A multi-vision system calibration method based on Levy flight

Abstract

The present invention discloses a multi-vision system calibration method based on Levy flight, comprising the following steps: acquiring a spherical calibration device image through a camera and extracting a rectangular calibration image attached to the spherical calibration device; extracting the centroid of the triangular code points and the centroid of the circular template in each acquired rectangular calibration image; calculating the matrix representation corresponding to each rectangular calibration image; calculating the similarity between the matrices acquired from the two spherical calibration device images, thereby completing the matching of the rectangular calibration images and determining the common view key points; calculating the extrinsic parameters of the camera in the world coordinate system based on the coordinates of the common view key points; constructing an adjacency matrix containing specific parameters; determining the flight step size and flight direction of the Levy flight algorithm based on the weight values ​​of the common view key points and the camera connection, thereby obtaining the extrinsic parameter relationship between the initial camera and the target camera. This method utilizes a small calibration object with low precision to calibrate high-precision measurement system parameters under a large field of view.

Description

Technical Field

[0001] The present invention relates to the field of machine vision technology, and in particular to a multi-vision system calibration method based on Levy flight. Background Art

[0002] The technology for achieving high-efficiency and high-precision calibration of multi-vision measurement systems with a large field of view (5m*5m) has broad application prospects. However, there are currently technical challenges, which are mainly reflected in the following aspects:

[0003] The first challenge stems from the low efficiency and limited applicability of large-FOV measurement systems. Large-FOV calibration requires more calibration feature points, which increases the number of calibration images required, making the calibration process more complex and inefficient. Furthermore, due to the time-consuming calibration process, this approach is not suitable for systems requiring rapid calibration.

[0004] The second challenge comes from the high calibration error for large field of view. On the one hand, the calibration distance of large field of view measurement systems is usually long (>5m). Existing calibration methods usually have the problem of high calibration accuracy in the two-dimensional plane and low calibration accuracy in the depth direction when performing long-distance calibration. As a result, the calibrated measurement system parameters have large differences in accuracy in the three-dimensional direction, which cannot meet the calibration accuracy requirements of large field of view measurement systems in field environments. On the other hand, small-sized calibration plates can achieve high-precision calibration for large fields of view: From the analysis of measurement system calibration methods based on calibration objects, although such methods can achieve high calibration accuracy, when calibrating large field of view measurement systems, due to manufacturing difficulties and other issues, there are often difficult-to-solve problems such as the mismatch between the calibration plate size and the measurement field of view of the measurement system, resulting in calibration accuracy that cannot meet the requirements of large field of view calibration.

[0005] Therefore, how to break the contradiction between the field of view range and the size of the calibration object and use a small calibration object with lower precision to calibrate high-precision measurement system parameters under a large field of view is a difficult problem that needs to be solved in the calibration of large field of view measurement systems. Summary of the Invention

[0006] In response to the above two technical difficulties in multi-vision system calibration in large field of view measurement applications, the present invention proposes a multi-vision system calibration method based on Levy flight, providing a method architecture with practical reference value for the implementation of target measurement applications of multi-vision systems in large field of view.

[0007] In order to solve the above technical problems, the technical solution of the present invention is:

[0008] A multi-vision system calibration method based on Levy flight includes the following steps:

[0009] Step 1: Place the custom spherical calibration device in the public field of view of all cameras, obtain the image of the spherical calibration device through the camera, and extract the rectangular calibration image of the cover on the spherical calibration device;

[0010] Step 2: Extract the centroid of the triangle coding points and the centroid of the circular template in each obtained rectangular calibration image;

[0011] Step 3: Calculate the matrix representation corresponding to each rectangular calibration image according to the custom encoding rules;

[0012] Step 4: Calculate the similarity between the matrices obtained from the two spherical calibration device images to complete the rectangular calibration image matching. The two spherical calibration device images are obtained by two arbitrary cameras, and the centroids of all triangle code points and the centroid of the circular template are used as the common view key points between the two cameras that obtained the matching spherical calibration device images.

[0013] Step 5: Based on the coordinates of the common view key points, the extrinsic parameters of the camera in the world coordinate system are calculated using epipolar geometry and SVD decomposition.

