of each part image to construct a projection equation set, takes the key point world coordinate set and the distortion parameters of each part image as the optimization variables, takes the minimum reprojection error as the optimization objective, uses a nonlinear optimization algorithm, initializes the optimization variables and calculates the reprojection error, updates the optimization variables through gradient descent and calculates the reprojection error again until the reprojection error converges, and obtains the optimal key point world coordinate set of the part with the assembly number n, wherein the reprojection error is the average error between the key point predicted pixel coordinate set of each part image obtained by inversely projecting the key point world coordinate set through the projection equation set and the corresponding key point pixel coordinate set. 10.The identification control system for smart wear assembly of claim 8, wherein, The OBB method is used to take the mean value of the J key point world coordinates in the key point world coordinate set of each part as the centroid world coordinate of the bounding box of each part, the covariance matrix of the key point world coordinate set of each part is calculated through principal component analysis, and the three principal axis directions of the bounding box of each part are determined based on the eigenvectors corresponding to the three largest eigenvalues, the length, width and height of the bounding box of each part are calculated according to the projection point ranges of the J key point world coordinates in the principal axis directions, and the bounding box of each part is fitted and generated.

Images