JP2009095631A - Golf swing measuring system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To enable golf swing measurement by using a swing image photographed outdoors without being affected by background changes. <P>SOLUTION: This golf swing measuring system includes a computing means which captures at least one color image of a golfer who is swinging with a golf club in his/her hands. The computing means is characterized by having image processing functions such as extracting a plurality of marker points which help specify the swing position of the golfer from the swing image, acquiring their positional coordinates, drawing an initial outline in accordance with each body regional shape after tracing the positional coordinates of the plurality of the extracted marker points, processing the initial outline to optimize it, and extracting silhouettes regarding key regions on the golfer's swing position. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a golf swing measurement system, and in particular, from a color image of a golfer during a swing photographed outdoors such as where a golf swing measurement device is not installed, from a series of operations during swing, an address position, an impact A silhouette of a golfer at an arbitrary position such as a position or a finish position can be efficiently extracted by the calculation control means.

  2. Description of the Related Art Conventionally, various devices have been provided that can capture a swing when a golfer is hit, and automatically calculate various information such as a flying distance and a trajectory of the hit ball and display it on a golfer. Although these devices allow golfers to examine the quality of their own flying balls, they have not been able to provide useful information to improve the essential form. Therefore, it is required to accurately measure a golfer's form, that is, a swinging posture, and to improve the form based on this.

Japanese Patent No. 2794018 (Patent Document 1) provides an operation diagnosis device as a device capable of performing swing diagnosis by measuring the swing posture of the golfer, and includes a plurality of operation points on the golf club head and the body of the diagnosis subject. Is provided to acquire the coordinates of the operating point in the swing moving image and perform swing diagnosis.
However, the device cannot grasp the body movement on the image unless an operating point such as a mark is attached to the body of the person to be diagnosed, and the user who is the person to be diagnosed is troubled with preparation work for receiving a diagnosis. There is a disadvantage that gives it. In addition, it is necessary to perform a calculation for extracting the coordinates of a large number of operation points for all the frames of the swing moving image, which causes a problem that the calculation amount is enormous and the misidentification rate of the position coordinates of the operation points is increased.

  In order to solve the above problem, the present applicant, in Japanese Patent Application Laid-Open No. 2005-270480 (Patent Document 2), takes a picture of a swing at the time of hitting a golfer, and subtracts the skin color area of the golfer from a background image from the acquired color image. Extract the silhouette of the golfer, extract the attention point of the golfer's body from the skin color area, that is, extract the joint part necessary for swing diagnosis, acquire the position coordinates, and address from the acquired position coordinates Swing diagnosis is performed by analyzing the joint position at the time of impact, impact and finish.

The use of the swing measurement system disclosed in Patent Document 2 has an advantage that the mark required in Patent Document 1 does not need to be attached to the golfer in advance.
The method of extracting a golfer's silhouette by subtracting a swing image from a background image in Patent Document 2 is based on the premise that the background on which the golfer is photographed does not change. When a color image is obtained by swinging in front of the attached camera, the background hardly changes, and therefore, the swing measurement can be accurately performed with the system of Patent Document 2.

However, in the case of a color image taken outdoors such as a driving range, there is a problem that the silhouette of the golfer cannot be accurately extracted if the background changes due to weather changes.
In addition, when the installation location of the swing diagnostic device is moved, the background is different, so it is necessary for the dedicated operator to change the threshold for extracting silhouettes from the background, which is very troublesome in operation. There is.

  Because of the above-described problems, the swing diagnosis needs to be performed at a specific indoor location where the swing diagnosis apparatus is installed, and a swing image taken outdoors by a golfer, such as a golf driving range, is taken, and the swing image is captured. Even if it is taken into the computer of the measurement system of Patent Document 2, there is a problem that the measurement accuracy is deteriorated.

Japanese Patent No. 2794018 JP 2005-270480 A

  The present invention improves the golf swing measurement system affected by the background image of Patent Document 2 and extracts feature points to be noted on the image even when using a color image taken outdoors where the background changes. It is an object of the present invention to provide a golf swing measurement system that can perform and accurately extract a golfer's silhouette at an arbitrary position during swing from the feature point.

