JP2018137667A - Camera calibration method, program and device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To accurately calculate a homography matrix from a camera image without requiring complete model matching while using a number of intersections.SOLUTION: A feature vector element generation part 52 generates a feature vector element for each detected intersection. The presence/absence of a field line extending in each direction for each of intersections is allocated to each of bits of the feature vector element consisting of binary data of four bits, and the presence/absence in a "UP" direction is allocated to a most significant bit (MSB1). The presence/absence in a "RIGHT" direction is allocated to an MSB2, the presence/absence in a "DOWN" direction is allocated to an MSB3, and the presence/absence in a "LEFT" direction is allocated to a least significant bit (MSB4). If a field line is present, "1" is set to each of bits and if a field line is absent, "0" is set. A feature vector element at an intersection where a vertical field line in the "UP" direction and a vertical field line in the LEFT" direction are detected becomes "1001".SELECTED DRAWING: Figure 7

Description

  The present invention relates to a camera calibration method and apparatus, and more particularly to a method, program, and apparatus for performing camera calibration using a camera image including a field line, such as baseball, soccer, or tennis.

  Camera calibration is a very important technology in 3D computer vision, and the relationship between the world coordinate system of the real scene and the pixel coordinate system of the camera image depends on the internal and external parameters (or homography matrix) specific to the camera. Based on.

  In particular, in navigation from a free viewpoint in sports videos, the position of each player on the actual field is important, and can be obtained by projective transformation of the coordinates of each player in the camera image using a homography matrix. Therefore, a technology capable of automatically performing accurate camera calibration is required for real-time applications such as free viewpoint navigation and video analysis.

  Non-Patent Document 1 discloses an automatic camera calibration method based on a field model for free viewpoint navigation in sports events. In Non-Patent Document 1, a robust field line is extracted using a histogram of line segments in a camera image, four intersections of two horizontal lines and two vertical lines are extracted, and a homography matrix is extracted. Estimated.

  In Patent Document 1, a field line image is extracted from a camera image, a line segment is extracted from the field line image to recognize a field line, and each field line is classified into a vertical and horizontal field line group. A technique is disclosed in which each field line is evaluated, a corresponding point is set based on the intersection of the highly evaluated field lines, and a homography matrix is determined based on the set corresponding point.

JP 2016-220129 A

Qiang Yao, Hiroshi Sankoh, Keisuke Nonaka, Sei Naito, "Automatic Camera Self-Calibration for Immersive Navigation of Free Viewpoint Sports Video", Multimedia Signal Processing (MMSP), 2016 IEEE International Workshop on. Montreal, Canada, IEEE, 2016. 09 .

  In Non-Patent Document 1, only four intersections can be extracted from a camera image, so complete model matching is required. Not only does matching require a long time, but homogeneity is not required unless the four intersections cover a wide area of the camera image. There has been a technical problem that the accuracy of the graphic matrix decreases.

  In Patent Document 1, perfect model matching is necessary to calculate a correct homography matrix, and there is a technical problem in that the accuracy of the homography matrix decreases if there is an intersection that cannot be recognized from the camera image.

  An object of the present invention is to solve the above technical problem and to provide a camera calibration method, program, and apparatus capable of accurately calculating a homography matrix from a camera image without using perfect model matching while using a large number of intersections. It is to provide.

  In order to achieve the above object, the present invention is characterized in that an apparatus for performing camera calibration based on a homography matrix between a field model and a camera image has the following configuration.

  (1) means for extracting a field line from a camera image, means for calculating an intersection of field lines, means for generating a camera intersection feature vector element based on the number and direction of field lines passing through the intersection for each intersection Means for arranging camera intersection feature vector elements in a predetermined order to generate a camera intersection feature vector Vc; means for generating a model intersection feature vector element for each intersection of each field line of a field model; and camera intersection feature Each model intersection feature vector element is arranged in the predetermined order for each combination of the model intersection feature vector elements having the intersection arrangement similar to each camera intersection feature vector element of the vector Vc to generate a plurality of model intersection feature vectors Vm The camera intersection feature vector Vc and each model intersection feature vector Vm The model intersection feature vector Vm with high similarity is extracted, and the corresponding points are set based on the correspondence between the model intersection feature vector Vm with high similarity and the camera intersection feature vector Vc, and the homography matrix is calculated. Means.

