A 3D candidate positioning method based on multiple fields of view in the examination room

Abstract

The present invention provides a 3D examinee positioning method based on multiple fields of view in an examination room, which relates to the field of computer vision technology. It solves the problem that traditional deep learning-based methods in the prior art usually require the construction of a custom data set, which is very complex and difficult to obtain. Some studies also use depth cameras to obtain three-dimensional information of the target, but the price of depth cameras is too high compared to ordinary monocular cameras, which brings unnecessary cost and technical problems. The technical solution is: use the examinee detection information in multi-view images to restore the 3D position of the examinee in the examination room; the present invention can estimate the examinee's 3D position from a 2D image only by a geometric method, and does not require a large amount of labeled data for training, so it has more advantages in terms of computational efficiency and computational cost.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to a 3D examinee positioning method based on multiple fields of view in an examination room. Background Art

[0002] With the development of computer vision and image processing technology, target positioning technology under multiple monocular cameras has become an important research direction. In recent years, with the modernization of education, the traditional way of invigilating examinations has problems of low efficiency and high labor costs. Therefore, it is of great practical significance to apply 3D candidate identification and positioning technology to the examination room. This method can accurately identify and locate candidates and improve the efficiency and fairness of examination room management. In a complex examination room environment, since the detection space of a monocular camera is too small and it is difficult to complete the task of accurately locating candidates, multiple monocular cameras are introduced. Joint processing based on multiple monocular cameras can expand the detection range and realize multi-angle three-dimensional space modeling, so as to more accurately identify and locate candidates. Through cross-validation of multiple perspectives, false detection can be reduced, and the accuracy and robustness of candidate positioning can be significantly improved.

[0003] Although multi-camera target localization technology has received long-term attention from the academic community, it still faces many challenges. Traditionally, target localization is mainly used on the two-dimensional image plane, but in recent years, many research methods have extended this technology to three-dimensional space, thus transforming the target detection problem into a three-dimensional target localization problem.

[0004] Traditional deep learning-based methods usually require the construction of custom datasets, which are very complex and difficult to obtain. Some studies also use depth cameras to obtain the three-dimensional information of the target, but the price of depth cameras is too high compared to ordinary monocular cameras, which brings unnecessary costs. Summary of the invention

[0005] Therefore, the present invention solves the problem that traditional deep learning-based methods in the prior art usually require the construction of custom data sets, which are very complex and difficult to obtain; there are also some studies that use depth cameras to obtain three-dimensional information of the target, but the price of depth cameras is too high compared to ordinary monocular cameras, which brings unnecessary cost and technical problems; the candidate consistency determination strategy and 3D positioning method based on multiple fields of view in the examination room provided by the present invention use the candidate detection information in the multi-view image to restore the 3D position of the candidate in the examination room; the method restates the 3D candidate positioning problem as an optimization problem in the dual space, so that the 3D position of each candidate in the examination room can be obtained with only three viewing angles; unlike the method based on deep learning, the present invention can estimate the 3D position of the candidate from the 2D image only by geometric methods, and does not require a large amount of labeled data for training, so it has more advantages in terms of computational efficiency and computational cost.

[0006] The inventive concept of the present invention is: the present invention uses the 2D candidate boundary box identified by the yolov8 algorithm and the geometric relationship of multiple views to achieve accurate 3D candidate positioning. First, each camera is calibrated to obtain the matrix relationship between each camera. Then, the yolov8 algorithm is used in the general image sequence to identify the candidate boundary box and convert it into an inscribed ellipse. Then, through the matrix relationship between multiple cameras, the corresponding three-dimensional ellipsoid surface is estimated through the inscribed ellipse under multiple views in the dual space, and finally the positioning of the candidate in the 3D space is achieved. The invention also introduces data preprocessing technology to perform standardization of the inscribed ellipse and centralization of the ellipsoid, thereby improving the accuracy of the calculation of the three-dimensional position of the candidate.