In order to solve the above-mentioned problems, the present invention comprises an arithmetic processing means for capturing at least one color image obtained by photographing a golfer who holds a golf club and swings.
The arithmetic processing means includes:
Extracting a plurality of feature points that can identify a golfer's swinging posture from the golfer's swing image, obtaining position coordinates of the feature points, and connecting the extracted position coordinates of the plurality of feature points to each body part The image processing function is provided to extract the silhouette of the required part of the golfer's swing posture by forming an initial contour line corresponding to the shape of the golf ball, optimizing the initial contour line and converging to the optimal shape. A golf swing measurement system is provided.

  Specifically, a swing image and a background image are acquired from the color image captured by the arithmetic processing unit, a difference value obtained by subtracting the swing image and the background image, a gray image from the swing image, and a gray image based on the gray image. An edge image is acquired by the Sobel method, an evaluation value is created using the difference image and the edge image, and a line connecting the feature points is changed to a Bezier curve so that the evaluation value is minimized. ing.

  As described above, when the silhouette (ie, the contour line of the golfer's posture during swing) is extracted using only the difference image obtained by subtracting the swing image and the background image, the accuracy of the silhouette extraction is improved when the background changes. Although it becomes worse, in the present invention, the difference image and the edge image are used together, an evaluation formula is constructed based on these two images, and the approximation curve is changed by an optimization algorithm so that the evaluation formula is minimized. Thus, silhouette extraction that fits the golfer's silhouette can be performed. As a result, even if the shooting environment changes and the background changes, it is possible to extract a silhouette with high accuracy, and a golf swing measurement system that is resistant to changes in the shooting environment can be obtained. Therefore, the swing diagnostic device can be installed outdoors such as a driving range, the installation position can be moved indoors, and there is no need to change the threshold for silhouette extraction, so a dedicated operator is unnecessary. it can.

The golfer's color image during the swing may be taken by an analog or digital dedicated camera or a camera attached to an electronic device including a mobile phone.
That is, any color image can be used as long as it can be captured by the arithmetic processing means. For example, a swing image captured by a camera of a mobile phone may be transferred. Furthermore, if the software of the arithmetic processing means is stored in the mobile phone, the captured image can be processed and the silhouette image can be displayed on the screen by the mobile phone.
In this way, by using a golf swing measurement system that is resistant to changes in the background, it can be made very convenient.

Further, the color image captured by the arithmetic processing means is at least one still image acquired from a moving image continuous from an address during swinging indoors or outdoors to the finish, or at least one still image between the address and the finish. Consists of.
For example, when it is desired to know only the form at the impact position at the time of swing, it is possible to extract only the color image at the impact position and extract the silhouette at the time of impact.
When a moving image obtained by capturing a series of motions during swing is captured, silhouettes are extracted based on each still image as a plurality of still images at a required position, for example, address, takeback shaft 8 o'clock , Take back shaft 9 o'clock, take back non-dominant arm level, top, down swing non-dominant arm level, down swing shaft 9 o'clock, impact, follow shaft 3 o'clock, the silhouette at each position of the finish can be extracted, You can check the form at each position.

The edge image is obtained by acquiring the swing image, acquiring a grayscale image from the acquired information, and acquiring an edge image by a Sobel method based on the grayscale image.
The edge strength is composed of the intensity of each pixel on the edge image, and the evaluation value is set using the difference value and the edge strength. The edge strength is extracted from the feature point, the initial contour. It is used for line formation and the optimization process.

The positions of the feature points extracted from the color image are preferably the rear neck, the back and front end positions of the waist position, the toe position of the shoe, the heel position, and the knee position.
By defining the above-described six positions as feature points, the above-described initial contour line is created, and the initial contour line is further optimized to perform accurate silhouette extraction. Note that feature points may be set at positions other than those described above. However, if the number of feature points is increased, calculation processing time is required. Therefore, silhouette extraction can be performed by specifying the positions of the six locations.

  The difference value obtained by subtracting the swing image and the background image around the body part including each feature point and the color information obtained from the golfer's clothing color area and the difference value obtained by subtracting the swing image are obtained, and the two obtained difference values A grayscale image is acquired from the information, and the feature points are specified and extracted from a binarized image obtained by binarizing the grayscale image.

