CN106408609A - Parallel mechanism end motion pose detection method based on binocular vision - Google Patents

Abstract

The invention discloses a binocular vision-based detection method for the terminal motion and posture of a parallel mechanism. First, the collected images of the parallel mechanism are subjected to image denoising preprocessing based on wavelet transform; images for feature matching. The matching algorithm first extracts the image feature points through the Harris operator, and then uses the SIFT feature descriptor to match the feature points extracted by the Harris operator; then, a new purification algorithm is proposed to purify the matching results. The method of taking points in blocks and taking points in advance to check the temporary model; finally, bring the pair of characteristic points at the end of the parallel mechanism after matching and purification into the binocular vision model, and obtain the three-dimensional pose of the end of the mechanism through coordinate transformation. The invention can greatly reduce the image processing time, and further improve the matching accuracy rate through the proposed new purification algorithm to purify the matching result, thereby improving the real-time performance and accuracy of the terminal pose detection of the parallel mechanism.

Description

A binocular vision-based detection method for terminal motion pose of parallel mechanism

technical field

The invention relates to a pose detection system based on machine vision, in particular to a binocular vision-based detection method aimed at the motion pose of an end of a parallel mechanism.

Background technique

In the control of parallel mechanisms, the terminal motion pose is an important parameter that reflects the motion state of the mechanism. Accurately measuring the end motion pose of the mechanism can effectively avoid the errors caused by the kinematic model solution, which will facilitate the realization of the parallel mechanism. High performance controls. Compared with other detection methods, machine vision has the advantages of non-contact, strong applicability, and high cost performance. It is especially suitable for the detection of motion poses at the end of parallel mechanisms with multiple degrees of freedom, complex motion trajectories, and difficult direct detection. Visual pose detection systems can generally be divided into monocular, binocular, and multi-eye detection systems. Because the robustness and accuracy of the binocular vision algorithm are superior to the monocular algorithm, and the real-time performance of the algorithm has a greater advantage compared with the multi-eye vision, therefore, the binocular vision has been widely used. For the detection of the terminal motion pose of the parallel mechanism, the difficulty of using the binocular vision detection method is how to quickly and accurately obtain the terminal pose information of the mechanism after image processing of the image of the parallel mechanism collected by the vision system. In the binocular vision pose detection process, the stereo matching process is the most important. The speed and accuracy of matching directly determine the speed and accuracy of the entire pose detection system. External factors will cause the image matching of the binocular vision system to be difficult to achieve accurately, which will affect the speed and accuracy of the visual inspection system.

The document "A New Type of Stereo Vision Detection System for Parallel Robot Pose" (Wu Difei, Ding Yongsheng, etc. Computer Engineering and Application, 2007, 43(33):190-192) established a scale-invariant feature transformation based on ( The binocular vision detection system of SIFT) detects the pose of parallel robots, and proves the robustness and feasibility of the detection algorithm by simulation experiments, but because the SIFT algorithm used in the matching process of the detection system will generate a large number of Large eigenvectors increase the complexity of the algorithm, which affects the real-time performance of pose detection; the document "Harris-SIFT Algorithm and Its Application in Binocular Stereo Vision" (Zhao Qinjun, Zhao Dongbiao, etc. Journal of University of Electronic Science and Technology of China , 2010, 39(4):546-550) combined Harris saliency and SIFT descriptors, proposed a Harris-SIFT algorithm for binocular vision matching, and verified the effectiveness of the method, which effectively improved the The matching algorithm is real-time, but because the corner points detected by Harris contain less information than the feature points detected by SIFT, the Harris-SIFT algorithm has the problem of wrong matching and error matching.

In view of the mismatching problem existing in the Harris-SIFT algorithm, a purification algorithm can be considered to purify the matching results. Conventional purification algorithms usually calculate the target model of matching point pairs based on a set of data sets containing abnormal points, and then use the remaining points to test the model to obtain valid data samples. Therefore, the key to the purification algorithm is to quickly find the correct target model . However, there are usually some problems in the acquisition of the target model: for example, when selecting a random sample, there are two candidate points that are too close to be considered as one point, resulting in an inaccurate target model; and each time a random sample is selected set, it is necessary to find the support set corresponding to the candidate model, which will lead to spending too much time looking for the corresponding support point set with more error observation data sets.

Contents of the invention

In view of the problems existing in the above-mentioned binocular vision detection, the present invention proposes a binocular vision-based detection method for the terminal motion pose of a parallel mechanism. This method proposes a new matching algorithm, and the algorithm realizes image features through the fast Harris-SIFT algorithm. Point matching, and a new purification algorithm is proposed to purify the matching results, so as to solve the problems that the traditional binocular vision algorithm is complicated, the space matching is difficult to achieve accurately, and the real-time detection is poor and the detection accuracy is not high.