[0014] Step 6: Construct an adjacency matrix containing specific parameters, where the elements of the adjacency matrix include external parameters, common view key points, and weight values ​​of camera connections in the common field of view;

[0015] Step 7: Determine the flight step size and flight direction of the Levy flight algorithm based on the weight values ​​of the common view key points and camera connections;

[0016] Step 8: Use the Lévy flight algorithm in step 7 to find the next transition camera from the initial camera. After finding the next camera, continue to execute the Lévy flight algorithm until the target camera is found.

[0017] Step 9: After searching for a complete transition path, determine a complete extrinsic parameter transformation path to obtain the extrinsic parameter relationship between the initial camera and the target camera.

[0018] Preferably, in step 1, each camera has a common field of view with at least one other camera.

[0019] Preferably, the spherical calibration device satisfies:

[0020] The spherical calibration device is used to place a number of rectangular calibration patterns;

[0021] The rectangular calibration pattern includes 6 circular templates and 2 triangular coding points;

[0022] The circular template forms the identification points of a rectangular calibration pattern, wherein one triangular coding point is located at the center point of the rectangle;

[0023] The triangular coding points are freely arranged in the rectangular calibration pattern, and any two triangular coding points do not overlap with each other.

[0024] As a preference, in step 3, the coding rule is as follows: set the upper left corner of the rectangular calibration pattern as the coordinate origin, right as the positive direction of the X axis, downward as the positive direction of the Y axis, define the coordinates of the center of the triangle coding point (X, Y), take the center of gravity of the central triangle coding point as point 1, take the center of gravity of the corner triangle coding point as point 2, define the coordinates of the center of gravity of the circular template (X, Y). i , Y i )(i=1...6), arranged in ascending order of X-axis coordinates, X-coordinate deviation: ΔX i =X i -X(i=1...6); Y coordinate deviation: ΔY i =Y i -Y t (i=1...6), calculate the distance between the centers of the circular templates to form a 4×6 matrix.

[0025] Preferably, in step 4, the root mean square error or correlation coefficient is used to judge the similarity. If it is less than a threshold, the two rectangular calibration images are considered to be matched; otherwise, step 4 is continued to complete the matching of the two rectangular calibration images, and the above matching steps are repeated until all rectangular calibration images acquired by the two cameras are matched.

[0026] Preferably, in step 5, the method for calculating the external parameters of the camera in the world coordinate system is:

[0027] Based on the mathematical model of pinhole camera perspective imaging, the mapping relationship between space points and image points is determined, which involves the camera intrinsic parameter matrix, unknown scale factors and homogeneous vectors of space points;

[0028] Determine the relationship between the scale factors of spatial points in different cameras, which involves the basic matrix and the pole position;

[0029] Thus, the projection equation of the key points in the common field of view of the cameras is established, and the camera extrinsic parameters of the projection space can be recovered through singular value decomposition.

[0030] Preferably, the adjacency matrix is ​​constructed by:

[0031] Each non-diagonal element of the adjacency matrix represents the association strength or information transmission cost between a pair of cameras. The element L of the adjacency matrix is i，j Stores the weight information between nodes i and j. The information of each edge is expressed as L i，j ={H i，j ,Ω i，j , w i，j}, where H i，j Indicates that from C i to Cj The transformation matrix, Ω i，j Refers to C i with C j The point set in the common view area, w i，j It represents the evaluated weight value.

[0032] As an example, when constructing the adjacency matrix, since the node graph formed by the multi-camera system is an undirected graph, for each variable H i，j and w i，j , they all satisfy the symmetry relationship, that is, H i，j =H j，i and ω i，j =ω j，i .

[0033] Preferably, the weight value is calculated as follows:

[0034] The two camera nodes C are represented by the reprojection error and the number of common view feature points. i with C j The association weight w between i，j , the weight of each edge can be represented by the sum of two components, namely:

[0035]

[0036] Among them, λ1 and λ2 represent the coefficients of the two components respectively. It can be expressed as:

[0037]

[0038] Among them i and Er j Represents camera C i and C j The average value of the reprojection error, TH is the reprojection error threshold.

[0039] Preferably, when calculating the weight value, the inverse of the number of common feature points is used to construct a second component of the association weight, and the second component is expressed as:

[0040]

[0041] where N i，j Indicates the number of feature points shared by the two cameras.