[0007] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solution: a 3D examinee positioning method based on multiple fields of view in an examination room, comprising the following steps:

[0008] S1: Calibrate each camera to obtain the camera’s intrinsic parameter K f , rotation matrix R f and the translation vector t f ;

[0009] S2: Acquire multi-view image frames f=1, 2, ..., F, which represent the examination room scenes under different cameras;

[0010] S3: Use the YOLOv8 algorithm to detect candidates in each image frame and obtain the 2D bounding box B of each identified candidate. if ;

[0011] S4: B if Convert to inscribed ellipse And represent each ellipse in homogeneous coordinate form

[0012] S5: In three-dimensional space, the homogeneous quadratic form of the quadratic equation is used to represent the ellipsoidal surface in three-dimensional space:

[0013] z T Q i z=0 (1)

[0014] Where z is the symmetric matrix Q i The homogeneous vector of a general three-dimensional point on the ellipsoid surface defined by Q i is the symmetric matrix defining the ellipsoid;

[0015] S6: Construct the projection matrix P f , the projection matrix contains the camera's intrinsic parameters K f , rotation matrix R f and the translation vector t f :

[0016] P f =K f [R f |t f ] (2)

[0017] S7: Combine 3D Euclidean space and Q in 2D image i and Reconstruct into dual space for processing. In dual space, the ellipse in 2D image can be represented by the envelope of all straight lines tangent to it, and the ellipsoid in 3D Euclidean space can be represented by the envelope of all planes tangent to it.

[0018] S8: From a set of dual ellipses Recover the dual ellipsoid The formula needs to be rearranged into a linear system;

[0019] S9: When reconstructing the dual ellipsoid, at least three examination room image frames f=1, 2, ..., F, F ≥ 3 are required from different perspectives. By stacking the linear equations of these image frames, we obtain a maximum linear system, which is expressed as:

[0020] M i w i =0 6F (3)

[0021] S10: solve the linear system;

[0022] S11: Restore the vectorized ellipsoid to its dual ellipsoid. The specific steps are as follows:

[0023] S111: Solved The first 10 elements of are the vectorized elements of the estimated dual ellipsoid, expressed as

[0024] S112: Restore to the dual ellipsoid:

[0025]

[0026] S12: After the dual ellipsoid is initially estimated, the translation vector T i Further move these ellipsoids to the center of the 3D space to obtain a new projection matrix

[0027] S13: Construct a new matrix related to the central ellipsoid Repeat the process from step S10 to step S11 to restore the new dual ellipsoidal surface

[0028] S14: Transform the vector matrix T i Apply to Get the final dual ellipse The expression is as follows:

[0029]

[0030] Where T i T is for T i The transposed matrix of

[0031] S15: The ellipsoid in the original 3D space is restored from the dual space through the inverse adjoint operation of the dual matrix to obtain the three-dimensional position information of each candidate. The expression is as follows:

[0032]

[0033] Preferably, the specific calibration steps in step S1 are as follows:

[0034] S101: Use the chessboard as the calibration board;

[0035] S102: taking multiple pictures of a chessboard from different angles for each camera that needs to be calibrated;

[0036] S103: Automatically detect the corner points of the chessboard in each image using OpenCV;

[0037] S104: Calculate the intrinsic parameter K of each camera using the detected chessboard corner points and the known physical size of the chessboard f , rotation matrix R f and the translation vector t f .

[0038] Preferably, the specific steps of step S4 are as follows:

[0039] S41: The inscribed ellipse is centered at the center of the bounding box, aligned with the image axes, and the axis length is equal to the width and height of the bounding box;

[0040] S42: Express each ellipse using a homogeneous quadratic form of a quadratic equation:

[0041]

[0042] Where u is a symmetric matrix Homogeneous vectors of general two-dimensional points defining the ellipse.

[0043] Preferably, the specific processing steps in step S7 are as follows:

[0044] S71: A 2D ellipse in the dual space can be represented by the envelope of all lines tangent to it, represented by the matrix Definition, where adj is the adjoint operator;

[0045] S72: Apply the homogeneous transformation matrix H to the dual ellipse if , and obtain the standardized dual ellipse

[0046] S73: H if The inverse transformation is And combined with the previous projection matrix P f Reconstruct the new projection matrix

[0047] S74: The 3D ellipsoid in the dual space can be represented by the envelope of all planes tangent to it, represented by the matrix definition;

[0048] S75: Introducing the scaling factor β if For adjustment The scale of the dual ellipsoid is matched with the projection result of the dual ellipsoid. and dual ellipsoid The relationship can be expressed as:

[0049]

[0050] in and They are P f and H if The transposed matrix of .