  (2) The means for extracting the model intersection feature vector Vm having a high similarity is a means for extracting the model intersection feature vector Vm of the top N best model having a high similarity and evaluating the model intersection feature vector Vm of the N best. And the means for evaluating includes means for projecting the left and right boundary lines of the frame image onto a field model with each homography matrix, and obtaining each corresponding line segment, and using the intersection of each corresponding line segment as a camera point Means for registering, and means for calculating a displacement between a camera point registered based on the nth frame image and a camera point registered based on the mth frame image, and the displacement exhibits a minimum value The highest evaluation of the homography matrix was made.

  (3) The feature vector element is binary data whose bit length is the number of field lines extending from each intersection, and each bit value is binary depending on whether or not a field line extending in each direction is extracted. Was set to one of the following.

  (4) The means for extracting the model intersection feature vector Vm having a high degree of similarity arranges the feature vector elements of each intersection in the order in which the respective intersections are raster-scanned to form the camera intersection feature vector Vc.

  (5) The means for extracting the model intersection feature vector Vm having a high degree of similarity includes the camera intersection feature vector Vc if the intersection of the field lines extracted from the camera image includes an intersection that cannot be seen on the camera image. The feature vector element of the corresponding intersection within the matching range of the model intersection feature vector Vm is copied to the feature vector element of the intersection of the invisible.

  According to the present invention, the following effects are achieved.

  (1) The number of corresponding points for calculating a homography matrix is not limited to four points, and perfect matching at all corresponding points is not required, so that the homography matrix can be calculated accurately in a short time. Become.

  (2) Since binary data (feature vector elements) based on the direction and number of lines extending from the intersection is used as the feature quantity at each intersection, the data size of the feature quantity at each intersection is reduced and the matching process is simplified. As a result, the processing speed is improved.

  (3) The camera intersection feature vector Vc is generated by arranging the feature values (feature vector elements) of each intersection extracted from the camera image in a one-dimensional manner in a predetermined rule, for example, in the order of raster scanning. The model intersection feature vector Vm is generated by extracting only the combination of each intersection point of the field model whose array is similar to that of the field model and arranging the feature values of each intersection point in one dimension according to the same rule for each combination. The number of model intersection feature vectors is reduced, enabling high-speed matching processing.

  (4) When an invisible intersection is detected in the camera image, the feature quantity (feature vector element) is interpolated with the feature quantity of the corresponding intersection in the field model. Reduction is prevented.

It is the functional block diagram which showed the structure of the principal part of the camera calibration apparatus which concerns on 1st Embodiment of this invention. It is the figure which showed an example of the camera image I _RGB . 5 is a diagram showing an example of a field ground image G. FIG. 5 is a diagram showing an example of a field line image F. FIG. It is the figure which showed an example of the look-up table (LUT). It is the figure which showed the direction of the field line extended from each intersection. It is the figure which showed the setting method of the feature vector element. It is the figure which showed the example of a setting of the feature vector element. It is the figure which showed the production | generation method of the camera intersection feature vector. It is the figure which showed the matching method of feature vector Vc, Vm. It is a figure explaining the method of interpolating the feature vector element of an invisible intersection. It is the figure which showed the example of the intersection which cannot be seen. It is the functional block diagram which showed the structure of the principal part of the camera calibration apparatus which concerns on 2nd Embodiment of this invention. It is a figure for demonstrating the evaluation method of a homography matrix. It is the flowchart which showed the evaluation procedure of the homography matrix. It is the figure which showed the method of calibrating the homography matrix of two cameras cam1 and cam2.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 is a functional block diagram showing the configuration of the main part of a camera calibration apparatus according to an embodiment of the present invention, and here, the configuration unnecessary for the description of the present invention is omitted.