The technical scheme that the present invention adopts is to adopt following steps:

A binocular vision-based detection method for terminal motion pose of a parallel mechanism, comprising the steps of:

Step 1) Use the binocular camera to collect the original image of the parallel mechanism, and perform image denoising preprocessing based on wavelet transform on the collected mechanism image;

Step 2) using the corner point extraction method based on the Harris operator to extract the mechanism image features;

Step 3) use the Harris-SIFT algorithm to perform feature matching on the parallel mechanism image; the matching algorithm first extracts image feature points through the Harris operator, and then uses the SIFT feature descriptor to match the feature points extracted by the Harris operator;

Step 4) Aiming at the problem of wrong matching and error matching in the Harris-SIFT algorithm, a new purification algorithm is used to further purify the matching results: to improve the purification algorithm by taking points in blocks and checking the temporary model in advance, wherein, block Picking points: take the method of dividing the image into blocks and randomly selecting a matching point in each block; taking points in advance to check the temporary model: randomly select 9 matching pairs, and use 8 matches to find the temporary model parameters, and the ninth matching pair Used to verify the correctness of the temporary model to quickly determine the image model;

Step 5) According to the principle of binocular vision, the three-dimensional coordinates of the terminal feature points are calculated through the terminal feature points extracted by the previous image processing, and the terminal attitude angle of the parallel mechanism is calculated, and finally the terminal pose information of the parallel mechanism is obtained.

Further, described step 2) in, the corner extraction method of Harris operator can directly call OpenCV storehouse function and realize, and in Harris feature point detection, the second-order moment of pixel point can be expressed as:

In the formula, I _x , I _y represent the gradient of the gray level of the pixel point in the x and y directions; I _xy is the second-order mixed partial derivative; the Harris response function of the pixel point is:

R＝det Mk(traceM) ²

Where det M is the determinant of matrix M, traceM is the authentic trace of matrix M; k is an empirical constant, usually ranging from 0.04 to 0.06; set the threshold t, and calculate the response value R according to the above Harris operator, if R<t, Then this point is a feature point.

Further, in the step 4), the purification algorithm calculates the target model of matching point pairs through a set of data sets containing abnormal points, and then uses the remaining points to test the model, thereby obtaining valid data samples; the specific steps are as follows:

4.1) According to the coordinate boundaries of the matching points in the first picture, divide the picture into b×b blocks on average, where b>3, randomly select 9 different blocks and randomly select a matching point in each block to form 9 A random sample set S of matching point pairs;

4.2) Randomly select 8 matching point pairs from the random sample set S to obtain the temporary candidate model F;

4.3) Detect whether the ninth point is the support set of the temporary candidate model F: if yes, then the temporary candidate model F is the candidate model F; otherwise, reselect 9 pairs of matching points, and repeat 4.1), 4.2) process;

4.4) Detect all matching point pairs through the candidate model F and the threshold k, and obtain the support set logarithm m of the candidate model D;

4.5) If m≥threshold t, then obtain the target model F; otherwise, reselect 9 pairs of matching points and repeat 4.1), 4.2), 4.3), 4.4) process;

4.6) Optimize the target model F, and finally determine the target model H;

4.7) Each time a random sample set is selected, the counter is incremented by 1. If the model parameters are not found after k times of repetition, the program is terminated.

Further, in the step 5), the three-dimensional coordinates (X, Y, Z) of the characteristic points at the end of the parallel mechanism are obtained according to the projection theorem, and then the attitude angle of the end of the parallel mechanism is obtained through the camera rotation matrix R; wherein the terminal position information of the parallel mechanism as follows:

P _w ＝(A ^T A) ^-1 A ^T B

in, (u ₁ , v ₁ ), (u ₂ , v ₂ ) represent the image coordinates of the end feature points on the left and right camera projection points respectively; Indicates the i-th row and j-th column of the left and right projection matrix;

At the same time, the terminal attitude angle of the parallel mechanism is:

Among them, R _ij (i=1,2,3; j=1,2,3) represents the i-th row and j-th column of the rotation matrix R, and α is the pitch angle of rotation around the X axis, and the unit is: rad, β It is the roll angle around the Y axis, the unit is: rad, γ is the heading angle around the Z axis, the unit is: rad.

Further, in said step 5), it also includes that during three-dimensional reconstruction, at first the camera parameters of the built binocular vision model need to be obtained, and the camera parameters can be obtained through camera calibration; adopt Zhang Zhengyou plane calibration method to carry out camera calibration.