[0051] Preferably, the specific arrangement steps in step S8 are as follows:

[0052] S81: Vectorizing symmetric matrices, definition definition f v is the vectorized representation of a symmetric matrix;

[0053] S82: Construct vectorized matrix G f , with a dimension of 6×10, and The element-wise products of are arranged in a matrix:

[0054]

[0055] in is the Kronecker product, the matrix D is used to convert the ordinary matrix vectorization form into the symmetric matrix vectorization form, and the matrix E is used to convert the symmetric matrix vectorization form into the ordinary matrix vectorization form;

[0056] S83: Through the vectorized matrix G fand symmetric matrix, formula (8) can be reformulated as a linear system:

[0057]

[0058] Preferably, the specific stacking steps in step S9 are as follows:

[0059] S91: Construct the linear system matrix M i , with dimensions of 6F×(10+F), is defined as follows:

[0060]

[0061] S92: Build w i , w i It is a vectorized representation of the dual ellipsoid and the scaling factor β if The set of , with dimension 10+F×1, is defined as follows:

[0062]

[0063] Preferably, the specific solution steps of step S10 are as follows:

[0064] S1001: In order to avoid the influence of the inaccuracy of the ellipse on the fitting of the ellipsoid, the linear system matrix obtained is Use the minimum error method to solve the system, and the calculation formula is as follows:

[0065]

[0066] The constraints are To avoid trivial zero solutions;

[0067] S1002: The minimization problem in the formula can be solved by The matrix is ​​solved using singular value decomposition, taking the right singular vector associated with the smallest singular value.

[0068] Preferably, the 3D Euclidean space in step S7 and the Q in the 2D image are i and Reconstruct into the dual space for processing. In the dual space, the ellipse in the 2D image can be represented by the envelope of all straight lines tangent to it, and the ellipsoid in the 3D Euclidean space can be represented by the envelope of all planes tangent to it. The specific steps are as follows:

[0069] S71: The ellipse in the 2D image in the dual space can be represented by the envelope of all the straight lines tangent to it, represented by the matrix Definition, where adj is the adjoint operator;

[0070] S72: Apply the homogeneous transformation matrix H to the dual ellipse if , and obtain the standardized dual ellipse

[0071] S73: Get the new projection matrix

[0072] S74: The ellipsoid in the 3D Euclidean space in the dual space can be represented by the envelope of all planes tangent to it, represented by the matrix definition;

[0073] S75: Introducing the scaling factor β if For adjustment , so that it matches the projection result of the dual ellipsoid.

[0074] Preferably, because when processing an ellipse in the dual space, the displacement of the ellipse and the mixing of the size parameters, as well as the geometric differences between the viewing angles, lead to the inaccuracy of the calculated ellipsoid, and this problem is particularly obvious when the ellipse is far away from the center of the image, the ellipse in step S72 is standardized, and the specific steps are as follows:

[0075] S721: Obtain the center coordinates of each inscribed ellipse (t c1 ,t c2 ) and semi-axis length l 1 ,l 2 ;

[0076] S722: For each ellipse, calculate a scaling factor h to normalize the size of the ellipse.

[0077] S723: Constructing a homogeneous transformation matrix H if Move the ellipse from its original position to the center of the image and normalize its size. The matrix form is:

[0078]

[0079] S724: Apply transformation matrix H if and its transpose To the dual ellipse To achieve the standardization of ellipses, is the normalized dual ellipse;

[0080] S725: Repeat steps 721 to 724 in all views to ensure that the ellipse in each view is subjected to the same normalization process.

[0081] Preferably, in order to alleviate the problem of inaccurate shape of the reconstructed ellipsoid due to the center of the ellipsoid being far away from the origin, the ellipsoid initially obtained in step S12 is translated to the 3D origin, and the specific steps are as follows:

[0082] S121: Since the exact ellipsoid is unknown, it is assumed that the translation parameters approximate the center position of the true ellipsoid (x c ,y c ,z c );

[0083] S122: Calculate the translation vector T from the center position of the dual ellipsoid to the 3D origin. c ,-y c ,-z c );

[0084] S123: Construct translation matrix T i :

[0085]

[0086] S124: Apply the translation matrix to each point of the ellipsoid surface obtained by preliminary estimation, and calculate the ellipsoid surface under the updated 3D space center.