[0042] Preferably, the flight step length and flight direction are calculated as follows:

[0043] There is a key variable in the Levy flight algorithm, namely the jump step length L, which is constructed with the weight value:

[0044]

[0045] Where α and σ follow normal distribution, the variable in flight step length It has the following forms:

[0046]

[0047] Select the node with the highest transfer probability from the current node as the next target node. The transition function is:

[0048]

[0049] where ρ i→j It represents the transition probability from the current node i to the neighboring node j, κ is a parameter related to the Levy flight step, defined as κ = c·L, c is a constant greater than zero, and the step length L is a non-negative random variable.

[0050] The present invention has the following characteristics and beneficial effects:

[0051] In the era of intelligent manufacturing, machine vision technology plays a very important role. Researching and optimizing multi-viewing calibration technology is in line with the key topics and breakthrough directions of target detection and product measurement research in the current industrial field, and is of great significance. The multi-viewing system calibration method based on Levy flight provided in the embodiment of the present invention is:

[0052] Spherical objects are readily available, and a globe used in education is a perfect fit. Because certain feature points on a sphere are always orthographically projected, this helps improve feature point detection accuracy. Furthermore, it can adapt to various camera arrangements.

[0053] In the calibration of multi-camera vision systems, the Lévy flight algorithm mimics the ability of organisms to effectively search in complex, dynamic environments. This property allows the algorithm to perform fine-tuned local searches most of the time, while making large cross-domain jumps in specific situations to effectively avoid local optima. These jumps allow the algorithm to cross inefficient regions of the solution space that typically entrap search methods guided by local gradient information. The efficiency advantage of the Lévy flight algorithm becomes more pronounced as the number of cameras in the vision system increases. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0055] Figure 1 The self-encoding spherical calibration target provided by the present invention;

[0056] Figure 2 The camera relationship adjacency matrix of the multi-viewing system provided by the present invention;

[0057] Figure 3 Flowchart of external parameter calculation provided by the present invention;

[0058] Figure 4 The present invention provides a flowchart of node path optimization based on Levy flight. DETAILED DESCRIPTION

[0059] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0060] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings.

[0061] On the contrary, the present invention covers any alternatives, modifications, equivalents, and solutions that fall within the spirit and scope of the present invention as defined by the claims. Furthermore, to facilitate a better understanding of the present invention, certain specific details are described in detail below in the detailed description of the present invention. Those skilled in the art will be able to fully understand the present invention without these details.

[0062] The present invention provides a multi-vision system calibration method based on Levy flight, comprising the following steps:

[0063] Step 1: Place the custom spherical calibration device in the common field of view of all cameras. Each camera has a common field of view with at least one other camera. Use the camera to obtain the image of the spherical calibration device and extract the rectangular calibration image of the cover on the spherical calibration device.

[0064] This embodiment of the present invention designs a custom-coded spherical calibration device to enable calibration data collection for a multi-vision system at different viewing angles. A rectangular calibration image is formed by combining eight discrete geometric patterns (including two triangles and six circles). A calibration target is formed by integrating 200 rectangular calibration images and attaching them to an ellipsoidal object.

[0065] Specifically, such as Figure 1As shown, the spherical calibration device consists of a sphere with a diameter of approximately 500 mm, covered with 200 independently coded calibration elements. Each calibration element is a 30mm x 30mm rectangle. Eight geometric shapes are distributed within the rectangular calibration image, providing eight features: {M1, M2, A, B, C, D, E, F}. M1 and M2 are template points, with M1 located at the center of the rectangle and M2 at a corner. A, B, C, D, E, and F are encoding points randomly distributed within the rectangle, with no overlap. Each feature point has a diameter of approximately 5 mm.

[0066] Furthermore, the spherical calibration device satisfies:

[0067] The spherical calibration device is used to place a number of rectangular calibration patterns;

[0068] The rectangular calibration pattern includes 6 circular templates and 2 triangular coding points;

[0069] The circular template forms the identification points of a rectangular calibration pattern, wherein one triangular coding point is located at the center point of the rectangle;

[0070] The triangular coding points are freely arranged in the rectangular calibration pattern, and any two triangular coding points do not overlap with each other.

[0071] Step 2: For each rectangular calibration image obtained, extract the centroid of the triangle coding points and the centroid of the circular template.