When extracting the feature point of the back neck, a search range around the body part where the back neck point is predicted to exist, specifically, the head and neck portions are set as the search range, and within the search range Clothing color extraction is performed.
In Patent Document 2, skin color extraction is used to extract a golfer's face and neck. However, in skin color extraction, the face area becomes extremely small due to the influence of the hairstyle. The head region is extracted.
In addition, since the search range is set in the present invention, even if a color approximate to the golfer's clothes color is included in the still image, this can be eliminated to prevent misrecognition, and the calculation time can also be reduced. It can be shortened.
Forming a gray image of the golfer's head acquired by extraction of the clothing color region, binarizing the gray image, subjecting the binary image to a shrinking process as necessary, and then expanding it, The lowermost point of the head is obtained to extract the feature point of the rear neck, and the pixel of the image at the lowermost position is specified as the position coordinate of the feature point of the rear neck.

Extraction of the feature points of the back side of the waist position and both end positions of the front surface is performed by the same method as the extraction of the feature point of the back neck, and the periphery of the waist position is set as a search range. Since the golfer usually attaches a belt to the waist position, the golfer acquires a difference value obtained by subtracting the color information acquired from the color area of the belt and the swing image, and acquires a grayscale image from the acquired information of the two difference values. The binarized image is binarized to obtain a binarized image, the binarized image is contracted, and then expanded to obtain the rightmost point and the leftmost point. The front and rear end points are extracted, and the pixels at the front and rear end points are specified as the position coordinates of the feature points.
Note that belt end point extraction filter processing is performed when the both end positions of the belt, in particular, the front end position on the abdomen side cannot be extracted depending on the illumination conditions and the arm position.

  The feature points of the toe position and the heel position of the shoe are extracted by obtaining a difference value obtained by subtracting a swing image around the shoe including the feature point and a background image, obtaining a grayscale image from the obtained information, and obtaining the grayscale image. The feature points are identified and extracted from the binarized image obtained by binarizing the image.

  Since the shape of the shoe can be acquired from a difference image obtained by subtracting the swing image and the background image, it is not necessary to acquire a golfer's clothing color area, and a search area surrounding the shoe shape portion is set, and the search area Obtaining a grayscale image from the difference image, obtaining a binarized image obtained by binarizing the grayscale image, shrinking the binarized image, and then expanding the binarized image, the lowermost right point Are extracted with the toe position of the shoe and the lower leftmost point as the heel position, and the position coordinates of the feature point are specified.

  Extraction of the feature point at the position below the knee is performed by obtaining an evaluation value composed of a difference value between a swing image around the body part having the feature point and a background image and the edge strength, and specifying a portion where the evaluation value is maximized And extracted.

The feature point at the lower knee position is used for extracting the silhouette of the lower limb by combining with the feature point of the toe and the heel.
First, a search area around the knee point periphery is set, an evaluation value consisting of the difference value between the swing image and the background image of the search area and the edge strength is obtained, and the evaluation value that maximizes the evaluation value is determined and specified. Then, the specified position is extracted as a feature point below the knee, and the position coordinates of the feature point are specified.

For the formation of the initial contour line, around the line connecting the feature points, a swing image and a background image are subtracted to obtain an evaluation value consisting of the difference value and the edge strength,
A plurality of points existing between the feature points as control points,
An approximate curve generated by moving the control point so that the evaluation value is minimized is formed.

The initial contour line of the back connecting the feature points first connects the feature point of the back neck and the feature point on the back side of the waist with a straight line, and provides, for example, two control points between the straight lines. The pixel at the position is moved so that the evaluation value is minimized, and the approximate curve is formed by forming a Bezier curve from the feature point of the rear neck, the control point, and the feature point on the back side of the waist.
In the case of the Bezier curve having a complicated silhouette shape between the rear neck point and the waist rear side feature point, an approximate curve formed using a higher-order Bezier curve is preferably used as the initial contour line.

  As described above, the evaluation value is a value indicating an optimum silhouette using the difference image and the edge image. Specifically, the feature value of the evaluation value is obtained by calculating the average value of edge strength on the contour line acquired from the edge image, the background difference value average value in the region inside the contour line acquired from the difference image, and the contour outside the contour line. A background difference average value is used, and a weighting coefficient is added to each of these feature amounts, and an equation obtained by multiplying the total value by −10000 is used as an evaluation equation, and a value acquired from the evaluation equation is used as an evaluation value. The position coordinate of the control point that minimizes the evaluation value is moved, the Bezier curve is formed from the control point and the feature point, and the initial contour line is determined.