[0087] Compared with the prior art, the present invention has the following advantages:

[0088] 1. The 3D candidate positioning method based on multiple fields of view in the examination room provided by the present invention utilizes the candidate's two-dimensional bounding box from multiple perspectives for three-dimensional positioning, and does not rely on traditional depth information or other advanced sensor inputs.

[0089] 2. The 3D candidate positioning method based on multiple fields of view in the examination room provided by the present invention transforms the three-dimensional candidate positioning problem into an optimization problem of a two-dimensional ellipse to a three-dimensional ellipsoid, and converts it into the dual space, providing a closed-form solution in the dual space.

[0090] 3. The 3D candidate positioning method based on multiple fields of view in the examination room provided by the present invention introduces data preprocessing technology to improve the robustness to inaccurate bounding boxes and the accuracy of three-dimensional reconstruction of the ellipsoid surface, so that the calculation of the three-dimensional position of the candidate is more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0092] Figure 1 This is a flow chart of a 3D examinee positioning method based on multiple fields of view in an examination room according to the present invention;

[0093] Figures 2a to 2c A schematic diagram of the bounding box and the corresponding inscribed ellipse of each group of examinees in Example 1 of the present invention;

[0094] Figure 3 It is a schematic diagram of the ellipsoidal surface fitted corresponding to each examinee in the original 3D space in Example 1 of the present invention. DETAILED DESCRIPTION

[0095] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0096] Example 1

[0097] This embodiment is a specific application of a 3D candidate positioning method based on multiple fields of view in a simulated examination room. Three examination room image frames under different viewing angles are taken. There are three candidates in the examination room. The YOLOv8 algorithm is used to detect the position and size of the candidate's boundary box in each picture, and converted into an inscribed ellipse. The three-dimensional ellipsoid surface is estimated by combining the matrix information between the three cameras, and finally the 3D position information of the candidate is obtained. In addition, data preprocessing technology is introduced to perform inscribed ellipse standardization and ellipsoid centering processing, which improves the accuracy of the candidate's three-dimensional position calculation. The specific steps are as follows:

[0098] Step 1): Use the chessboard as a calibration board, take multiple pictures of the chessboard from different angles for the three cameras to be calibrated, use OpenCV to automatically detect the corners of the chessboard in each image, and calculate the intrinsic parameter K of each camera based on the detected chessboard corners and the known physical size of the chessboard. f , rotation matrix R f and the translation vector t f .

[0099]

[0100] Step 2): Obtain multi-view image frames f=1, 2, 3, which represent examination room scenes from different viewpoints.

[0101] Step 3): Use the yolov8 algorithm to detect three examinees in three image frames respectively, and obtain the 2D bounding box B of each examinee detected. ifAnd calculate the center point, width and height of each bounding box, respectively: B 11 =[303.37,300.9,69.31,72.43],B 21 =[383.52, 321.45, 76.74, 70.60], B 31 =[243.5, 386.3, 65.63, 94.21], B 12 =[357.5,309.1,76.8,61.6],B 22 =[423.0,358.2,87.5,56.8],B 32 =[258.7,364.30,82.4,84.9],B 13 =[354.8,342.2,79.1,52.1],B 23 =[393.8,416.0,93.1,74.5],B 33 =[319.09,278.32,72.3,70.4], where the first two digits are the center point coordinates and the last two digits are the width and height of the bounding box.

[0102] Step 4): Bounding box B of each detected candidate if Convert to inscribed ellipse The inscribed ellipse is centered on the center of the bounding box, aligned with the image axis, and the axis length is equal to the width and height of the bounding box. Each inscribed ellipse is calculated and expressed in homogeneous coordinates. The inscribed ellipses obtained are: The first two digits are the center of the inscribed ellipse, and the last two digits are the major and minor axes of the inscribed ellipse respectively.

[0103] Step 5): Express each ellipse using the homogeneous quadratic form of the quadratic equation,

[0104]

[0105] Where u is a symmetric matrix Homogeneous vectors of general two-dimensional points that define a quadratic curve.

[0106] Step 6): In three-dimensional space, the ellipsoidal surface in three-dimensional space is represented by the homogeneous quadratic form of the quadratic equation.