[0072] Specifically, ellipse extraction is performed on each circular template, and a binary image is obtained, and the center of gravity coordinates corresponding to the binary image are calculated. The center point of the rectangle is point 1, and the other point is point 2; polygon fitting is performed on each triangle coding point, and a binary image is obtained, and the center of gravity coordinates corresponding to the binary image are calculated.

[0073] Step 3: Calculate the matrix representation corresponding to each rectangular calibration image according to the custom encoding rules.

[0074] Specifically, under each camera perspective, record the center of gravity points corresponding to the binary images of 8 geometric figures. When matching the same calibration individual, in the absence of an absolute world coordinate system, the operation is performed by recording the relative position of the circle center under each camera perspective. Let the upper left corner of the rectangle be the coordinate origin, the right is the positive direction of the X axis, and the downward is the positive direction of the Y axis. The triangle center of gravity coordinates (X, Y), with the center triangle point of gravity as point 1, and the corner triangle point of gravity as point 2. The circle center of gravity coordinates (X, Y) i , Y i )(i=1...6), arranged in ascending order of X-axis coordinates. X-coordinate deviation: ΔX i =X i -X(i=1...6); Y coordinate deviation: ΔYi =Y i -Y t (i=1...6). Calculate the distance between the center points to form a 4×6 matrix:

[0075]

[0076] Step 4: Calculate the similarity between the matrices obtained in the two spherical calibration device images to complete the rectangular calibration image matching. The two spherical calibration device images are obtained by two arbitrary cameras, and the centroids of all triangle code points and the centroid of the circular template are used as the common view key points between the two cameras that obtain the matching spherical calibration device images.

[0077] In this embodiment, the root mean square error or correlation coefficient is used to determine the similarity. If the root mean square error is less than a threshold, the two rectangular calibration images are considered to be matched. Otherwise, step 4 is continued to complete the matching of the two rectangular calibration images. The above matching steps are repeated until all the rectangular calibration images acquired by the two cameras are matched.

[0078] Step 5: Based on the coordinates of the common view key points, the extrinsic parameters of the camera in the world coordinate system are calculated using epipolar geometry and SVD decomposition.

[0079] Specifically, such as Figure 3 As shown, first obtain the coded target on the calibration sphere under each camera's view angle. Match the coded target under the common view area according to the coding rule. According to the mathematical model of pinhole camera perspective imaging, the spatial point P i With the image point x, i，j The mapping relationship between them can be expressed as:

[0080]

[0081] where K j For camera C j The internal parameter matrix, μ i，j is an unknown scale factor, which represents the spatial point P i to C j The projection depth, Indicates camera C j Point on the image plane The corresponding homogeneous vector, Represents the homogeneous vector corresponding to the target point. Point P i In C j and C g The relationship between the scaling factors in can be defined as:

[0082]

[0083] Among them F j，g and They refer to the fundamental matrix and pole positions between the two cameras respectively.

[0084] Assume P i In C j The projection depth in the three cameras is 1, and the projection depth of the point in the other two cameras is calculated. Based on this, the projection equation of n feature points in the common field of view of the three cameras is established in this embodiment:

[0085]

[0086] Using the SVD-based method to decompose the matrix on the left side of the following formula, whose rank is no more than 4, the 3D coordinates in the projection space can be restored and camera pose [Q s1 , Q s3 , Q s3 ] T In order to further obtain the corresponding 3D coordinates and camera poses in the Euclidean space, the transformation matrix H from the projective space to the Euclidean space is introduced. The matrix should satisfy the following constraints:

[0087]

[0088] where α i , (i=1, 2, 3) are non-zero real numbers.

[0089] Therefore, the transformation relationship from projective reconstruction to Euclidean reconstruction can be expressed as follows:

[0090]

[0091] By combining the matrix constraint formula and the transformation formula from projective reconstruction to Euclidean reconstruction, H can be solved. QH and H (-1) X corresponds to the camera pose and feature point coordinates in Euclidean space, respectively.

[0092] The calculation method of the external parameters in the above formula represents the relative pose between cameras in the initial calibration stage, as well as the transformation relationship between the spatial positions of the same point in different camera coordinates. After obtaining the above relationship, the common view relationship matrix is ​​constructed together with the common view key point information to provide the structural information of the visual system for the subsequent Levy flight algorithm. The execution process of the Levy flight algorithm will be described in Figure 4 In display.