Further, the process of Zhang Zhengyou’s planar calibration method is as follows: first fix the camera, then take pictures of the calibration plate, then input the image of the calibration plate into the computer and obtain the pixel coordinates of each grid point on the calibration plate, and then take the known three-dimensional coordinates of the calibration plate The coordinates are loaded into the computer, and by substituting the pixel coordinates and three-dimensional coordinates of the calibration board into the calibration model, the internal and external parameters of the camera are solved, and the left and right cameras are calibrated separately.

The present invention proposes a binocular vision-based detection method for the terminal motion and posture of a parallel mechanism. After adopting the above technical solution, it has the following beneficial effects:

1. The present invention adopts the Harris-SIFT algorithm to realize stereo matching in the stereo matching stage, extracts image feature points through a simple Harris operator, and then uses SIFT feature descriptors to match images, so that the matching results are both real-time and stable ;

2. Aiming at the error matching and wrong matching problems in the matching algorithm, the present invention proposes a new purification algorithm, that is, obtaining more accurate image model parameters by taking points in blocks and checking the temporary model by taking points in advance to quickly determine the image model The method is used to improve the purification algorithm to purify the Harris-SIFT matching results and eliminate the wrong matching points existing in the Harris-SIFT algorithm. While ensuring the real-time performance of the matching algorithm, the accuracy of the matching is improved, thereby improving the three-dimensional pose detection at the end of the mechanism. accuracy.

Description of drawings

The present invention will be described in further detail below in conjunction with the accompanying drawings and specific embodiments.

Fig. 1 is a flow chart of a binocular vision-based terminal pose detection system for a parallel mechanism proposed by the present invention.

Figure 2 is an image denoising effect diagram of parallel mechanism based on wavelet transform method. Among them: a is the image with noise, and b is the image after denoising.

Figure 3 is a Harris feature point extraction diagram.

Figure 4 is an effect diagram of feature point stereo matching based on the Harris-SIFT algorithm.

Fig. 5 is an effect diagram after the matching result is purified.

Fig. 6 is a principle diagram of binocular vision imaging of the method of the present invention.

Fig. 7 is the fixed coordinate system and the moving coordinate system established on the mechanism platform.

Fig. 8 is a tracking error graph of the end pose parameters of the mechanism obtained by comparing the experimental results of the method of the present invention with the measurement results of the laser rangefinder and the electronic compass.

detailed description

The invention provides a binocular vision-based detection method for the terminal pose of a parallel mechanism. In order to solve the problem of poor real-time performance in the binocular vision matching process, the invention uses the Harris-SIFT algorithm to achieve stereo matching; however, there are errors in the Harris-SIFT algorithm For the problem of matching and error matching, the present invention proposes a purification algorithm to eliminate the mismatching points existing in the Harris-SIFT algorithm, while ensuring the real-time performance of the matching algorithm, improve the matching accuracy rate, and then make the real-time performance of pose detection and accuracy are improved. Specific steps are as follows:

1. Referring to Figure 1, the original image of the parallel mechanism is collected through the binocular vision left and right cameras. The image information of the parallel mechanism is obtained in real time through the binocular camera, and the camera should be adjusted to a suitable position to ensure that the end of the parallel mechanism is always within the camera's field of view during the movement of the parallel mechanism.

2. Referring to Fig. 1 and Fig. 2, the original images collected by the left and right cameras are subjected to image denoising preprocessing operations based on wavelet transform. The human-machine interface of the detection system is developed by using the Visual Studio integrated development tool, and the C/C++ language is used for programming. At the same time, combined with OpenCV (open source cross-platform computer vision library), the programming of the related functions of each module is realized. In the denoising process, the image wavelet threshold denoising (WaveDec()) method is used to preprocess the original institutional image. By setting the threshold (g_HighPassFilterValue), the high-frequency wavelet coefficients are thresholded, and finally the image is reconstructed by wavelet (WaveRec ()) to achieve image denoising.

3. Referring to Figure 1 and Figure 3, use the corner point extraction method based on the Harris operator to extract the characteristics of the institutional image. In order to ensure the rapidity and stability of feature extraction, considering that the parallel mechanism image is a rigid image with multiple corners, Harris corner detection is not only simple and stable, but also not easily affected by illumination and rigid geometric deformation. Therefore, The Harris corner detection algorithm is used to extract the feature points of the institutional image. Harris feature point extraction can be implemented by directly calling the OpenCV library function ((conerHarris()). In Harris feature point detection, the second-order moment of a pixel can be expressed as:

In the formula, I _x , I _y represent the gradient of the pixel gray level in the x and y directions; I _xy is the second-order mixed partial derivative. Therefore, the Harris response function of the pixel is:

R＝det Mk(traceM) ² (2)

Where det M is the determinant of matrix M, traceM is the true trace of matrix M; k is an empirical constant, usually in the range of 0.04 to 0.06. Set the threshold t, and calculate the response value R according to the above Harris operator. If R<t, the point is a feature point.