[0107] z f Qiz=0 (1)

[0108] where z is a homogeneous vector of a general three-dimensional point on a quadratic surface defined by the symmetric matrix Qi, and Qi is an ellipsoid expressed in homogeneous coordinate form.

[0109] Step 7): Construct 3 transformation matrices P 1 , P 2 , P 3 , the transformation matrix contains the intrinsic parameters K of each camera f , rotation matrix R f and the translation vector t f :

[0110] P f =K f [R f |t f ] (2)

[0111] Step 8): Convert Q in the original space (3D Euclidean space and 2D image) i and Reconstruct into dual space for processing. In dual space, an ellipse in 2D can be represented by the envelope of all straight lines tangent to it, and an ellipsoid in 3D can be represented by the envelope of all planes tangent to it;

[0112] Step 9): For each ellipse, calculate the scaling factor h to normalize the size of the ellipse, Construct the homogeneous transformation matrix H if , moves the ellipse from its original position to the center of the image and normalizes its size, the matrix form is:

[0113]

[0114] Each transformation matrix is ​​calculated as follows:

[0115]

[0116] Get the new projection matrix

[0117] Step 10): Introduce scaling factor β if For adjustment The scale factor β if The role of the projected and transformed ellipse is to ensure that the geometric properties of the original three-dimensional ellipsoid can be accurately reflected. and dual ellipsoid The relationship can be expressed as:

[0118]

[0119] Step 11): From the dual ellipse Recover the dual ellipsoid The formula needs to be rearranged into a linear system, the specific steps are as follows:

[0120] Step 11-1): Vectorize the symmetric matrix and define definition f v is the vectorized representation of a symmetric matrix;

[0121] Step 11-2): Construct the vectorized matrix G 1 , G 2 , G 3 , with a dimension of 6×10, and The element-wise products of are arranged in a matrix:

[0122]

[0123] in is the Kronecker product, the matrix D is used to convert the ordinary matrix vectorization form into the symmetric matrix vectorization form, and the matrix E is used to convert the symmetric matrix vectorization form into the ordinary matrix vectorization form.

[0124] Step 11-3): Through the matrix G f and the vectorized symmetric matrix, formula (3) can be reformulated as a linear system:

[0125]

[0126] The linear system corresponding to Candidate 1 in three perspectives is expressed as:

[0127] From the first perspective:

[0128] From the second perspective:

[0129] From the second perspective:

[0130] The linear system corresponding to Candidate 2 in three perspectives is expressed as:

[0131] From the first perspective:

[0132] From the second perspective:

[0133] From the third perspective:

[0134] The linear system corresponding to Candidate 3 from three perspectives is expressed as:

[0135] From the first perspective:

[0136] From the second perspective:

[0137] From the third perspective:

[0138] Step 12): When reconstructing the dual ellipsoid through the examination room image frames f=1, 2, 3 under three different viewing angles, in order to obtain a unique solution, it is necessary to stack the linear equations of these image frames. We can get a maximum linear system, which is expressed as:

[0139] M i w i =0 6F (3)

[0140] The specific steps are as follows:

[0141] Step 12-1): Construct the linear system matrix M i , the dimension is 6F×(10+F), and the three linear systems are as follows:

[0142]

[0143] Step 12-2): Build w i , w i It is a vectorized representation of the dual ellipsoid and the scaling factor β if A collection of dimensions 10+F, 3 w i as follows:

[0144]

[0145] Step 13): Solve the linear system. The specific steps are as follows:

[0146] Step 13-1): In actual situations, the inscribed ellipse detected by the general yolov8 may have inaccuracies in position and size, which will affect the fitting of the ellipsoid. Therefore, an error is introduced to solve the system using the minimum error method. The calculation formula is as follows:

[0147]

[0148] The constraints are To avoid trivial zero solutions.

[0149] Step 13-2): The minimization problem in the formula can be solved by The matrix is ​​solved using singular value decomposition (SVD) and the right singular vector associated with the smallest singular value is taken.