The optimization process of the initial contour line obtains an evaluation value composed of a difference value between the swing image and the background image of the initial contour line and the edge intensity, and a curve having the minimum evaluation value is obtained by a quasi-Newton method Is used to converge to the optimal shape.
That is, the optimization process is performed using the quasi-Newton method so that the initial contour line formed as described above is approximated by a golfer's silhouette.

  If the arithmetic processing means is a computer, the silhouette image obtained by the optimization process is displayed on the display attached to the computer or printed on a printer and provided to the golfer. Further, it can be displayed on the screen of the mobile phone device.

As described above, in the golf swing measurement system of the present invention, even if the background changes, the influence point is removed to extract the feature point of the golfer, the contour line of the golfer is obtained from the feature point, and the contour line is obtained. Since it converges to the optimal shape, it is possible to accurately extract the silhouette of the golfer when swinging.
In addition, when a silhouette image is extracted by using a difference image obtained by subtracting a swing image and a background image and the above-described edge image, a background change can be removed and a golfer silhouette image can be extracted. Therefore, even when the background changes, such as when using a color obtained by shooting a swing outdoors such as a golf course where environmental conditions such as sunshine change, or when the installation position of the golf swing measuring device is moved. The silhouette at the time of swing can be extracted accurately without being affected by the change of the background.

  Further, in the golf swing measurement system of the present invention, the position coordinates of the golfer's feature point on the image can be derived without a plurality of marks or the like for recognizing the golfer's feature point. Therefore, the golfer who performs the swing measurement can easily install the measuring instrument and the like, and as described above, the shooting location can be taken at any position regardless of whether it is outdoors or indoors, so the convenience is improved. To do.

  Further, from a plurality of still images constituting the color moving image, the address, takeback shaft 8 o'clock, takeback shaft 9 o'clock, takeback non-dominant arm horizontal, top, down swing non-dominant arm horizontal, down swing shaft 9 o'clock, impact At the time of the follow shaft 3, at least one or more swing posture silhouette images selected from the finish can be acquired.

  Furthermore, if a procedure for obtaining the position coordinates of the feature point of the golfer after selectively extracting still images useful for golf swing diagnosis, for example, address, impact, and finish, It is not necessary to acquire the position coordinates of the golfer's feature points for all frames (still images), and the calculation cost can be reduced. In addition, by limiting the still images to be calculated to only the necessary images, the misidentification rate of feature points can be reduced.

Embodiments of the present invention will be described with reference to the drawings.
FIG. 1 is a schematic diagram of a golf swing measurement system, and includes a computer 16 serving as arithmetic processing means, a monitor 17 serving as display means connected to the computer 16, a keyboard 18 serving as input means connected to the computer 16, and A mouse 19 and a color CCD camera 14 (hereinafter referred to as the camera 14) connected to the computer 16 and installed at the front position of the golfer 11 are provided.

In this embodiment, the swing of the golfer 11 is taken with a camera 14 outdoors such as a golf course or a golf practice place, and a color image taken with the camera 14 is stored in a memory of a hard disk and installed indoors at a distance. The memory is taken into the computer 16. Or you may electronically transmit to the said computer 16 from the computer which a golfer possesses.
The camera 14 may be a camera owned by a golfer, or may be a mobile phone camera. In addition, the camera 15 may be disposed on the side surface of the golfer, and further, may be disposed on the back side, and may be photographed at the same time or during different swings.
Further, when the measurement is performed at a predetermined indoor golf swing measurement place, the camera 14 and the camera 15 are connected online to the computer 16.

A golfer 11 (to be right-handed) whose swing is measured holds the golf club 13 with the outer garment U, the trousers D, the shoes S, and the belt B fastened.
The golfer 11 does not necessarily need to wear dedicated clothes for measurement, but if the photographer wears the dedicated clothes and shoots, the amount of calculation with a computer can be reduced, and it is not necessary to add marks or the like to the clothes or skin. A mark may be attached to a location in order to shorten the point extraction time. In addition, when the golfer is tightening a belt of the same color as the clothes, it is preferable to wear a belt for measuring different colors.
When swinging, it is preferable to swing on the mat 12 from the viewpoint of measurement.
Since the back surface 20 on which the golfer 11 is photographed is outdoors, it may change depending on the sunshine conditions.