4. Referring to Figure 1 and Figure 4, the Harris-SIFT algorithm is used to achieve stereo matching of feature points. By using the Harris feature points to replace the extreme points in SIFT, and defining the main direction for each feature point, generating feature vector descriptors based on the feature points, and then matching the feature points according to the Euclidean distance of the SIFT matching algorithm, the Harris corner points The detection algorithm is organically combined with the SIFT scale-invariant feature transformation, so that the obtained feature points have the common characteristics of the two algorithms, which can not only improve the real-time performance of the SIFT operator, but also improve the stability of the feature points. The feature matching based on the Harris-SIFT algorithm needs to be implemented by programming, and the following formula is used when writing the program:

Among them: Formula (3) is the Taylor expansion of the DoG operator of the feature point x in the scale space, D(x) represents the spatial scale function, △x is the offset of the feature point x; formula (4) is the exact The offset of the position; the formula (5) represents the judgment condition of the non-edge point, r is the set threshold, H is the Hessian matrix of 2×2, Tr(H) is the true trace of the H matrix, and Det(H) is H The determinant of the matrix; formula (6) is a 2×2 Hessian matrix, D _xx represents the second-order partial derivative of the DOG operator of the feature point along the x direction, and D _yy represents the second-order partial derivative of the DOG operator of the feature point along the y direction Partial derivative, D _xy represents the second-order partial derivative of the DOG operator of the feature point along the xy direction; formula (7) represents the modulus m(x,y) and direction θ(x,y) of the Gaussian gradient at the feature point, L (x, y) represents the scale space value where the feature point is located.

5. Referring to Figure 1 and Figure 5, a new purification algorithm is used to purify the matching results and eliminate false matching pairs. Aiming at the problem of wrong matching and error matching in the Harris-SIFT algorithm, the present invention adopts a new purification algorithm to further purify the matching results, so as to eliminate the wrong matching points existing in the Harris-SIFT algorithm, while ensuring the real-time performance of the matching algorithm, Improve the accuracy of matching, thereby improving the accuracy of three-dimensional pose detection at the end of the mechanism. When performing the purification process, since the key of the purification algorithm is to quickly find the correct target model, and the target model is obtained by optimizing the candidate model containing the largest number of internal points, in order to quickly obtain the candidate model parameters, the present invention randomly selects 9 matching pairs , where 8 matching pairs are used to obtain the temporary model parameters, and the ninth matching pair is used to verify the correctness of the temporary model, thereby quickly determining the target model; at the same time, the present invention adopts image segmentation and randomly selects The method of one matching point avoids the problem of inaccurate basic matrix caused by the fact that two matching points are too close to be considered as one point. The specific steps to obtain the target model are as follows:

[1] According to the coordinate boundaries of the matching points in the first picture, the picture is divided into b×b (b>3) blocks on average, and 9 different blocks are randomly selected and a matching point is randomly selected in each block to form 9 A random sample set S of matching point pairs.

[2] Randomly select 8 matching point pairs from the random sample set S to obtain the temporary candidate model F.

[3] Check whether the ninth point is the support set of the temporary candidate model F: if yes, then the temporary candidate model F is the candidate model F; otherwise, reselect 9 pairs of matching points and repeat the process of [1] and [2].

[4] Detect all matching point pairs through the candidate model F and the threshold k, and obtain the support set logarithm m of the candidate model D.

[5] If m≥threshold t, get the target model F; otherwise, reselect 9 pairs of matching points and repeat the process of [1], [2], [3], [4].

[6] Optimize the target model F, and finally determine the target model H.

[7] Each time a random sample set is selected, the counter is incremented by 1. If the model parameters are not found after k times of repetition, the program is terminated.

6. Referring to Figure 1 and Figure 6, realize 3D reconstruction according to the binocular vision principle, bring the end matching point pairs extracted from the previous image processing into the binocular vision model, and finally obtain the terminal pose information of the parallel mechanism. During three-dimensional reconstruction, firstly, the camera parameters of the built binocular vision model need to be obtained. The camera parameters can be obtained through camera calibration. The present invention adopts the Zhang Zhengyou calibration method to perform camera calibration on the binocular vision system. According to the Zhang Zhengyou plane calibration process: first, fix the camera , then take pictures of the calibration plate, input the image of the calibration plate into the computer and obtain the pixel coordinates of each grid point on the calibration plate, and then load the known three-dimensional coordinates of the calibration plate into the computer, by combining the pixel coordinates of the calibration plate and the three-dimensional The coordinates are substituted into the calibration model, and the internal and external parameters of the camera are solved. After the calibration of the left and right cameras, the camera parameters of the binocular vision system can be obtained according to formula (8):

In the formula, R _l , T _l and R _r , T _r represent the rotation matrix and translation matrix of the left and right cameras, respectively.