[0150] Step 14): Restore the dual ellipsoid from the vectorized ellipsoid. The specific steps are as follows:

[0151] Step 14-1): Solve The first 10 elements of are the vectorized elements of the estimated dual ellipsoid, expressed as

[0152] Step 14-2): Restore to the dual ellipsoid:

[0153]

[0154] Step 15): After the dual ellipsoid is preliminarily estimated, the translation vector T from the center position of the preliminarily obtained dual ellipsoid to the 3D origin is calculated. c ,-y c ,-z c ), and construct the translation matrix T i :

[0155]

[0156] Step 16): Apply the translation matrix to each point of the ellipsoid obtained by the preliminary estimation to obtain the ellipsoid under the updated 3D space center;

[0157] Step 17): Get the new projection matrix

[0158] Step 18): Construct a new matrix associated with the central ellipsoid Repeat steps 12) to 13) to restore the new dual ellipsoidal surface.

[0159] Step 19): Convert the vector matrix T i Apply to Get the final dual ellipse The expression is as follows:

[0160]

[0161] Step 20): Recover the ellipsoid in the original 3D space from the dual space through the inverse adjoint operation of the dual matrix Get the three-dimensional position information of each candidate.

[0162]

[0163] Through practical verification, this method can achieve accurate 3D positioning of candidates by using only the 2D bounding box of the given candidate and the multi-view geometric relationship. By converting these bounding boxes into inscribed ellipses, combining the known camera matrix parameters and the inscribed ellipses under multiple views, the three-dimensional ellipsoid surface is estimated in the dual space, and finally the 3D position information of the candidate is obtained.

[0164] Obviously, the above embodiments are merely examples for the purpose of clear explanation, and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived therefrom are still within the scope of protection of the invention.

Claims

1. A 3D examinee positioning method based on multiple fields of view in an examination room, characterized in that: The steps include: S1: Calibrate each camera to obtain the camera’s intrinsic parameter K f , rotation matrix R f and the translation vector t f ; S2: Acquire multi-view image frames f=1, 2, ..., F, which represent the examination room scenes under different cameras; S3: Use the YOLOv8 algorithm to detect candidates in each image frame and obtain the 2D bounding box B of each identified candidate. if ; S4: B if Convert to inscribed ellipse And represent each ellipse in homogeneous coordinate form S5: In three-dimensional space, the homogeneous quadratic form of the quadratic equation is used to represent the ellipsoidal surface in three-dimensional space: z T Q i z=0 (1) Where z is the symmetric matrix Q i The homogeneous vector of a general three-dimensional point on the ellipsoid surface defined by Q i is the symmetric matrix defining the ellipsoid; S6: Construct the projection matrix P f , the projection matrix contains the camera's intrinsic parameters K f , rotation matrix R f and the translation vector t f : P f =K f [R f |t f ] (2) S7: Combine 3D Euclidean space and Q in 2D image i and Reconstruct into dual space for processing. In the dual space, the ellipse in the 2D image is represented by the envelope of all straight lines tangent to it, and the ellipsoid in the 3D Euclidean space is represented by the envelope of all planes tangent to it; S8: From a set of dual ellipses Recover the dual ellipsoid Rearrange the formula into a linear system; S9: When reconstructing the dual ellipsoid, at least three examination room image frames from different perspectives are required, f = 1, 2, ..., F, F ≥ 3. By stacking the linear equations of these image frames, a maximum linear system is obtained, which is expressed as: M i w i =0 6F (3) S10: solve the linear system; S11: Restore the vectorized ellipsoid to its dual ellipsoid. The specific steps are as follows: S111: Solved The first 10 elements of are the vectorized elements of the estimated dual ellipsoid, expressed as S112: Restore to the dual ellipsoid: S12: After the dual ellipsoid is initially estimated, the translation vector T i Further move these ellipsoids to the center of the 3D space to obtain a new projection matrix S13: Construct a new matrix related to the central ellipsoid Repeat the process from step S10 to step S11 to restore the new dual ellipsoidal surface S14: Transform the vector matrix T i Apply to Get the final dual ellipse The expression is as follows: Where T i T is for T i The transposed matrix of S15: The ellipsoid in the original 3D space is restored from the dual space through the inverse adjoint operation of the dual matrix to obtain the three-dimensional position information of each candidate. The expression is as follows: The specific processing steps in step S7 are as follows: S71: A 2D ellipse in dual space is represented by the envelope of all lines tangent to it, represented by the matrix Definition, where adj is the adjoint operator; S72: Apply the homogeneous transformation matrix H to the dual ellipse if , and obtain the standardized dual ellipse S73: H if The inverse transformation is And combined with the previous projection matrix P f Reconstruct the new projection matrix S74: The 3D ellipsoid in dual space is represented by the envelope of all planes tangent to it, represented by the matrix definition; S75: Introducing the scaling factor β if For adjustment The scale of the dual ellipsoid is matched with the projection result of the dual ellipsoid. and the dual ellipsoid The relationship is expressed as: in and They are P f and H if The transposed matrix of .