[0093] Step 6: Construct an adjacency matrix containing specific parameters, where the elements of the adjacency matrix include external parameters, common view key points, and weight values ​​of camera connections in the common field of view.

[0094] An embodiment of the present invention provides an adjacency matrix that describes the relationship between cameras in a multi-vision system. In a multi-vision system, cameras are abstracted as vertices of a graph, and the geometric relationship between two cameras constitutes the edge of the graph. Each edge has a weight value, which also represents the common view relationship between the two cameras, and the common view relationship between all cameras constitutes an adjacency matrix. The matrix is ​​a symmetric matrix, in which each element consists of three parts: the spatial pose between the two cameras, which is empty if there is no common field of view; the feature point set under the common field of view, which is empty if there is no common calibration point; the weight value between the two cameras, the weight takes into account both the reprojection error between the two vertices (cameras) and the number of common view feature points between them. The number of common view feature points is relatively large.

[0095] Specifically, such as Figure 2 As shown, each non-diagonal element of the matrix represents the correlation strength or information transmission cost between a pair of cameras, and its value directly reflects the difficulty of these cameras cooperating in feature matching and 3D reconstruction. i，j Stores the weight information between nodes i and j. The information of each edge is expressed as L i，j ={H i，j ,Ω i，j , w i，j}, where H i，j Indicates that from C i to C j The transformation matrix, Ω i，j Refers to C i with C j The point set in the common view area, w i，j represents the weight value of the evaluation. Since the node graph of the multi-camera system is an undirected graph, for each variable H i，j and ω i，j , they all satisfy the symmetry relationship, that is, H i，j =H j，i and ω i，j =w j，i .

[0096] Further, the calculation of weight value:

[0097] The two camera nodes C are represented by the reprojection error and the number of common view feature points. i with C j The association weight w between i，j The weight of each edge can be represented by the sum of two components, namely:

[0098]

[0099] Where λ1 and λ2 represent the coefficients of the two components respectively. It can be expressed as:

[0100]

[0101] Among them i and Er j Represents camera C i and C j The average value of the reprojection error, TH is the reprojection error threshold. Generally, when the number of common feature points contained in the field of view of the two cameras increases, the accuracy of the camera pose estimation also improves. Therefore, the inverse of the number of common feature points is used to construct the second component of the association weight, so that in the process of graph construction, this embodiment takes the sparsity of common view key points into consideration. This component is expressed as:

[0102]

[0103] where N i，j Indicates the number of feature points shared by the two cameras.

[0104] Step 7: Determine the flight step size and flight direction of the Levy flight algorithm based on the weight values ​​of the common view key points and camera connections.

[0105] It is understandable that if there is a common field of view between camera C1 and camera C6, and the calibration error is relatively large, the path optimization process can be described as: determine the nodes adjacent to node C1, Figure 4 denoted as 2, 5, 8, and 7 in the example. Each time, the node with the highest transition probability is selected from the set of nodes adjacent to the current node as the next transition target. Upon transitioning to a new node, if its neighboring nodes have already appeared on the previous path, they are excluded from the selection. This process continues until the predetermined target node is reached.

[0106] Specifically, the Levy flight algorithm has a key variable, namely the jump step length L, which is constructed with the weight value:

[0107]

[0108] Where α and σ follow normal distribution, the variable in flight step length It has the following forms:

[0109]

[0110] Select the node with the highest transfer probability from the current node as the next target node. The transition function is:

[0111]

[0112] where ρ i→jIt represents the transition probability from the current node i to the neighboring node j, κ is a parameter related to the Levy flight step, defined as κ = c·L, c is a constant greater than zero, and the step length L is a non-negative random variable.

[0113] Step 8: Use the Lévy flight algorithm in step 7 to find the next transition camera from the initial camera. After finding the next camera, continue to execute the Lévy flight algorithm until the target camera is found.

[0114] Step 9: After searching for a complete transition path, determine a complete extrinsic parameter transformation path to obtain the extrinsic parameter relationship between the initial camera and the target camera.