After the camera calibration is completed, the binocular vision projection matrix can be obtained. Assuming that the image points of the terminal feature point P of the parallel mechanism in the left and right cameras are respectively Pl and Pr, the left and right cameras have been calibrated and their projection matrices are respectively Ml and Mr, then according to the projection theorem:

Among them, s _l and s _r represent scale factors, and the above formula can be expanded as follows:

In the formula, (u ₁ ,v ₁ ,1), (u ₂ ,v ₂ ,1) represent the homogeneous coordinates of the projection points P _l and P _r respectively; (X,Y,Z,1) are the terminal feature points at Homogeneous coordinates in the world coordinate system; Indicates the i-th row and j-th column of the left-right projection matrix.

After formula (10) and formula (11) are combined, the following equation can be obtained:

make: Then the least square method can be used to obtain the three-dimensional coordinates P _W of the terminal position of the parallel mechanism, which is expressed as follows:

P _w ＝(A ^T A) ^-1 A ^T B (13)

Then the rotation matrix is expressed as shown in formula (14):

Then the terminal attitude angle of the parallel mechanism is shown in formula (15):

In the formula, α is the pitch angle of rotation around the X axis (unit: rad), β is the roll angle of rotation around the Y axis (unit: rad), and γ is the heading angle of rotation around the Z axis (unit: rad) .

So far, the solution of the terminal pose parameters of the parallel mechanism has been completed.

Example

The present invention emphatically proposes a new binocular vision-based detection method for the terminal motion pose of a parallel mechanism to solve the problems of poor real-time performance and low detection accuracy existing in traditional binocular vision algorithms in the detection of the terminal pose of a parallel mechanism. The end of a new type of 3-DOF parallel mechanism is used as the detection object below, and the specific implementation of the detection method is as follows:

1. Collect the original image of the parallel mechanism. The image of the new 3-DOF parallel mechanism is collected by the Weishi binocular camera. The camera model is MV-1300FM and the lens model is AFT-0814MP. The camera is installed in front of the parallel mechanism, and the height and angle from the ground can be flexibly adjusted. In addition, the detection system uses a computer operating system of Windows 7, a processor model of Intel(R) Core(TM) 2Duo, a main frequency of 2.66GHz, and a memory of 2GB.

2. Denoise the original image based on wavelet transform. In order to reduce the adverse effect of noise on the detection of the terminal pose of the parallel mechanism, the original image is preprocessed with a denoising method based on wavelet transform, and the image noise is effectively suppressed through the denoising process, so that the next feature extraction operation, denoising The effect is shown in Figure 2.

3. Use the corner point extraction method based on Harris operator to extract the image characteristics of the mechanism. Considering that the new 3-DOF parallel mechanism has obvious rigid body characteristics, and the Harris feature point detection not only has a high re-inspection rate in terms of rigid body geometric deformation and brightness change, but also has a simple and stable algorithm, so the Harris corner detection algorithm is selected for the mechanism image Extract feature points. The Harris response function of the pixel is:

R＝det Mk(traceM) ² (16)

Where det M is the determinant of matrix M, traceM is the true trace of matrix M; k is an empirical constant, usually in the range of 0.04 to 0.06. Set threshold t, calculate response value R according to above-mentioned Harris operator, if R<t, then this point is feature point, present embodiment gets empirical constant k=0.04, and directly by calling OpenCV library function ((conerHarris ()) The feature point extraction of the new 3-DOF parallel mechanism is realized, and the feature point extraction diagram of the new 3-DOF parallel mechanism is shown in Figure 3.

4. The Harris-SIFT algorithm is used to realize the stereo matching of feature points. In this embodiment, the Harris-SIFT algorithm is used to match the image feature points of the novel 3-DOF parallel mechanism. The matching algorithm uses Harris feature points to replace the extreme points in SIFT, and uses the method of fitting a three-dimensional quadratic function to achieve precise positioning of feature points, and removes low-contrast points that are sensitive to noise and points on the edge of the image that are difficult to locate , so as to filter out stable points as the SIFT feature points of the image. Stereo matching based on the Harris-SIFT algorithm is divided into the following steps:

(1) Perform Taylor expansion on the DoG operator of the feature point x:

Among them, △x is the offset of the feature point x, when x is the extreme point of the DoG operator, the first-order partial derivative of D(x) is 0, then the offset △x of the precise position of the feature point can be obtained have to:

After multiple iterations, the precise position and scale of the measured point can be obtained, and the DoG value of the point can be obtained by substituting it into formula (16), and its absolute value can be obtained, and the contrast threshold Tc can be set, so as to eliminate the contrast whose absolute value is less than the threshold Tc low-contrast points.