2. The 3D examinee positioning method based on multiple fields of view in the examination room according to claim 1 is characterized in that: The specific calibration steps in step S1 are as follows: S101: Use the chessboard as the calibration board; S102: taking multiple pictures of a chessboard from different angles for each camera that needs to be calibrated; S103: Automatically detect the corner points of the chessboard in each image using OpenCV; S104: Calculate the intrinsic parameter K of each camera using the detected chessboard corner points and the known physical size of the chessboard f , rotation matrix R f and the translation vector t f .

3. The 3D examinee positioning method based on multiple fields of view in the examination room according to claim 1 is characterized in that: The specific steps of step S4 are as follows: S41: The inscribed ellipse is centered at the center of the bounding box, aligned with the image axes, and the axis length is equal to the width and height of the bounding box; S42: Express each ellipse using a homogeneous quadratic form of a quadratic equation: Where u is a symmetric matrix Homogeneous vectors of general two-dimensional points defining the ellipse.

4. The 3D examinee positioning method based on multiple fields of view in the examination room according to claim 1, characterized in that: The specific arrangement steps in step S8 are as follows: S81: Vectorizing symmetric matrices, definition definition f v is the vectorized representation of a symmetric matrix; S82: Construct vectorized matrix G f , with a dimension of 6×10, and The element-wise products of are arranged in a matrix: in is the Kronecker product, the matrix D is used to convert the ordinary matrix vectorization form into the symmetric matrix vectorization form, and the matrix E is used to convert the symmetric matrix vectorization form into the ordinary matrix vectorization form; S83: Through the vectorized matrix G f and symmetric matrix, formula (8) can be reformulated as a linear system:

5. The 3D examinee positioning method based on multiple fields of view in an examination room according to claim 1, characterized in that: The specific stacking steps in step S9 are as follows: S91: Construct the linear system matrix M i , with dimensions of 6F×(10+F), is defined as follows: S92: Build w i , w i It is a vectorized representation of the dual ellipsoid and the scaling factor β if The set of , with dimension 10+F×1, is defined as follows:

6. The 3D examinee positioning method based on multiple fields of view in an examination room according to claim 1, characterized in that: The specific solution steps of step S10 are as follows: S1001: For the actual linear system matrix Use the minimum error method to solve the system, and the calculation formula is as follows: The constraints are S1002: The minimization problem in the formula is solved by The matrix is ​​solved using singular value decomposition, taking the right singular vector associated with the smallest singular value.

7. The 3D examinee positioning method based on multiple fields of view in an examination room according to claim 1, characterized in that: The ellipse in step S72 is standardized, and the specific steps are as follows: S721: Obtain the center coordinates of each inscribed ellipse (t c1 ,t c2 ) and semi-axis lengths l1, l2; S722: For each ellipse, calculate a scaling factor h to normalize the size of the ellipse. S723: Constructing the homogeneous transformation matrix H if Move the ellipse from its original position to the center of the image and normalize its size. The matrix form is: S724: Apply transformation matrix H if and its transpose To the dual ellipse To achieve the standardization of ellipses, is the normalized dual ellipse; S725: Repeat steps 721 to 724 in all views to ensure that the ellipse in each view is subjected to the same normalization process.

8. The 3D examinee positioning method based on multiple fields of view in an examination room according to claim 1, characterized in that: The ellipsoid surface initially obtained in step S12 is translated to the 3D origin. The specific steps are as follows: S121: Since the exact ellipsoid is unknown, it is assumed that the translation parameters approximate the center position of the true ellipsoid (x c ,y c ,z c ); S122: Calculate the translation vector T from the center position of the dual ellipsoid to the 3D origin. c ,-y c ,-z c ); S123: Construct translation matrix T i : S124: Apply the translation matrix to each point of the ellipsoid surface obtained by preliminary estimation, and calculate the ellipsoid surface under the updated 3D space center.

Images