Welding pose matching method for blow molded parts based on parting line and wall thickness distribution characteristics

Abstract

The present invention discloses a method for matching the welding posture of blow molded parts based on the characteristics of the mold line and wall thickness distribution, comprising the following steps: 1) continuous multi-frame image input; 2) matching the feature points of blow molded parts based on the mold line and wall thickness distribution characteristics; 2.1) extracting the mold line features in the input image; 2.2) extracting the feature points on the mold line based on Gaussian downsampling; 2.3) performing inter-frame estimation based on the feature point method; 2.4) performing matching tracking based on the wall thickness distribution characteristics; 2.5) performing mismatch elimination based on ORB‑SLAM2; 3) data output of camera posture estimation. The present invention estimates the posture of the camera based on the mold line and wall thickness distribution characteristics of the blow molded welded parts and the invariance of the feature points on the mold line of the input images between two frames. The method can quickly match the feature points and complete the punching welding of the blow molded welded parts; effectively improves the matching efficiency of the welding posture of the blow molded welded parts and further improves the welding quality.

Description

Technical Field

[0001] The invention relates to the field of synchronous positioning and mapping methods of punching and welding blow molding robots, and in particular to a blow molding welding posture matching method based on a parting line and wall thickness distribution characteristics. Background Art

[0002] Traditional metal parts are difficult to form and have low strength at the welds. In order to reduce the weight and cost of products, products need to be transformed from metal materials to plastic resin materials. Plastic parts can be processed into various complex shapes due to their large degree of freedom in shape design and high space utilization.

[0003] In 2018, Zhou Jinchun et al. proposed a metal welding splicing device (patent publication number: CN108637582A) for metal welding parts. The metal pipe is fixed by a clamping device, and the height of the metal pipe is changed by adjusting the lifting assembly, which can be welded automatically. This device improves the welding accuracy of metal pipes and solves the problem of manual welding. However, this method can only solve the welding of metal pipes with simple shapes and designated welding points. It is not suitable for blow molded parts with complex shapes, and the position of the welding point is limited by the tooling mechanism.

[0004] In 2015, Yu Hua et al. proposed a positioner suitable for transformer oil tank welding (patent publication number: CN104607857B), which drives the transformer blow-molded welded parts to rotate through a servo motor to achieve precise rotation and facilitate automatic welding by welding robots. This solves the problem of welding complex welding points of transformer blow-molded welded parts. However, this method uses a fixture device to position and clamp the blow-molded welded parts, which will cause a certain degree of damage to the surface of the blow-molded welded parts; and when the entire blow-molded welded parts support frame is driven to rotate together, the risk factor is high.

[0005] To this end, Shen Zhi et al. proposed an intelligent flexible production line and its working method in 2018 (invention patent publication number: CN109693140B). Through the cooperation between the host computer and each processing station, multiple production modes can be realized, but the transfer manipulator used in this method cannot perform self-feedback of the robot, cannot detect the welding posture, and cannot realize scene reconstruction.

[0006] Visual SLAM (Simultaneous Localization and Mapping) can use the robot's camera as a receiving source to perform autonomous positioning and map construction during movement. Figure 2As shown in the figure, the mobile robot SLAM system based on monocular vision first obtains information from the camera; estimates the camera motion between adjacent pictures through visual odometer (Visual Inertial Odometry); then determines whether the camera has been to the previous position based on loop detection; sends the content of visual odometer and loop detection to the backend for optimization; finally, reconstructs the scene based on the estimated camera trajectory and posture.

[0007] Among them, the SLAM front-end estimation solves the motion estimation of the robot within the time interval of obtaining the sensor information of the previous and next two frames. The back-end receives data from the camera and IMU (inertial measurement unit), which requires nonlinear optimization and solves the optimization problem of the historical trajectory after the robot detects the path closure. The conditional probability distribution of the robot state space is iteratively updated to estimate the robot's position and posture.