For the elimination of edge points, it can be realized by comparing the main curvature of the extreme point of the DoG function with the threshold. Usually, the extreme point of the DoG function of the edge is larger than the main curvature of the non-edge point. Therefore, the main curvature ratio can be greater than the threshold. The edge points are eliminated. The non-edge points satisfy the relationship as shown in formula (19):

Wherein, r is the set threshold, and the present embodiment takes r=10; H is a 2×2 Hessian matrix, and the H matrix is expressed as shown in formula (20):

(2) Determine the direction of the feature point. Let the gradient modulus of point L(x, y) be m(x, y) and the direction be θ(x, y), then the Gaussian gradient modulus and direction of this point are expressed as formula (21):

The gradient distribution of pixels in the neighborhood of the key point is counted, and the main peak of the histogram is selected as the direction of the key point. So far, each feature point has three important information: position, scale and direction.

(3) Generate feature descriptors. To generate a feature descriptor, you need to rotate the coordinate axis to the direction of the feature point, and evenly divide the 16×16 pixel area adjacent to the feature point into 8 4×4 sub-regions, and then calculate on each 4×4 sub-region Gradient histograms in 8 directions, so that each feature point can generate 128 data to form a 128-dimensional SIFT feature vector, and normalize the length of the generated feature vector to further reduce the impact of illumination changes.

(4) After the feature descriptor is generated, the feature points can be matched according to the Euclidean distance of the SIFT matching algorithm. The specific implementation is as follows: First, calculate the Euclidean distance between a certain feature point p in the left image and all the feature points in the right image to be matched; Two points p1 and p2; finally, calculate the ratio of the closest Euclidean distance to the next closest Euclidean distance and compare it with a threshold. If the ratio is less than the threshold, the match is successful. At this time, (p1,p2) is a sequence of images Match point pairs, otherwise, the match fails. In this embodiment, the threshold is set to 0.5, and the matching effect of the new 3-DOF parallel mechanism based on the Harris-SIFT algorithm is shown in FIG. 4 .

5. Use the purification algorithm proposed by the present invention to eliminate the mismatching points existing in the matching result. During the purification process, the target model of matching point pairs can be calculated based on a set of data sets containing abnormal points, and then the remaining points can be used to test the model to obtain valid data samples. In order to quickly find out the candidate model parameters, the picture is divided into 16 blocks on average according to the coordinate boundaries of the matching points in the left figure, and 9 different blocks are randomly selected, and a matching point is randomly selected in each block, and from the 9 matching pairs Eight matching pairs are randomly selected to obtain the temporary model parameters, and the ninth matching pair is used to verify the correctness of the temporary model, so as to quickly determine the target model. The effect of the purified matching result is shown in Figure 5. The selection of the target model H and the Sampson distance are expressed as follows:

(1) Selection of the target model H between the two images: Randomly select 8 matching point pairs according to the above block-taking method to calculate the temporary candidate model F, and calculate the Sampson distance d of the matching point pair corresponding to the model F, set Threshold k, the points that meet the condition of d<k are used as inliers, and the candidate model F with the largest number of inliers is further optimized to obtain the target model H. Finally, the matching points are re-estimated through the target model H, and false matching pairs are eliminated. The representation of the projection transformation model is shown in formula (22):

In the formula, (x, y, 1) and (x’, y’, 1) represent the homogeneous coordinates of matching point pairs m(x, y) and m(x’, y’), respectively.

(2) Representation of the Sampson distance d: For any pair of matching points m(x, y) and m(x’, y’), its Sampson distance d in the candidate model F can be expressed as:

6. Solve the terminal motion pose of the new 3-DOF parallel mechanism. According to the motion characteristics of the new 3-DOF parallel mechanism, the fixed coordinate system and the dynamic coordinate system of the mechanism platform are established as shown in Figure 7, where {B}={O-XYZ}, {T}={O'-X'Y 'Z'} represent the fixed coordinate system and the moving coordinate system respectively, then the change of the end pose of the parallel mechanism can be expressed as the change of the pose of the center point O' of the moving platform. In this embodiment, Zhang Zhengyou calibration method is used to realize the camera calibration of the binocular vision system, and the terminal position information of the new 3-DOF parallel mechanism is obtained through the early image processing and the later coordinate transformation as follows:

P _w ＝(A ^T A) ^-1 A ^T B (24)

in, (u ₁ , v ₁ ), (u ₂ , v ₂ ) represent the image coordinates of the end feature points on the left and right camera projection points respectively, Indicates the i-th row and j-th column of the left-right projection matrix.