  The computer 16 has a built-in program for extracting a golfer's silhouette from the captured color image during swing, performs image processing in accordance with the program process, and displays the extracted silhouette on an attached monitor 17 as necessary. Is printed. Further, the acquired silhouette may be transmitted by telegraph to the golfer's computer.

  A program built in the computer 16 extracts feature points M1 to M6, which will be described later, from the swing image of the golfer 11, and can extract the position coordinates of these feature points M1 to M6. An initial contour line corresponding to the shape of each body part is formed by connecting the position coordinates of a plurality of feature points, the initial contour line is optimized and converged to the optimal shape, and the required part of the golfer's swing posture It has an image processing function to extract silhouettes.

Specifically, the program built in the computer 16 is
A processing unit that selectively extracts a still image necessary for swing measurement from a color image captured by the computer 16;
A processing unit for obtaining a difference image obtained by subtracting the swing image and the background image from the selected color image;
A difference value information between the swing image and the background image, and a processing unit for acquiring a grayscale image from difference value information between the clothing color information of the golfer and the swing image;
A processing unit for binarizing the grayscale image;
A processing unit for extracting the feature points;
A processing unit that acquires an edge image by a Sobel method based on the difference image and the swing image; and an evaluation value is created using the difference image and the edge image, and the feature points are set such that the evaluation value is minimized. A processing unit for forming a contour line that changes an approximate curve made of a Bezier curve with a line connecting
It has a processing unit that outputs the acquired silhouette.
The evaluation value is set using the difference value and the edge strength, with the intensity of each pixel on the edge image as the edge strength.

  As shown in FIG. 2, the feature points extracted by the processing unit 54 are points that can be easily extracted on the outline of the human body, such as the neck M1, the back side of the waist position, and both end points of the front side (taken from the front of the golfer). In the image, the leftmost end point M2 of the belt, the rightmost end point M3), the heel point M4 and the toe point M5 of the shoe, and the lower knee point M6 are six locations.

  The rear neck point M1, the leftmost end point M2 of the belt, and the rightmost end point M3 are extracted by setting a head as a search area from the lower neck part, a difference value obtained by subtracting a swing image from the search area and a background image, and a golfer. A binarized image obtained by obtaining a difference value obtained by subtracting the color information obtained from the clothing color area and the swing image, obtaining a grayscale image from the obtained information of the two difference values, and binarizing the grayscale image M1, M2, and M3 of each feature point are identified and extracted, and position coordinates are set on the image.

  The shoe toe point M5 and the heel point M4 are extracted by setting the shoe periphery as a search range, obtaining a difference value between the swing image of the search range and the background image, and obtaining a grayscale image from the obtained information. Then, the feature points M4 and M5 are specified and extracted from the binarized image obtained by binarizing the grayscale image, and the position coordinates are set on the image.

  For the extraction of the lower knee point M6, the lower limbs are set as a search area, an evaluation value consisting of the difference value between the swing image and the background image of the search area and the edge strength is obtained, and the evaluation value is maximized. The position is identified and the position of M6 is extracted, and the position coordinates are set on the image.

The initial contour line is formed after extracting the feature points M1 to M6 and specifying the position coordinates. For the formation of the initial contour line, the evaluation value composed of the difference value and the edge strength is obtained by subtracting the swing image and the background image around the line connecting M1 to M6,
A plurality of points existing between the feature points are defined as control points,
An approximate curve generated by moving the pixel of the control point so that the evaluation value is minimized is formed.

  After forming the initial contour line, optimization processing is performed in order to increase the accuracy of the contour line, and an evaluation value composed of a difference value between the swing image and the background image of the initial contour line and the edge strength is obtained. The curve having the smallest evaluation value is converged to the optimum shape using the quasi-Newton method.

  Hereinafter, a method of extracting feature points M1 to M6, forming an initial contour line, and optimization processing will be described. In this description, the still image selected from the color image is an image at the time of golf swing address.

  First, extraction of the feature points M1 to M6 that are nodes (end points) forming the outline will be described.

The extraction of the feature point of the back neck M1 will be described in detail below.
As described above, the back neck M1 sets the head from the lower neck portion as a search area, the difference value obtained by subtracting the swing image and the background image of the search area, and the color information acquired from the golfer's clothing color area and the swing A difference value obtained by subtracting the images is acquired, a grayscale image is acquired from information of the acquired two difference values, and a feature point M1 is extracted from the binarized image obtained by binarizing the grayscale image. The position coordinates are set.