[0008] An inertial measurement unit (IMU) is a device that uses an accelerometer and a gyroscope to measure the three-axis attitude angle (or angular rate) and acceleration of an object. The accelerometer detects the acceleration signals of the object in the three independent axes of the carrier coordinate system, while the gyroscope detects the angular velocity signals of the carrier relative to the navigation coordinate system. After processing these signals, the attitude of the object can be solved. In some places where the GPS signal is weak, the IMU provides a relative positioning information.

[0009] In view of the problem that the visual odometer has a large deviation in motion information due to the insufficient number of corner points, it is necessary to select more corner points from the feature image. Traditional corner point extraction methods include: Harris corner point detection method, SIFT corner point detection method, and SURF corner point detection method. These three methods have good robustness, but they have large computational complexity and long speed response time in visual SLAM. ORB (Oriented FAST and Rotated BRIEF) is a binary descriptor for fast feature point extraction and description. This descriptor is used for image stitching and has good rotation and scale invariance, which greatly improves computational efficiency.

[0010] In 2016, Mur-Artal et al. proposed ORB-SLAM, which uses ORB to extract features (MUR-ARTAL R et al, ORB-SLAM2: An open source slam system for monocular, stereo, and RGB-D cameras[J]. IEEE Transactions on Robotics, 2016, 33(5): 1255.); ORB-SLAM was improved by using a three-thread method, reducing the response time of the SLAM system. To further improve the system, Mur-Artal et al. proposed ORB-SLAM2 (Mur-Artal R, Tardos J D. ORB-SLAM2: an Open-Source SLAM System for Monocular, Stereo and RGB-D Cameras [J]. IEEE Transactions on Robotics, 2017, 33 (05): 1255-1262.) in 2017 and ORB-SLAM3 (Campos C, et al. ORB-SLAM3: An Accurate Open-Source Library for Visual, Visual-Inertial and Multi-Map SLAM [J]. 2020, 1-17.) in 2020, which extended the visual odometer to stereo cameras and RGB-D cameras, and improved the map accuracy by extracting the local geometric consistency of three co-visible key frames. However, the above methods do not take into account the special characteristics of blow-molded welded parts, and require the extraction of sufficient feature points from the environmental scene, so the computational cost is slightly high.

[0011] ORB-SLAM2 is a real-time monocular / stereo SLAM system based on feature points, proposed in 2015 (R Mur-Artal, and JDT Ar Dos. "ORB-SLAM2:an Open-Source SLAM System for Monocular, Stereo and RGB-D Cameras." (2016).). ORB-SLAM2 improves the ORB homogenization feature extraction strategy in ORB-SLAM, and improves the accuracy and robustness of the system compared to the method in OpenCV.

[0012] In 2020, Pang Zhi-hua et al. (Pang Zhi-hua, Qi Chen-kun. A robust monocular visual odometry algorithm [J]. Computer Engineering, 2020, 46 (12): 80-87.) proposed a robust monocular visual odometry method using a tree-structured feature matching and real-time map update scheme. However, when the surrounding environment texture is missing, too few point feature extractions will result in poor algorithm accuracy.

[0013] In summary, although the existing visual SLAM method can perform autonomous positioning and map construction in the process of blow molding welding, the problem that the matching speed is low due to the insufficient number of feature points of the visual odometer and the system fails to achieve real-time response has not yet been solved. Summary of the invention

[0014] The purpose of the present invention is to overcome the defects of the prior art and provide a method for matching the welding posture of blow molded parts based on the mold line and wall thickness distribution characteristics. The method can estimate the posture of the camera based on the mold line and wall thickness distribution characteristics of the blow molded parts and utilize the invariance of the feature points between two frames; the method can quickly match the feature points and complete the punching and welding of the blow molded parts.

[0015] The object of the present invention is achieved by: a method for matching the welding posture of a blow molded part based on the mold line and the wall thickness distribution characteristics, comprising the following steps:

[0016] Step 1) inputting multiple frames of images continuously;

[0017] Step 2) matching feature points of the blow molded part based on the mold line and wall thickness distribution characteristics;

[0018] Step 2.1) extracting the parting line features in the input image;

[0019] Step 2.2) extracting feature points on the mold line based on Gaussian downsampling;

[0020] Step 2.3) perform inter-frame estimation based on feature point method;

[0021] Step 2.4) performing matching tracking based on wall thickness distribution characteristics;

[0022] Step 2.5) Eliminate false matches based on ORBSLAM2;

[0023] Step 3) Data output of camera pose estimation.