[0115] Finally, this embodiment provides a multi-vision system path optimization algorithm based on Levy flight based on the extrinsic parameter relationship between the initial camera and the target camera, thereby realizing the optimization of the extrinsic parameters of the multi-vision system. For binocular cameras with large parameter errors in the multi-vision system, the node camera with small reprojection error is found by continuously searching for intermediate node cameras, and then the accurate extrinsic parameter transformation is reconstructed. First, the key variables of the Levy flight algorithm are clarified, and the flight step length is constructed by considering factors such as the number of feature points in the common viewing area and the reprojection error. The flight direction is constructed based on the probability distribution and the flight step length. Secondly, the iterative stopping condition of the Levy flight algorithm is established: (1) For camera combinations with a common calibration field of view, the relative change of the reprojection root mean square function value is considered. If the relative change of the objective function value is less than the given threshold δ after multiple consecutive iterations, such as K consecutive iterations, that is, Then the iteration can be stopped. This means that when the improvement of the objective function value is relatively small, the algorithm may have approached the optimal solution. (2) For camera combinations without a common calibration field of view, calculate the Euclidean distance or other appropriate distance metric of the parameter vectors between adjacent iterations. If the change distance of the parameter vector is less than a given threshold θ after K consecutive iterations, that is, ||x i+K -x i If || < θ, the iteration is terminated. This indicates that the parameters have changed little after multiple iterations, and the algorithm is likely converging. By combining these two conditions, we can determine the convergence of the Lévy flight algorithm in multi-camera vision system parameter optimization, thereby terminating the iteration in a timely manner, saving computing resources and obtaining better parameter results.

[0116] Specifically, assuming that the acquired camera transformation path is Path={L s，s , L s，1 ,…,L s，e , L e，e}, where L s，s and L e，e Represent the starting node and target node in the adjacency matrix respectively. If the path is determined, the pose transformation between cameras will have the same transformation method, that is:

[0117] H s，e =H s，1 H 1，2 ,...,H n，e

[0118] In order to obtain more accurate camera poses and point positions, the BA algorithm is introduced to optimize the reprojection error and further refine the parameter results. According to calibration theory, there is a scale difference between the camera poses and spatial point positions obtained by global optimization and the actual values. Usually, this scale difference can be obtained by calculating the ratio between the reconstructed length of the reference rod and its theoretical length. Since the distance between the target points on the calibration object is a constant value, the two farthest points, denoted as point A and point B, are used to calculate the scale factor, that is:

[0119]

[0120] where d A，B Represents the actual distance between space point A and space point B, Represent the spatial coordinates of point A and point B after reconstruction, N AB Represents the number of calculations. After obtaining the scale, calculate the space point and camera pose in the global coordinate system:

[0121]

[0122] Where T j，r and P i，r Respectively represent the camera translation vector and space point coordinates before scale transformation. j and P i Represent the translation vector and spatial point coordinates after scale transformation respectively.

[0123] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. It will be apparent to those skilled in the art that various changes, modifications, substitutions, and variations of these embodiments, including components, without departing from the principles and spirit of the present invention are still within the scope of protection of the present invention.

Claims

1. A multi-vision system calibration method based on Levy flight, characterized in that: The steps include: Step 1: Place the custom spherical calibration device in the public field of view of all cameras, obtain the image of the spherical calibration device through the camera, and extract the rectangular calibration image of the cover on the spherical calibration device; Step 2: Extract the centroid of the triangle coding points and the centroid of the circular template in each obtained rectangular calibration image; Step 3: Calculate the matrix representation corresponding to each rectangular calibration image according to the custom encoding rules; Step 4: Calculate the similarity between the matrices obtained from the two spherical calibration device images to complete the rectangular calibration image matching. The two spherical calibration device images are obtained by two arbitrary cameras, and the centroids of all triangle code points and the centroid of the circular template are used as the common view key points between the two cameras that obtained the matching spherical calibration device images. Step 5: Based on the coordinates of the common view key points, the extrinsic parameters of the camera in the world coordinate system are calculated using epipolar geometry and SVD decomposition. Step 6: Construct an adjacency matrix containing specific parameters, where the elements of the adjacency matrix include external parameters, common view key points, and weight values ​​of camera connections in the common field of view; The weight value is calculated as follows: The two camera nodes C are represented by the reprojection error and the number of common view feature points. i with C j The association weight w between i，j , the weight of each edge can be represented by the sum of two components, namely: Among them, λ1 and λ2 represent the coefficients of the two components respectively. It can be expressed as: Among them i and Er j Represents camera C i and C j The average value of the reprojection error, TH is the reprojection error threshold; When calculating the weight value, the inverse of the number of common feature points is used to construct the second component of the association weight, which is expressed as: where N i,j Indicates the number of common feature points in the fields of view of the two cameras; Step 7: Determine the flight step size and flight direction of the Levy flight algorithm based on the weight values ​​of the common view key points and camera connections; Step 8: Use the Lévy flight algorithm in step 7 to find the next transition camera from the initial camera. After finding the next camera, continue to execute the Lévy flight algorithm until the target camera is found. Step 9: After searching for a complete transition path, determine a complete extrinsic parameter transformation path to obtain the extrinsic parameter relationship between the initial camera and the target camera.