First, as shown in FIG. 3, the right shoulder point A1 is extracted and the search range is determined. The search range includes a background color area H1 and a clothing color area H2.
In order to clarify the right shoulder point, a mark may be attached to the right shoulder. In this case, a pixel is a search range with a required range centered on the coordinate position of the extracted right shoulder mark.
The background color region H1 is (x-10, y + 10), (x-10, y) where x is the horizontal direction of the image, y is the vertical direction, and the coordinates of the extracted right shoulder point are (x, y). −70), (x + 70, y−70) (x + 70, y + 10), which is in the range of 80 pixels × 80 pixels. The clothing color region H2 is a range of 20 × 20 pixels of (x−10, y + 10) (x−10, y + 30) (x−30, y + 30) (x−30, y + 10).

ここで、０≦Ｒ、Ｇ、Ｂ≦２５５、Ｓ、Ｖともに値の範囲０〜１になる。The norm between the clothes color image and the swing image is calculated.
The hue (H), saturation (S), and brightness (V) of each pixel in the clothing color region H2 are calculated from Formulas 1 to 3 using a cylindrical model.

Here, 0 ≦ R, G, B ≦ 255, S, and V all fall within the range of values 0 to 1.

このように、ＳおよびＶをＬｏｇ化することによりＳ、Ｖが小さいときに値の変化が大きくなるようになる。なお、Ｓ、Ｖともに値の範囲は０〜ｌｏｇ_１０１０１となる。Further, when S and V are converted by Log, Expression 4 is obtained.

Thus, by making S and V log, the change in value becomes large when S and V are small. Note that the range of values for both S and V is 0 to log ₁₀ 101.

Next, the HSV is converted into three-dimensional space coordinates. Here, the length of the vector is S, and the angle between the X axis and the vector is H.

Average values X _ave , Y _ave , and Z _ave are calculated by calculating X, Y, and Z in the range of 20 pixels × 20 pixels in the clothing color region H2. Using the average value of the clothing color, each pixel x, y, z is calculated using the swing image in the H1 region, and the norm is calculated. However, normalization is performed with the maximum length of the XYZ space. Therefore, the norm range is 0-1.

Next, the norm between the background color image and the swing image is calculated.
First, with respect to the background image, the range of 80 pixels × 80 pixels in the background color region H1 is set to X _heikei , Y _heikei , and Z _{heikei for} each pixel from Equations 1-5. For the swing image, x, y, and z are obtained for each pixel, and the background color norm is obtained.

Next, in Equations 6 and 7, the minimum value of either the clothing norm or background color norm is selected for each pixel. This selected value becomes the value of the grayscale image, and the grayscale image shown in FIG. 4 is obtained. However, a grayscale image is acquired except for the mark area under the jaw (30 <H <80, S> = 0.15, V> = 100) and the “right shoulder” of the red mark extracted when the search range is determined. Yes.
After obtaining the grayscale image, the head 21 is extracted and binarized as shown in FIG. The threshold used in the binarization process is determined by the P tile method so that the head 21 is 30% of H1. Further, as shown in FIG. 5B, the binarized one region is repeated twice in the order of contraction and expansion of the head 21. After that, as shown in FIG. 5C, an orthogonal line d2 that is perpendicular to the straight line d1 drawn from the center of gravity of the head 21 to the lowest point of the head 21 is drawn. Then, an intersection M1 between the orthogonal line d2 and the end of one region of the back of the head is defined as a feature point M1 of the back neck.

Next, extraction of the leftmost end point M2 and the rightmost end point M3 of the belt will be described in detail.
The search range of the belt color region is searched from almost the entire image. At this time, the golfer is photographed almost at the center of the image. The belt color region H3 is a red band as shown in FIG. In the belt color region H3, as shown in FIG. 7, the leftmost end point M2 and the rightmost end point M3 are extracted.

In the extraction of M2 and M3, x = 150 to 490 and y = 160 to 320 are set as the search range (the size of the video is x for 640 and y for 480). The red information of the belt region H3 is H = 330 to 360, 0 to 30, S = 0.2 or more, and V = 15 to 150.
Similarly to the rear neck, the belt region H3 acquires a grayscale image from the background color image and the clothing color image, binarizes the grayscale image, and acquires the leftmost end point M2 and the rightmost end point M3 of the belt region H3.