[0024] As a further limitation of the present invention, the step 2.1) comprises:

[0025] Step 2.1.1) extracting the parting line features in the feature image based on LSD;

[0026] Step 2.1.2) Obtain radial wall thickness distribution characteristics based on ray regression.

[0027] As a further limitation of the present invention, the step 2.1.1) comprises:

[0028] (1) Reduce the size of the captured image by 15% through Gaussian sampling to reduce the jagged effect in the image;

[0029] (2) Calculate the gradient value and gradient direction for each point in the image;

[0030] (3) Sort the gradients;

[0031] (4) Suppress small gradient values;

[0032] (5) Combine pixels with similar gradient directions to generate a straight line support region;

[0033] (6) forming a rectangular envelope;

[0034] (7) Use the NFA uncertain finite automaton algorithm to calculate the straight line;

[0035] (8) Screen out the straight line and fit the parting line of the blow molded part, and extract the parting line where the upper box and the lower box are separated in the feature image as A 1 (x,y).

[0036] As a further limitation of the present invention, the step 2.1.2) includes: using an online eccentric wall thickness detection system X-RAY2000 to obtain the wall thickness of the blow-molded welded part in a very short time through instantaneous exposure imaging technology in two axes, and after multiple shots, the overall wall thickness distribution curve B(x, y) of the blow-molded welded part can be obtained.

[0037] As a further limitation of the present invention, the step 2.2) comprises: assuming that there are N layers of Gaussian pyramids, the joining line A of the Gaussian pyramid of the first layer from the bottom of the pyramid to the top of the pyramid l (x,y) can be expressed as:

[0038]

[0039] Among them, (x, y) is the pixel point on the Gaussian pyramid mold line, is the Gaussian filter window function; 0≤a≤h, 0≤b≤w; h l ,w l They are the height and width of the lth Gaussian pyramid layer, and the height and width of each Gaussian pyramid layer are different.

[0040] As a further limitation of the present invention, the step 2.2) specifically comprises the following steps:

[0041] Step 2.2.1) Calculate the pixel points A on the top parting line of the pyramid N The horizontal gradient of (x,y) and longitudinal gradient

[0042] Among them, the lateral gradient The parting line A N (x,y) partial derivative in the x direction, longitudinal gradient The parting line A N (x,y) calculates the partial derivative in the y direction; pixel (x,y) is the smallest processing unit;

[0043] Step 2.2.2) Calculate the Q matrix for each pixel;

[0044] Using the lateral gradient obtained in step 2.2.1) and longitudinal gradient The matrix Q is represented as:

[0045]

[0046] Where ζ(x, y) is the Gaussian filter window function; the eigenvalues ​​of the calculated matrix Q are recorded as λ1,λ2; if min{λ1,λ2}≥δ0 is satisfied, the corner point will be obtained and the corresponding position will be marked; where δ0 is the critical threshold, x, y are the horizontal and vertical coordinates of the pixel point whose mold line is less than the critical threshold, respectively; the obtained corner point position data set is recorded as Ω, Ω={(x k ,y k )|k=1,2,...,n1}; where x,y are the indices of the current pixel position; n1 is the number of corner points.

[0047] As a further limitation of the present invention, the step 2.3) comprises:

[0048] The multi-view geometry method is used to extract feature points between adjacent frame images, and the camera pose is regressed from the geometric relationship between feature points. The classic minimization reprojection error model is used to optimize the initial pose. The spatial coordinates of a point P in the three-dimensional space of the n1 data set Ω are expressed as (X W ,Y W ,Z W ,1), its coordinates in the pixel coordinate system are (u,v,1), and the relationship between the spatial point and the pixel point is:

[0049]

[0050] Where s is the proportionality factor; is the internal parameter matrix, (u0,v0) is the coordinate of the distortion center, the unit is pixel; Where, f is the focal length, f x is the scale factor of the horizontal axis of the image, f y is the scale factor of the vertical axis of the image, in millimeters; ξ x and y Respectively represent the physical size of a single pixel in the x-axis and y-axis directions on the imaging plane, in millimeters;