At the same time, the terminal attitude angle of the new 3-DOF parallel mechanism is shown in formula (25):

Since the parallel mechanism performs three-degree-of-freedom motion, its translational displacement on the Y-axis and Z-axis, and the roll angle β around the Y-axis can be obtained, and the pose parameters of the mechanism are (Y, Z, β).

In this embodiment, 12 images are regularly collected, and then the three-dimensional pose parameters of the mechanism are calculated at each moment, and the results are compared with the actual pose parameters of the end of the mechanism measured by the laser rangefinder INR-Laser Scanner H and the electronic compass Honeywell-HMR3100 In contrast, the tracking error is shown in Figure 8. Analyzing Figure 8, it can be seen that the tracking errors of measurement points P1 and P7 in the Y direction are relatively large, while the tracking errors of P4 and P10 in the Z direction are relatively large, which is related to the symmetry and running track of the mechanism. Changes in velocity and acceleration lead to large changes in velocity and acceleration, plus the influence of mechanism inertia, resulting in large errors. This shows that the experimentally measured end pose motion state of the parallel mechanism is consistent with the actual motion state; at the same time, according to the tracking error can be calculated The maximum deviation of each measurement point is Y direction: 0.622mm, Z direction: 0.782mm, β rotation angle: 0.677°, indicating that the detection has high precision. The experimental results show that the pose detection method based on binocular vision proposed by the present invention can better realize the real-time detection of the motion pose of the end of the parallel mechanism.

To sum up, a binocular vision-based terminal motion pose detection method of a parallel mechanism of the present invention, firstly, performs image denoising preprocessing based on wavelet transform on the collected parallel mechanism image to eliminate the impact of image noise on subsequent feature extraction. The influence caused; then, the Harris-SIFT algorithm is used to perform feature matching on the image of the parallel mechanism. Different from the general matching algorithm, this matching algorithm first extracts the feature points of the image through the Harris operator, and then uses the SIFT feature descriptor to match the feature points extracted by the Harris operator, so that the matching results are both real-time and stable; then , in order to further improve the matching accuracy, a new purification algorithm is proposed to purify the matching results. The purification algorithm solves the problem of time-consuming acquisition of the target model and model Solve the inaccurate problem; finally, bring the pair of feature points at the end of the parallel mechanism after matching and purification into the binocular vision model, and obtain the three-dimensional pose of the end of the mechanism through coordinate transformation. The binocular vision-based terminal pose detection method of the parallel mechanism proposed by the present invention first extracts image feature points through the Harris operator in the matching stage, and then uses the SIFT feature descriptor to match the feature points extracted by the Harris operator. The image processing time can be greatly reduced, and the matching result can be purified through the proposed new purification algorithm, which can further improve the matching accuracy rate, thereby improving the real-time and accuracy of the terminal pose detection of the parallel mechanism.

In the description of this specification, references to the terms "one embodiment," "some embodiments," "exemplary embodiments," "example," "specific examples," or "some examples" are intended to mean that the implementation A specific feature, structure, material, or characteristic described by an embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

Although the embodiments of the present invention have been shown and described, those skilled in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principle and spirit of the present invention. The scope of the invention is defined by the claims and their equivalents.

Claims (6)

1. A binocular vision-based parallel mechanism tail end motion pose detection method is characterized by comprising the following steps:

step 1) acquiring an original image of a parallel mechanism by using a binocular camera, and performing image denoising pretreatment based on wavelet transformation on the acquired mechanism image;

step 2) extracting the characteristics of the mechanism image by adopting an angular point extraction method based on a Harris operator;

step 3) carrying out feature matching on the parallel mechanism image by adopting a Harris-SIFT algorithm; firstly, extracting image feature points through a Harris operator, and then matching the feature points extracted by the Harris operator by utilizing an SIFT feature descriptor;

step 4) aiming at the problems of error matching and error matching of the Harris-SIFT algorithm, a new purification algorithm is adopted to further purify the matching result: the purification algorithm is improved through a block point taking and a point taking in advance check calculation temporary model, wherein the block point taking: a mode that the image is divided into blocks and a matching point is randomly selected from each block is adopted; taking points in advance to check a temporary model: randomly selecting 9 matching pairs, solving parameters of the temporary model by using 8 matching pairs, and verifying the correctness of the temporary model by using the 9 th matching pair so as to quickly determine the image model;

and 5) calculating the three-dimensional coordinates of the tail end characteristic points through the tail end characteristic points extracted by the previous image processing according to a binocular vision principle, calculating tail end attitude angles of the parallel mechanism, and finally obtaining tail end pose information of the parallel mechanism.