The field ground image extraction unit 10 acquires a field ground image G by removing background regions other than the field ground region from the camera image I _RGB in which the field ground is captured as shown in FIG.

In this embodiment, first, the camera image I _RGB is converted from the RBG color space to the image I _HSV in the HSV color space. Next, several statistical thresholds σ _min ^H , σ _max ^H and σ _min ^V are estimated, and a field ground is extracted using a labeling process of the following equation (1). In the present embodiment, the field ground image G shown in FIG. 3 is extracted from the labeled mask image B.

  Based on the field ground image G, the field line image extraction unit 20 thresholds the pixel intensity (grayscale) and the pixel gradient according to the following equation (2), so that the field line image F shown in FIG. To extract.

Here, the field ground image G is first converted into a gray image. The symbols th _w and th _grad are white and pixel gradient thresholds, respectively. The symbol Δ is expressed by the following equation (3). τ is the distance used to calculate the gradient of each pixel G (x, y) of the field ground image G.

  In the field line analysis unit 30, the segment detection unit 31 applies a well-known Hough transform to the field line image F to detect a segment (line segment) that is a component of the field line. Each segment is defined by the start point coordinate P1 (x1, y1) and the end point coordinate P2 (x2, y2) and registered in the segment list.

  The segment parameter calculation unit 32 calculates two parameters of the detected slope k and offset (intercept) b of each segment by the following equations (4) and (5), respectively. These two parameters are used by the line pruning unit 33, the line classification unit 34, and the sorting unit 35 described later.

  Since the segment group detected by the Hough transform includes segments constituting the same field line, the line pruning unit 33 performs line pruning to remove redundant lines. At this time, the slope k and the offset b are used to compare two line segments, and if the following equations (6) and (7) are both established, the two segments li and lj constitute the same field line. As a result, the segment lj is removed from the segment list, and each field line is determined based on the segments remaining in the segment list.

  Here, ki and kj are the slopes of the segments li and lj, respectively, and bi and bj are the offsets of the segments li and lj, respectively. The threshold Δα is a threshold for determining the identity of the slopes ki and kj, and the threshold Δβ is a threshold for determining the identity of the offsets bi and bj.

Line classification unit 34, to calculate the intersection point from the determined field lines, each field line is classified into one of the horizontal field lines L _H and vertical field lines L _V based on the inclination k.

In the present embodiment, it has defined a line classification threshold gamma advance, | ki |> classified gamma if the vertical field lines L _V, | ki | ≦ γ if classified into horizontal field lines L _H. In addition, the line classification threshold value may differ depending on the position of the camera and the shooting direction. Therefore, after the line classification process, the horizontal field line L _H and the vertical field line L _V are in two subsets.

Sorting unit 35 sorts from above horizontal field lines L _H below, sort the vertical field lines L _V from left to right. Using an offset bi is when sorting horizontal field lines L _H, sorted from top to bottom the corresponding line. In the sorting of the vertical field lines L _V , paying attention to the top horizontal field line L _H , the intersection of the horizontal field line L _H and each vertical field line L _V is calculated, and each intersection is determined according to the x coordinate value of the intersection. to sort _the vertical field line L _V.

  The inclination threshold value Δα, the offset threshold value Δβ, and the line classification threshold value γ depend on the position and orientation of the camera. Therefore, in this embodiment, as shown in FIG. 5, a look-up table (LUT) 100 in which a threshold group unique to each camera position is registered is built in advance, and a threshold group corresponding to the camera for each camera. Is adopted.

Returning to FIG. 1, the intersection detection unit 40 detects an intersection between the vertical field line L _V and the horizontal field line L _H by a known method, and calculates the coordinates thereof.

  The vector information generation unit 50 generates a feature vector element as a detected feature amount of the intersection. The line direction analysis unit 51 detects the presence or absence of a field line extending from the intersection for each detected intersection in order to generate a feature vector element based on the number and direction of field lines extending from the intersection for each intersection.

In this embodiment, field lines are searched for four directions orthogonal to each intersection P (x, y). As shown in FIG. 6, each direction along the vertical field line Lv is “UP (upper ) ”,“ DOWN (down) ”, and the respective directions along the horizontal field line L _H are defined as“ RIGHT (right) ”and“ LEFT (left) ”.