[0051] R = [r1 r2 r3] is the rotation matrix, is the translation matrix, R and T are the external parameter matrices;

[0052] Since the error will cause the camera pinhole model to fail, a least squares problem is constructed. The error is minimized by adjusting the camera's posture, and the optimal estimate of the posture is obtained as follows:

[0053]

[0054] U σ is the coordinate of the σth point in the world coordinate system, R σ is the rotation matrix of the camera at the σth point, T σ is the translation matrix of the camera at the σth point, Pw is the coordinate in the world coordinate system, S σ is the σth point in the proportional coefficient.

[0055] As a further limitation of the present invention, the step 2.4) comprises:

[0056] Determine the parting line A in step 2.2) based on the wall thickness distribution curve B(x,y) l The seed point A0 of (x,y) and the lateral gradient of point A0 and longitudinal gradient The parting line A in step 2.2) 1 Build a rectangle around (x, y) and compare the gradients between adjacent points to obtain their respective feature vectors v. i Add the local gradient of each adjacent row, and for the jth row:

[0057]

[0058] where λ3=f g (k)f l (k) is the Gaussian term, g d1 and g d2 The horizontal and vertical directions are defined.

[0059] As a further limitation of the present invention, the step 2.5) includes: based on the ORB-SLAM2 algorithm, selecting whether to add a key frame to each frame of the image positioning camera, and extracting appropriate feature points; the feature point matching quality expression is:

[0060]

[0061] Among them, γ is the quality factor, and β is the ratio of the Euclidean distance d established between each pair of feature points.

[0062] The present invention adopts the above technical solution, and compared with the prior art, the beneficial effects are as follows: the present invention uses an RGB-D camera to acquire images, extracts the mold line and wall thickness distribution features in the feature image, then extracts feature points based on Gaussian downsampling, then estimates between frames based on the feature point method, then matches and tracks based on wall thickness distribution features, eliminates mismatches based on the ORB-SLAM2 algorithm, and finally estimates the camera pose. Based on the mold line and wall thickness distribution features of blow molded welded parts, the present invention uses the invariance of feature points on the mold line of the input image between two frames to estimate the camera pose, can quickly match feature points, complete the punching welding of blow molded welded parts, effectively improve the matching efficiency of the welding pose of blow molded welded parts, and further improve the welding quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a flow chart of the present invention.

[0064] Figure 2 This is a schematic diagram of the visual SLAM system framework flow.

[0065] Figure 3 Schematic diagram of blow molded parts.

[0066] Figure 4 This is the wall thickness distribution characteristic of the blow molded part of the present invention.

[0067] Figure 5 This is a schematic diagram of feature point extraction based on Gaussian pyramid in the present invention.

[0068] Figure 6 Schematic diagram of the steps of the Gauss-Newton method of the present invention.

[0069] Among them, 1 is the mold line, and 2 and 3 are the welding positions of the parts to be blow molded. DETAILED DESCRIPTION

[0070] like Figure 1 As shown, the method for matching the welding posture of blow molded parts based on the mold line and wall thickness distribution characteristics includes the following steps:

[0071] Step 1) Input multiple frames of images continuously; use a monocular camera to take multiple frames of photos, denoted as I0, I1, ..., I n .

[0072] The visual odometer blow molding robot punching and welding posture matching method uses a monocular camera, which has a simple structure, low cost, strong scalability, and is more suitable for miniaturized intelligent devices.

[0073] Step 2) matching feature points of the blow molded part based on the mold line and wall thickness distribution characteristics;

[0074] Step 2.1) extracting the parting line features in the input image;

[0075] like Figure 3 As shown, the parting line refers to the feature line formed by trimming the excess overflow generated in the mold structure of the blow molded part after blow molding, which is the parting line 1; the feature parting line 1 of the blow molded part after blow molding is located at the separation between the upper box and the lower box, and the closed length of the line is long and complete, which is easier to be recognized by the camera. The welding positions 2 and 3 of the blow molded welded parts are generally curves, corresponding to the shape of the welded parts. Multiple welds are required to complete a welded part; the depth of the welded part and the welding position requires the posture matching of the welding point.