2. The binocular vision-based parallel mechanism end motion pose detection method according to claim 1, and is characterized in that: in step 2), the angular point extraction method of the Harris operator may directly call an OpenCV library function to implement, and in Harris feature point detection, the second moment of the pixel point may be expressed as:

$M=\left[\begin{array}{cc}{I}_x^2& I_x{I}_y\\ I_x{I}_y& I_y^2\end{array}\right]$

in the formula I_x、I_yExpressing the gradient of the gray level of the pixel point in the x and y directions; i is_xyIs a second order mixed partial derivative; the Harris response function of the pixel point is as follows:

R＝detM-k(traceM)²

wherein detM is determinant of matrix M, traceM is true trace of matrix M; k is an empirical constant, and the value range is usually 0.04-0.06; and setting a threshold value t, calculating a response value R according to the Harris operator, and if R < t, taking the point as a characteristic point.

3. The binocular vision-based parallel mechanism end motion pose detection method according to claim 1, and is characterized in that: in the step 4), the purification algorithm calculates a target model of the matching point pairs through a group of data sets containing abnormal points, and then tests the model by using the remaining points, so as to obtain an effective data sample; the method comprises the following specific steps:

4.1) dividing the picture into b x b blocks according to the coordinate boundary of the matching points in the first picture, wherein b is more than 3, randomly selecting 9 different blocks and randomly selecting one matching point in each block to form a random sample set S of 9 matching point pairs;

4.2) randomly selecting 8 matching point pairs from the random sample set S to obtain a temporary candidate model F;

4.3) detecting whether the 9 th point is a support set of the temporary candidate model F: if yes, the temporary candidate model F is a candidate model F; otherwise, 9 pairs of matching points are reselected, and the processes of 4.1) and 4.2) are repeated;

4.4) detecting all matching point pairs through the candidate model F and a threshold k to obtain a support set logarithm m of the candidate model D;

4.5) if m is larger than or equal to the threshold t, obtaining a target model F; otherwise, 9 pairs of matching points are reselected, and the processes of 4.1), 4.2), 4.3) and 4.4) are repeated;

4.6) optimizing the target model F, and finally determining a target model H;

4.7) every time a random sample set is selected, the counter is incremented by 1, and if no model parameters are found after repeating k times, the procedure is terminated.

4. The binocular vision-based parallel mechanism end pose detection method according to claim 1, and is characterized in that: in the step 5), three-dimensional coordinates (X, Y, Z) of the characteristic points at the tail end of the parallel mechanism are obtained according to the projection theorem, and then the attitude angle of the tail end of the parallel mechanism is obtained through the camera rotation matrix R; the position information of the tail end of the parallel mechanism is as follows:

P_w＝(A^TA)^-1A^TB

wherein,(u₁,v₁)、(u₂,v₂) Respectively representing the image coordinates of the terminal feature points at the left and right camera projection points;the ith row and the jth column of the left and right projection matrixes are represented;

meanwhile, the tail end attitude angle of the parallel mechanism is as follows:

$\left\{\begin{array}{c}\mathrm{\&alpha;}=\arctan \left(\frac{R_{23}}{R_{33}}\right)\\ \mathrm{\&beta;}=-\arcsin R_{13}\\ \mathrm{\&gamma;}=\arctan \left(\frac{R_{12}}{R_{11}}\right)\end{array}\right.$

wherein R is_ij(i 1,2, 3; j 1,2,3) denotes the i-th row of the rotation matrix R, the j-th column, α, is the pitch angle of rotation about the X-axis in rad, β is the roll angle of rotation about the Y-axis in rad, and γ is the course angle of rotation about the Z-axis in rad.

5. The binocular vision-based parallel mechanism end pose detection method according to claim 1, and is characterized in that: in the step 5), the method further comprises the steps of, during three-dimensional reconstruction, firstly obtaining camera parameters of the established binocular vision model, wherein the camera parameters can be obtained through camera calibration; and calibrating the camera of the binocular vision system by adopting a Zhangyingyou plane calibration method.

6. The binocular vision-based parallel mechanism end pose detection method of claim 5, which is characterized in that: the Zhangzhengyou plane calibration method comprises the following steps: firstly fixing a camera, then shooting a calibration plate, inputting an image of the calibration plate into a computer, acquiring a pixel coordinate of each grid point on the calibration plate, loading a known three-dimensional coordinate of the calibration plate into the computer, solving internal and external parameters of a camera by substituting the pixel coordinate and the three-dimensional coordinate of the calibration plate into a calibration model, and respectively calibrating a left camera and a right camera.