However, depending on the illumination conditions, the end of the belt region H3 may not be extracted. In this case, the leftmost end point M2 and the rightmost end point M3 of the belt region H3 are extracted by the following method.
As shown in FIG. 8B, the end of the belt region H3 is set as a region C1 giving a positive weight, while the circular region C is set as a region C2 giving a negative weight so as to surround the positive region C1, and the positive region C1 A circular area C is formed from the minus area C2.
The width of the belt region H3 of the region C1 to which the positive weight is given is 5 pixels in the vertical direction, and the maximum in the horizontal direction of the positive region is 6 pixels. On the other hand, the size of the circle in the region C2 giving a negative weight is 11 pixels in diameter, the positive weight is 1.0, and the negative weight is -1.0.

なお、各ピクセルの色情報Ｈ_ｉｊ、Ｓ_ｉｊ、Ｖ_ｉｊ（ｌｏｇへの変換）は数式１〜４と同様に求める。ただし、Ｓ_ｉｊは２つの色情報の差を大きくするために、１．０に固定してもよい。
色類似度を求める範囲は、後首領域と同様、ｘ＝１５０〜４９０、ｙ＝１６０〜３２０とする。Next, the belt color similarity is calculated.
The red color of the belt region H3 is assumed to be (H, S, V) = (0.0, 0.6, 0.33), and converted to the XYZ space. The difference between the color information of each pixel (after XYZ conversion) and the red color of the belt region H3 is taken.

Note that the color information H _ij , S _ij , and V _ij (conversion to log) of each pixel are obtained in the same manner as Equations 1 to 4. However, S _ij may be fixed to 1.0 in order to increase the difference between the two color information.
The range for obtaining the color similarity is x = 150 to 490 and y = 160 to 320 as in the rear neck region.

ｄ_ｉｊは数式７の背景色ノルムと同じとする。In the background area, the background difference value is considered so that the color similarity is 0, and the norm is subtracted from 1 in order to increase the value closer to red.
Accordingly, the belt color similarity is given by the following equation.

Let d _ij be the same as the background color norm of Equation 7.

ここで、（ｘｇ、ｙｇ）は領域重心、Ｓは領域面積とし、ｍ２０は重心を原点にした場合のＹ軸についての二次モーメント、ｍ０２は重心を原点にした場合のＸ軸についての二次モーメンある。The search areas H4 and H5 at the end of the belt area are areas of 25 pixels square from the leftmost end point M2 and the rightmost end point M3 of the belt B.
As shown in the following equation, the inclination angle θ of the main axis L1 is obtained from the color extracted region.

Here, (xg, yg) is the region centroid, S is the region area, m20 is the second moment about the Y axis when the center of gravity is the origin, and m02 is the secondary moment about the X axis when the center of gravity is the origin. There is momentum.

ここで、Ｆ_ｉｊは円領域Ｃの重み、Ｒはベルト色類似度とする。
数式１１で求めた評価値が最大となるところを特徴点とする。Next, the circular region C that forms the plus region C1 and the minus region C2 is rotated so as to match the angle θ in the principal axis direction L1 obtained by Expression 10. After the rotation, the search areas H4 and H5 are folded from the upper left to the lower right. An evaluation value result at a specific pixel in the search areas H4 and H5 is expressed by the following equation.

Here, F _ij is the weight of the circular area C, and R is the belt color similarity.
A point where the evaluation value obtained by Expression 11 is maximized is defined as a feature point.

Next, extraction of feature points of the saddle point M4 and the toe point M5 will be described.
As described above, the heel point M4 and the toe point M5 are extracted by setting the shoe periphery as a search range, obtaining a difference value obtained by subtracting the swing image of the search range and the background image, and obtaining a grayscale image from the obtained information. And the feature points M4 and M5 are identified and extracted from the binarized image obtained by binarizing the grayscale image.

More specifically, the saddle point M4 and the toe point M5 require a gray image only from the background image.
First, as shown in FIG. 9, a point P on the shoe 23 obtained by projective transformation from the front right foot tip coordinates is extracted, and the entire foot is surrounded with the point P as a reference (x, y). Set the search range. The search range is (x + 40, y-25), (x + 40, y + 25) (x-80, y + 25) (X-80, y-25).
Next, as in the rear neck region, as shown in FIGS. 10A to 10C, a difference image is obtained to obtain a grayscale image, binarized, contracted and then expanded, The lower left point is acquired as the saddle point M4, and the lower rightmost point is acquired as the toe point M5.
For binarization, it is assumed that the foot area is a fixed ratio with respect to the area of the region. Specifically, binarization is performed by determining a threshold value of 1200 with a foot of 20% for a region area 120 × 50 from the P tile method.