[0076] Step 2.1.1) Extracting the mold line features in the feature image based on the LSD (Line Segment Detector) straight line detection algorithm includes the following steps:

[0077] (1) Reduce the size of the captured image by 15% through Gaussian sampling to reduce the jagged effect in the image;

[0078] (2) Calculate the gradient value and gradient direction for each point in the image;

[0079] (3) Sort the gradients;

[0080] (4) Suppress small gradient values;

[0081] (5) Combine pixels with similar gradient directions to generate a straight line support region;

[0082] (6) forming a rectangular envelope;

[0083] (7) Use NFA (Non-Deeterministic Finite State Automata) uncertain finite automaton algorithm to calculate the straight line;

[0084] (8) Screen out the straight line and fit the parting line of the blow molded part, and extract the parting line where the upper box and the lower box are separated in the feature image as A 1 (x,y).

[0085] Step 2.1.2) Obtain radial wall thickness distribution characteristics based on ray regression;

[0086] The new online eccentric wall thickness detection system X-RAY2000 developed by SIKORA is used. It adopts the ray regression analysis technology and the instantaneous exposure imaging technology in two axes to obtain the wall thickness of the plastic oil tank in a very short time; after multiple shots, the overall wall thickness distribution curve B(x,y) of the blow-molded welded parts can be obtained.

[0087] Step 2.2) extracting feature points on the mold line based on Gaussian downsampling;

[0088] like Figure 5 As shown, suppose there are N layers of Gaussian pyramid, and the mold line A of the lth layer of Gaussian pyramid from the bottom to the top of the pyramid is l (x,y) can be expressed as:

[0089]

[0090] Among them, (x, y) is the pixel point on the Gaussian pyramid mold line, is the Gaussian filter window function; 0≤a≤h, 0≤b≤w; h l ,w l They are the height and width of the lth Gaussian pyramid layer, and the height and width of each Gaussian pyramid layer are different.

[0091] Step 2.2.1) Calculate the pixel points A on the top parting line of the pyramid N The horizontal gradient of (x,y) and longitudinal gradient

[0092] Among them, the horizontal gradient refers to the characteristic of taking pixel as the smallest processing unit. N (x,y) finds the partial derivative in the x direction, and the vertical gradient refers to the gradient of A N Take the partial derivative of (x,y) in the y direction.

[0093] Step 2.2.2) Calculate the Q matrix for each pixel;

[0094] Using the lateral gradient obtained in step 2.2.1) and longitudinal gradient The matrix Q is represented as:

[0095]

[0096] Where ζ(x, y) is the Gaussian filter window function; the eigenvalues ​​of the calculated matrix Q are recorded as λ1,λ2; if min{λ1,λ2}≥δ0 is satisfied, the corner point will be obtained and the corresponding position will be marked; where δ0 is the critical threshold, x, y are the horizontal and vertical coordinates of the pixel point whose mold line is less than the critical threshold, respectively; the obtained corner point position data set is recorded as Ω, Ω={(xk ,y k )|k=1,2,...,n1}; where x,y are the indices of the current pixel position; n1 is the number of corner points.

[0097] Step 2.3) perform inter-frame estimation based on feature point method;

[0098] The multi-view geometry method is used to extract feature points between adjacent frame images, and the camera pose is regressed from the geometric relationship between feature points. The classic minimization reprojection error model is used to optimize the initial pose. The spatial coordinates of a point P in the three-dimensional space of the n1 data set Ω are expressed as (X W ,Y W ,Z W ,1), its coordinates in the pixel coordinate system are (u,v,1), and the relationship between the spatial point and the pixel point is:

[0099]

[0100] Where s is the proportionality factor; is the internal parameter matrix, (u0,v0) is the coordinate of the distortion center, the unit is pixel; Where, f is the focal length, f x is the scale factor of the horizontal axis of the image, f y is the scale factor of the vertical axis of the image, in millimeters; ξ x and y Respectively represent the physical size of a single pixel in the x-axis and y-axis directions on the imaging plane, in millimeters;