In the present embodiment, in order to maintain robustness, an accumulator that integrates pixel values in each direction is defined. d _i ^U is an accumulator in the “UP” direction, and each time a white pixel is found along the “UP” direction of the intersection P (x, y), “1” is added to the corresponding accumulator d _i ^U. The Furthermore, the cumulative range is defined as N (N is a reasonable search range). Therefore, the cumulative number represented by the following equation (8) takes a value between 0 and N.

In the present embodiment, presence / absence threshold value μ (0 <μ <1) is defined. If d _i ^U > μN, the presence of the field line in the “UP” direction is recognized, and if d _i ^U ≦ μN. The presence of the field line is denied. Further, in the present embodiment, another threshold value μ _rej is defined as a high-speed algorithm, and if the following equation (9) is continuously established, the presence of a field line in the direction is denied.

Such analysis, the intersection P (x, y) as the starting point, is performed for each direction based on the inclination of the field lines, the UP direction accumulator d _i ^U, the DOWN direction accumulator d _i ^D, RIGHT direction of the accumulator analysis of d _i ^R, LEFT direction of the accumulator d _i ^L is respectively performed based on the following equation (10) to (13).

  Each expression is “1” if a white pixel (pixel location = 255) is found in each direction, and “0” if not found.

  The feature vector element generation unit 52 generates a feature vector element for each detected intersection point based on the presence / absence of a field line extending from the intersection point in each direction.

  In the present embodiment, as shown in FIG. 7, the presence / absence of a field line extending in each direction at each intersection is assigned to each bit of a feature vector element composed of 4-bit binary data. For example, “UP” The presence or absence in the direction is assigned to the most significant bit (MSB1).

  Similarly, presence / absence in the “RIGHT” direction is assigned to MSB2, presence / absence in the “DOWN” direction is assigned to MSB3, and presence / absence in the “LEFT” direction is assigned to the least significant bit (MSB4). For each bit, “1” is set if the field line exists, and “0” is set if it does not exist. Therefore, as shown in FIGS. 7 and 8, the feature vector element of the intersection where the vertical field line extending in the “UP” direction and the vertical field line extending in the “LEFT” direction is detected is “1001”.

  The camera intersection feature vector generation unit 53 generates the camera intersection feature vector Vc by arranging the feature vector elements at each intersection in a predetermined order, for example, the raster scan order as shown in FIG. Therefore, in the example of FIG. 9, the camera intersection feature vector Vc is expressed by the following equation (14).

  Returning to FIG. 1, in the model intersection feature vector element storage unit 60, a feature generated in advance by the same procedure as described above based on the number and direction of the field lines passing through the intersection for each field line intersection of the field model. Vector elements are stored.

In the corresponding point setting unit 70, the matching candidate extraction unit 71 determines whether the camera intersection point is in accordance with the number of horizontal field lines L _H and vertical field lines L _V that define the intersection points related to the feature vector elements of the camera intersection feature vector Vc. A combination of model intersection feature vector elements having a similar intersection array with each feature vector element of the feature vector Vc is extracted. Then, each feature vector element is arranged for each combination of the feature vector elements to generate a plurality of model intersection feature vectors Vm.

  The matching unit 72 performs matching between the camera intersection feature vector Vc and each model intersection feature vector Vm, and responds based on the correspondence relationship between the model intersection feature vector Vm having the highest similarity and the camera intersection feature vector Vc. Set a point.

  FIG. 10 is a diagram schematically showing a matching method between the camera intersection feature vector Vc and the model intersection feature vector Vm.

If the camera intersection feature vector Vc is composed of, for example, six feature vector elements related to the intersection of three horizontal field lines L _H and two vertical field lines L _V in the raster scan order, a matching candidate is extracted. The unit 71 relates from the Vm element storage unit 60 to the intersection of the three horizontal field lines and the two vertical field lines of the field model, that is, the intersection of the camera image intersection and the arrangement (horizontal and vertical positions). All sets of six feature vector elements are extracted and connected to each other in raster scan order to generate a model intersection feature vector Vm.