2. The multi-vision system calibration method based on Levy flight according to claim 1, characterized in that: In step 1, each camera has a common field of view with at least one other camera.

3. The multi-vision system calibration method based on Levy flight according to claim 1, characterized in that: The spherical calibration device satisfies: The spherical calibration device is used to place a number of rectangular calibration patterns; The rectangular calibration pattern includes 6 circular templates and 2 triangular coding points; The circular template forms the identification points of a rectangular calibration pattern, wherein one triangular coding point is located at the center point of the rectangle; The triangular coding points are freely arranged in the rectangular calibration pattern, and any two triangular coding points do not overlap with each other.

4. The multi-vision system calibration method based on Levy flight according to claim 1, characterized in that: In step 3, the coding rule is as follows: set the upper left corner of the rectangular calibration pattern as the coordinate origin, the right as the positive direction of the X axis, and the downward as the positive direction of the Y axis, define the coordinates of the center of the triangle coding point (X, Y), take the center of gravity of the central triangle coding point as point 1, take the center of gravity of the corner triangle coding point as point 2, and define the coordinates of the center of gravity of the circular template (X, Y). i , Y i ), arranged in ascending order of X-axis coordinates, X-coordinate deviation: ΔX i =X i -X;Y coordinate deviation: ΔY i =Y i -Y t , calculate the distance between the centers of the circular templates to form a 4×6 matrix, where the value of i ranges from 1 to 6.

5. The multi-vision system calibration method based on Levy flight according to claim 1, characterized in that: In step 4, the root mean square error or correlation coefficient is used to determine the similarity. If the similarity is less than the threshold, the two rectangular calibration images are considered to be matched. Otherwise, step 4 is continued to complete the matching of the two rectangular calibration images. The above matching steps are repeated until all the rectangular calibration images obtained by the two cameras are matched.

6. The multi-vision system calibration method based on Levy flight according to claim 1, characterized in that: In step 5, the calculation method of the camera's external parameters in the world coordinate system is: Based on the mathematical model of pinhole camera perspective imaging, the mapping relationship between space points and image points is determined; Determine the relationship between the scale factors of spatial points in different cameras; Thus, the projection equation of the key points in the common field of view of the cameras is established, and the camera extrinsic parameters of the projection space can be recovered through singular value decomposition.

7. The multi-vision system calibration method based on Levy flight according to claim 1, characterized in that: The construction method of the adjacency matrix is ​​as follows: Each non-diagonal element of the adjacency matrix represents the association strength or information transmission cost between a pair of cameras. The element L of the adjacency matrix is i，j Stores the weight information between nodes i and j. The information of each edge is expressed as L i,j ={H i,j ,Ω i,j ,w i,j }, where H i，j Indicates that from C i to C j The transformation matrix, Ω i，j Refers to C i with C j The point set in the common view area, w i，j It represents the evaluated weight value.

8. The multi-vision system calibration method based on Levy flight according to claim 7, characterized in that: When constructing the adjacency matrix, since the node graph composed of the multi-camera system is an undirected graph, for each variable H i，j and w i，j , they all satisfy the symmetry relationship, that is, H i，j =H j，i and w i，j =w j,i .

9. The multi-vision system calibration method based on Levy flight according to claim 7, characterized in that: The calculation method of the flight step length and flight direction is: There is a key variable in the Levy flight algorithm, namely the jump step length L, which is constructed with the weight value: Where α and σ follow normal distribution, the variable in flight step length It has the following forms: Select the node with the highest transfer probability from the current node as the next target node. The transition function is: where ρ i→j It represents the transition probability from the current node i to the neighboring node j, κ is a parameter related to the Levy flight step, defined as κ = c·L, c is a constant greater than zero, and the step length L is a non-negative random variable.

Images