[0101] R = [r1 r2 r3] is the rotation matrix, is the translation matrix, R and T are the external parameter matrices;

[0102] Since the error will cause the camera pinhole model to fail, a least squares problem is constructed. The error is minimized by adjusting the camera's posture, and the optimal estimate of the posture is obtained as follows:

[0103]

[0104] U σ is the coordinate of the σth point in the world coordinate system, R σ is the rotation matrix of the camera at the σth point, T σ is the translation matrix of the camera at the σth point, P W is the coordinate in the world coordinate system, s σ is the σth point in the scale factor. Then the reprojection error can be calculated about the position of point P in the world coordinate system The Jacobian matrix of .

[0105] Combine feature points and parting line features to optimize camera pose, construct environmental feature maps, and reduce the cumulative error of the entire system.

[0106] Step 2.4) performing matching tracking based on wall thickness distribution characteristics;

[0107] like Figure 4 As shown, wall thickness distribution refers to the uneven blowing ratio of each part of the blow-molded welded parts. The parts with larger blowing ratio are thicker than the parts with smaller blowing ratio, which compensates for the wall thickness changes caused by self-weight sagging and shrinkage. In order to improve the uniformity of the wall thickness of the parts, efforts should be made to adjust the wall thickness distribution.

[0108] Blow molded welded parts have complex shapes, and their cross-sectional shapes vary not only in the axial direction, but also in the radial direction. The wall of the part with small deformation is thick, and the wall of the part with large deformation is thin. From the material point of view, it can be divided into 6 layers, from the inside to the outside, namely: new material layer (HDPE), bonding layer (LLDPE), barrier layer (EVOH), adhesive layer (LLDPE), recycled material layer, and new material layer (HDPE). The thickness of blow molded welded parts is: 2.5mm-5mm.

[0109] Determine the parting line A in step 2.2 based on the wall thickness distribution curve B(x,y) l The seed point A0 of (x,y) and the lateral gradient of point A0 and longitudinal gradient The parting line A in step 2.2 l Build a rectangle around (x, y) and compare the gradients between adjacent points to obtain their respective feature vectors v. i Add the local gradient of each adjacent row, and for the jth row:

[0110]

[0111] where λ3=f g (k)f l (k) is the Gaussian term, g d1 and g d2 The horizontal and vertical directions are defined.

[0112] Step 2.5) Eliminate false matches based on ORB-SLAM2 (Oriented FAST and Rotated BRIEF feature-based simultaneous localization and mapping system version 2);

[0113] Based on the ORB-SLAM2 algorithm, choose whether to add key frames to each frame of the image positioning camera and extract appropriate feature points; the feature point matching quality expression is:

[0114]

[0115] Among them, γ is the quality factor, and β is the ratio of the Euclidean distance d established for each pair of feature points. The smaller β is, the smaller the distance is, the larger the quality factor γ is, and the better the matching quality is. Based on ORBSLAM2, a large number of mismatched point pairs are eliminated, which improves the matching effect and the positioning accuracy.

[0116] Step 3) Data output of camera pose estimation;

[0117] By setting the initial camera pose of the current frame to match the previous frame, we can get the matching relationship between 3D features and 2D features, and estimate the pose of the camera by combining the point-line matching relationship. We use PnP (Perspective-n-Point) and solve the pose by minimizing the cost function; Solve by minimizing the cost function:

[0118]

[0119] in is the optimal solution of the camera pose in the current frame. C and D are the sets of matching points and matching line segments in the current frame. are the projection errors of the parting line and thickness distribution, ∑ c ,∑ d is the covariance matrix of the parting line and thickness distribution, w c ,w d is the feature weight of the parting line and thickness distribution, L c ,L d is the Huber robustness loss function. The Huber function expression is: Where a0 is the difference between the observed value and the estimated value. δ1 is a hyperparameter and ε is the loss coefficient, which determines the size of the outlier.