That is, as shown in FIG. 9, when the field model is composed of six horizontal field lines L _H and seven vertical field lines L _V , the field model is arbitrarily selected from the six horizontal field lines L _H. the aforementioned three (the ways ₆ C ₃ = 20) was a combination of six of the feature vector element for each intersection of the two selected arbitrarily from seven vertical field lines L _{V (7} C ₂ = 21 ways) Acquired from the Vm element storage unit 60 and connects six feature vector elements for each combination to generate 420 model intersection feature vectors Vm.

  The matching unit 72 obtains the Hamming distance D between the camera intersection feature vector Vc and each model intersection feature vector Vm. When the hamming distance D is obtained for all the combinations, the corresponding intersection of the model intersection feature vector Vm and the camera intersection feature vector Vc that gives the minimum hamming distance D is registered as a corresponding point.

  As shown in FIG. 12, if the detected intersection P includes an invisible intersection that is outside the range of the camera image, the corresponding intersection portion of the camera intersection feature vector Vc has a feature. Vector elements are not generated.

  In the present embodiment, assuming such a case, if the camera intersection feature vector Vc includes an invisible intersection (represented by a symbol?) And the 4-bit feature vector element v is not set, the model intersection feature The camera intersection feature vector Vc is interpolated by copying the 4-bit value of the corresponding intersection in the vector Vm. Then, the Hamming distance D is calculated based on the camera intersection feature vector Vc and the model intersection feature vector Vm after interpolation.

  Returning to FIG. 1, the homography matrix determination unit 80 determines a homography matrix based on at least four corresponding points. The camera parameter calculation unit 90 calculates a camera calibration matrix based on the determined homography matrix and camera internal parameters.

  FIG. 13 is a functional block diagram showing the configuration of the second embodiment of the present invention, and the same reference numerals as those described above represent the same or equivalent parts. The present embodiment is characterized in that a homography matrix evaluation unit 110 that evaluates a plurality of homography matrices and selects any one of the homography matrices is provided.

  In the first embodiment, only one maximum likelihood model intersection feature vector Vm is obtained based on the matching result between the camera intersection feature vector Vc and the plurality of model intersection feature vectors Vm, and the homography is obtained based on the corresponding points. In the present embodiment, the corresponding point setting unit 70 extracts the top N best model intersection feature vector Vm having a small Hamming distance D.

  The homography matrix determination unit 80 generates N homography matrices by calculating a homography matrix for each corresponding point set related to the top N best model intersection feature vector Vm. The most likely mammography matrix is selected from the homography matrices.

Back projection is known as an effective method for evaluating and verifying a homography matrix. However, in the case of a camera that is not particularly fixed (pan / tilt / zoom), an appropriate time is required for evaluation. Therefore, in this embodiment, a new verification method that is much simpler and faster than the back projection method is newly adopted. FIG. 14 is a diagram showing a homography matrix evaluation method performed by the homography matrix evaluation unit 110. It is.

  In this embodiment, first, a camera point (M, N) is estimated by applying a homography matrix to the nth frame image. Next, a camera point (M ′, N ′) is estimated by applying a homography matrix to the m-th frame image. Then, the homography matrix is evaluated by comparing the camera points (M, N) and (M ′, N ′).

FIG. 15 is a flowchart showing a homography matrix evaluation procedure. In step S1, two boundary lines are selected on the nth frame image. Here, description will be made assuming that the left boundary line W _1L and the right boundary line W _1R are selected.

In step S2, the boundary lines W _1L and W _1R are projected onto the field model using the extracted homography matrix. In step S3, the intersection coordinates of the two projected line segments W _1l and W _1r are obtained and registered as reference camera points (M, N).

In step S4, the same processing as described above is executed for the two boundary lines W _2L and the right boundary line W _2R selected on the mth frame image, and camera points (M ′, N ′) are registered. The In step S5, the displacement θ of each camera point (M, N), (M ′, N ′) is calculated based on the following equation.