[0120] like Figure 6 As shown, the Gauss-Newton method is used to minimize the target cost function and iteratively estimate the camera pose of the current frame.

[0121] The present invention uses an RGB-D camera to acquire images, extracts the mold line and wall thickness distribution features in the feature image, then extracts feature points based on Gaussian downsampling, and then estimates between frames based on the feature point method; then matches and tracks the wall thickness distribution features, eliminates mismatches based on ORB-SLAM2, and finally estimates the camera pose. Based on the mold line and wall thickness distribution features of the plastic fuel tank for automobiles, the invariance of the feature points between two frames is used to estimate the camera pose; the method of the present invention can quickly match the feature points and complete the punching and welding of the plastic fuel tank for automobiles; effectively improves the welding quality of the plastic fuel tank. The method can well solve the impact of the adaptive depth of the plastic fuel tank welding.

[0122] The present invention is not limited to the above-mentioned embodiments. On the basis of the technical solution disclosed in the present invention, technicians in this field can make some substitutions and deformations to some technical features therein according to the disclosed technical content without creative labor, and these substitutions and deformations are all within the protection scope of the present invention.

Claims

1. A method for matching the welding posture of blow molded parts based on the characteristics of the parting line and wall thickness distribution, characterized in that: The following steps are involved: Step 1) inputting multiple frames of images continuously; Step 2) matching feature points of the blow molded part based on the mold line and wall thickness distribution characteristics; Step 2.1) extracting the parting line features in the input image; Step 2.2) extracting feature points on the mold line based on Gaussian downsampling; The step 2.2) comprises: assuming that there are N layers of Gaussian pyramid, the joining line A of the Gaussian pyramid of the first layer from the bottom to the top of the pyramid l (x,y) is represented as: Among them, (x, y) is the pixel point on the Gaussian pyramid mold line, is the Gaussian filter window function; 0≤a≤h, 0≤b≤w; h l ,w l are the height and width of the first layer of the Gaussian pyramid layer, and the height and width of each layer of the Gaussian pyramid layer are different; l-1 Indicates the joining line of the l-1th layer of the Gaussian pyramid; a and b are the horizontal and vertical coordinates of the Gaussian filter function respectively; h and w are the total height and total width of the Gaussian pyramid layer respectively; Step 2.2.1) Calculate the pixel points A on the top parting line of the pyramid N The horizontal gradient of (x,y) and longitudinal gradient Among them, the lateral gradient The parting line A N (x,y) partial derivative in the x direction, longitudinal gradient The parting line A N (x,y) calculates the partial derivative in the y direction; pixel (x,y) is the smallest processing unit; Step 2.2.2) Calculate the Q matrix for each pixel; Using the lateral gradient obtained in step 2.2.1) and longitudinal gradient The matrix Q is represented as: in, and The horizontal gradient and longitudinal gradient The square root of ζ(x, y) is the Gaussian filter window function; the eigenvalues ​​of the calculated matrix Q are recorded as λ1,λ2; if min{λ1,λ2}≥δ0 is satisfied, the corner point will be obtained and the corresponding position will be marked; where δ0 is the critical threshold, x, y are the horizontal and vertical coordinates of the pixel point whose mold line is less than the critical threshold, respectively; the obtained corner point position data set is recorded as Ω, Ω={(x k ,y k )|k=1,2,...,n1}; where x,y are the indices of the current pixel position; x k and k are the horizontal and vertical coordinates of the current pixel position respectively; n1 is the number of corner points; Step 2.3) perform inter-frame estimation based on feature point method; The step 2.3) comprises: The multi-view geometry method is used to extract feature points between adjacent frame images, and the camera pose is regressed from the geometric relationship between feature points. The classic minimization reprojection error model is used to optimize the initial pose. The spatial coordinates of a point P in the three-dimensional space of the n1 data set Ω are expressed as (X W ,Y W ,Z W ,1), where X W , Y W and Z W are the coordinates of point P on the horizontal, vertical and vertical axes of space respectively; its coordinates in the pixel coordinate system are (u, v, 1), and the relationship between the spatial point and the pixel point is: Where s is the proportionality factor; is the internal parameter matrix, (u0,v0) is the coordinate of the distortion center, the unit is pixel; Where, f is the focal length, f x is the scale factor of the horizontal axis of the image, f y is the scale factor of the vertical axis of the image, in millimeters; ξ x and y Respectively represent the physical size of a single pixel in the x-axis and y-axis directions on the imaging plane, in millimeters; R = [r1 r2 r3] is the rotation matrix, is the translation matrix, R*T is the external parameter matrix; r1, r2, r3 are 1×3 rotation sub-matrices on the horizontal, vertical and vertical axes of the space respectively; t1, t2, t3 are 1×1 translation sub-matrices on the horizontal, vertical and vertical axes of the space respectively; Since the error will cause the camera pinhole model to fail, a least squares problem is constructed to minimize the error by adjusting the camera's position and posture to obtain the optimal estimate of the position and posture. for: is the optimal estimate of the pose of a point P in the three-dimensional space of the mold line feature point in the world coordinate system, U σ is the coordinate of the σth point in the world coordinate system, R σ is the rotation matrix of the camera at the σth point, T σ is the translation matrix of the camera at the σth point, Pw is the coordinate in the world coordinate system, S σ is the σth point in the proportional coefficient; Step 2.4) performing matching tracking based on wall thickness distribution characteristics; The step 2.4) comprises: Determine the parting line A in step 2.2) based on the wall thickness distribution curve B(x,y) l The seed point A0 of (x,y) and the lateral gradient of point A0 and longitudinal gradient The parting line A in step 2.2) 1 Build a rectangle around (x, y), obtain the feature vector v of each point by comparing the gradients between adjacent points, and convert the i-th strip B i Add the local gradient of each adjacent row, and for the jth row: where λ3=f g (k)f l (k) is the Gaussian term, g d1 and g d2 For the defined horizontal and vertical directions; are the local gradient sums of the feature vectors v1, v2, v3, and v4 between the jth pair of adjacent points in the horizontal direction of the i-th row strip whose corresponding gradient values ​​are greater than 0; f g (k), f i (k) are the Gaussian coefficients in the horizontal and vertical directions respectively; Step 2.5) Eliminate mismatches based on ORB-SLAM2; Step 3) Data output of camera pose estimation.

2. The method for matching the welding posture of blow molded parts based on the mold line and wall thickness distribution characteristics according to claim 1, characterized in that: The step 2.1) comprises: Step 2.1.1) extracting the parting line features in the feature image based on LSD; Step 2.1.2) Obtain radial wall thickness distribution characteristics based on ray regression.

3. The method for matching the welding posture of blow molded parts based on the mold line and wall thickness distribution characteristics according to claim 2, characterized in that: The step 2.1.1) comprises: (1) Reduce the size of the captured image by 15% through Gaussian sampling to reduce the jagged effect in the image; (2) Calculate the gradient value and gradient direction for each point in the image; (3) Sort the gradients; (4) Suppress small gradient values; (5) Combine pixels with similar gradient directions to generate a straight line support region; (6) forming a rectangular envelope; (7) Use the NFA uncertain finite automaton algorithm to calculate the straight line; (8) Screen out the straight line and fit the parting line of the blow molded part, and extract the parting line where the upper box and the lower box are separated in the feature image as A 1 (x,y).

4. The method for matching the welding posture of blow molded parts based on the mold line and wall thickness distribution characteristics according to claim 2, characterized in that: The step 2.1.2) includes: using an online eccentric wall thickness detection system to obtain the wall thickness of the blow-molded welded part through instantaneous exposure imaging technology in two axes, and after multiple shots, the overall wall thickness distribution curve B(x, y) of the blow-molded welded part can be obtained.

5. The method for matching the welding posture of blow molded parts based on the mold line and wall thickness distribution characteristics according to claim 1, characterized in that: The step 2.5) includes: based on the ORB-SLAM2 algorithm, selecting whether to add a key frame to each frame of the image positioning camera, and extracting appropriate feature points; the feature point matching quality expression is: Among them, γ is the quality factor, β is the minimum Euclidean distance d established for each pair of feature points min ratio.

Images