  In step S6, the displacements θ calculated using the N best homography matrices are compared, and the homography matrix that gives the minimum displacement θ is selected as the maximum likelihood homography matrix.

According to this embodiment, as a whole, an unfixed camera can be evaluated automatically, robustly and quickly by using the method described above.
The maximum likelihood homography matrix calculated as described above can also be used to calibrate two cameras cam1 and cam2 that are far away from each other, as shown in FIG. Usually, calibration between two cameras is achieved by using feature point detection and matching. However, if the baseline between the two cameras cam1 and cam2 is large, the calibration accuracy may be low, or calibration may not be possible.

Therefore, the present invention provides a bridge that links cam1 and cam2 as a field model-based calibration method. Specifically, the homography matrices H ⁽¹⁾ and H ⁽²⁾ related to the field model are calculated for each camera cam1 and cam2, and the homography matrix between the cameras Cam1 and Cam2 is calculated by the following equation (16).

In each embodiment described above, extracts the horizontal field lines L _H and vertical field lines L _V by extracting field lines from pre-field image, the intersection between the horizontal field lines L _H and the vertical field lines L _V However, the present invention is not limited to this. The feature points of the field line image and the field model are detected using a well-known feature detection method such as SIFT, and the detected feature points are detected as field lines. It may be considered as an intersection of.

  DESCRIPTION OF SYMBOLS 10 ... Field ground image extraction part, 20 ... Field line image extraction part, 30 ... Field line analysis part, 31 ... Segment detection part, 32 ... Segment parameter calculation part, 33 ... Line pruning part, 34 ... Line classification part, 35 ... Sorting unit, 40 ... intersection detection unit, 50 ... vector information generation unit, 51 ... line direction analysis unit, 52 ... feature vector element generation unit, 53 ... camera intersection point feature vector (Vc) generation unit, 60 ... Vm element storage unit, DESCRIPTION OF SYMBOLS 70 ... Corresponding point setting part, 71 ... Matching candidate extraction part, 72 ... Matching part, 80 ... Homography matrix determination part, 90 ... Camera parameter calculation part, 100 ... Look-up table (LUT), 110 ... Homography matrix evaluation part

Claims (14)

In a device that performs camera calibration based on a homography matrix between a field model and a camera image,
Means for extracting a field line from a camera image;
Means for calculating the intersection of the field lines;
Means for generating a camera intersection feature vector element based on the number and orientation of field lines passing through the intersection for each intersection;
Means for arranging each camera intersection feature vector element in a predetermined order to generate a camera intersection feature vector Vc;
Means for generating a model intersection feature vector element for each intersection of each field line of the field model;
Each model intersection feature vector element is arranged in the predetermined order for each combination of model intersection feature vector elements whose intersections are similar to each camera intersection feature vector element of the camera intersection feature vector Vc, and a plurality of model intersection feature vectors Means for generating Vm;
Means for comparing the camera intersection feature vector Vc and each model intersection feature vector Vm to extract a model intersection feature vector Vm having a high degree of similarity;
Means for setting corresponding points based on the corresponding relationship between the model intersection feature vector Vm having a high similarity and the camera intersection feature vector Vc;
And a means for calculating a homography matrix based on the corresponding points. The means for generating the model intersection feature vector Vm extracts the top N best model intersection feature vector Vm having a high similarity,
Means for evaluating the N-best model intersection feature vector Vm;
The camera calibration apparatus according to claim 1, wherein a homography matrix is calculated based on corresponding points in the model intersection feature vector Vm having the highest evaluation. The means for evaluating is
Means for determining the corresponding line segment by projecting the left and right boundary lines of the frame image onto the field model with each homography matrix;
Means for registering the intersection of each corresponding line segment as a camera point;
means for calculating a displacement between a camera point registered based on the nth frame image and a camera point registered based on the mth frame image;
The camera calibration apparatus according to claim 2, wherein evaluation of a homography matrix in which the displacement has a minimum value is made highest.   The means for generating the model intersection feature vector Vm is a model based on a model intersection feature vector element calculated for each combination of field lines on the field model corresponding to each field line for calculating each intersection of the camera image. 4. The camera calibration apparatus according to claim 1, wherein an intersection feature vector Vm is generated.   Each feature vector element is binary data whose bit length is the number of field lines extending from each intersection, and each bit value is binary depending on whether or not a field line extending in each direction is extracted. The camera calibration apparatus according to claim 1, wherein the camera calibration apparatus is set to be either.   6. The camera calibration according to claim 5, wherein each of the feature vector elements is binary data of 4 bits, and each bit position is associated with each of four directions extending from each intersection and orthogonal to each other. apparatus.   The means for generating the camera intersection feature vector element repeats the search for field line pixels within a predetermined range in each direction from each intersection on the camera image, integrates the number of discoveries, and if the integrated value exceeds a predetermined threshold, 7. The camera calibration apparatus according to claim 5, wherein a field line is determined to extend in a direction.   8. The camera calibration apparatus according to claim 7, wherein the means for generating the camera intersection feature vector element interrupts a search in a direction with a low discovery rate of field line pixels. The means for extracting the field lines comprises means for classifying each extracted field line into either a vertical field line or a horizontal field line;
9. The camera calibration apparatus according to claim 1, wherein the means for calculating the intersection calculates an intersection between each vertical field line and each horizontal field line.   The means for generating the camera intersection feature vector Vc arranges the camera intersection feature vector elements of each intersection in the order in which the intersections are raster-scanned to form a camera intersection feature vector Vc. The camera calibration device according to any one of 9.   The means for extracting the model intersection feature vector Vm obtains a Hamming distance D between the camera intersection feature vector Vc and each model intersection feature vector Vm, and extracts a model intersection feature vector Vm that minimizes the Hamming distance D. The camera calibration device according to claim 1, wherein:   The means for extracting the model intersection feature vector Vm is a feature of the invisible intersection of the camera intersection feature vector Vc when an intersection that is not visible on the camera image is included in the intersection of the field lines extracted from the camera image. 12. The camera calibration apparatus according to claim 1, wherein a feature vector element of an intersection corresponding to a model intersection feature vector Vm to be compared is copied to a vector element portion. In a method for performing camera calibration based on a homography matrix between a model of a field and a camera image,
Extract field lines from camera images,
Calculate the intersection of the field lines,
Generating feature vector elements for each intersection based on the number and orientation of field lines passing through the intersection;
A feature vector element of each feature point is arranged in a predetermined order to generate a camera intersection feature vector Vc,
Generate model intersection feature vector elements for each field line intersection in the field model,
Each model intersection feature vector element is arranged in the predetermined order for each combination of model intersection feature vector elements whose intersections are similar to each camera intersection feature vector element of the camera intersection feature vector Vc, and a plurality of model intersection feature vectors Generate Vm
A corresponding point is set based on a correspondence relationship between the model intersection feature vector Vm and the camera intersection feature vector Vc having a high similarity with the camera intersection feature vector Vc,
A camera calibration method, wherein a homography matrix is calculated based on the corresponding points. In a method for performing camera calibration based on a homography matrix between a model of a field and a camera image,
The procedure to extract the field line from the camera image;
The procedure to calculate the intersection of the field lines;
Generating a feature vector element for each intersection based on the number and orientation of field lines passing through the intersection;
A procedure for generating the camera intersection feature vector Vc by arranging the feature vector elements of the feature points in a predetermined order;
A procedure for generating a model intersection feature vector element for each intersection of each field line of the field model;
Each model intersection feature vector element is arranged in the predetermined order for each combination of model intersection feature vector elements whose intersections are similar to each camera intersection feature vector element of the camera intersection feature vector Vc, and a plurality of model intersection feature vectors The steps to generate Vm,
A procedure for setting corresponding points based on the correspondence between the model intersection feature vector Vm and the camera intersection feature vector Vc having a high degree of similarity with the camera intersection feature vector Vc;
A camera calibration program for causing a computer to execute a procedure for calculating a homography matrix based on the corresponding points.