なお、Ｄ_ｘはカラー画像もしくは濃淡画像におけるＸ方向のソーベル強度であり、Ｄ_ｙはカラー画像もしくは濃淡画像におけるＹ方向のソーベル強度である。Next, extraction of the lower knee point M6 that is the intersection of the foot and the mat will be described in detail.
As shown in FIG. 11, the search range H6 is 10 pixels × 40 pixels square with reference to the line L2 at the mat end.
(1) edge strength, (2) inner background difference value (red region), (3) outer background difference value (blue region), (1) × 1.0 + (2) × 1.0 + (3) × ( -1) is calculated pixel by pixel from the heel and toe center line L3 toward the toes, and the point having the maximum value is set as the foot / mat intersection M6 as shown in the table of FIG. 12).
The edge strength D is expressed by the following equation.

Incidentally, D _x is the Sobel intensity of X-direction in a color image or grayscale image, D _y is the Sobel intensity in the Y direction in a color image or grayscale image.

After extracting the features M1 to M6 as described above, initial contour lines having the feature points M1 to M6 as nodes are formed.
Below, the formation method of an initial stage outline is explained in full detail.
As for the back contour line, first, as shown in FIG. 13, points equally divided between the back neck point M1 and the belt leftmost point M2 are set as control points Q1 and Q2 of the back contour.
In the case of an address image, as shown in FIG. 13A, the points obtained by dividing the straight line A connecting the back neck point M1 and the belt leftmost point M2 into three equal parts are set as control points Q1 and Q2.
In the case of an impact image, as shown in FIG. 13B, the points obtained by dividing the straight line A connecting M1 and M2 into four equal parts are defined as control points Q1, Q2, and Q3.
The control point may be determined by connecting the rear neck point M1 and the rear belt point M2 with a low-order Bezier curve.

  In forming the rear leg contour line, first, as shown in FIG. 14, control points P1, P2, and P3 of the rear leg contour are obtained. The control points P1 to P3 are obtained by setting the length of the straight line B connecting the saddle point M4 and the leftmost end point M2 of the belt to L1, and on the tangent line C obtained from the back contour of M2, from M2 to the foot direction. A point extended by a distance of L1 / 3 is defined as a control point P1. Further, a midpoint Pa of a straight line B connecting the saddle point M4 and the M2 is obtained, and a straight line parallel to the tangent line C and passing through the midpoint Pa is defined as a straight line D. A point extended from the middle point Pa by a distance of L1 / 5 upward on the straight line D is defined as a control point P2, and a midpoint between the control point P2 and the saddle point M4 is defined as a control point P3.

The initial contour of the back is formed from feature points M1, control points P1, P2, and feature points M2, and will be described in detail below with reference to an address.
As shown in FIG. 15, when the length of the straight line A connecting the back neck point M1 and the leftmost end point M2 on the belt rear side is L2, the straight line A connecting the back neck point M1 and the leftmost end point M of the belt In the vertical direction, the control points Q1 and Q2 are moved pixel by pixel to a distance of (L2) / 3 on the back contour side to obtain a contour shape that minimizes the evaluation value.
Since the control point is moved by a minimum of L2 / 3 pixels, there are L2 / 3 evaluation values. From the L2 / 3 evaluation values, the curve having the minimum evaluation value is determined as the initial contour line.

The evaluation value is obtained from the difference value between the swing image and the background image (outside and inside of the contour line) and the edge strength on the contour line, and is expressed by Equation 13.

In Equation 13, Vedge is the average edge strength on the model curve shown in FIG. 16A, and is obtained by Vedge = (total edge strength in contour model pixels) / (number of contour model pixels).
Vin is an average of background difference values in the region inside the contour model shown in FIG. 16B, and is obtained by Vin = (total of background difference values in the contour model inner region) / (inner evaluation region area).
Vout is an average of background difference values in the region outside the contour model shown in FIG. 16B, and is obtained by Vin = (total of background difference values in the contour model outer region) / (outer evaluation region area).
In the present embodiment, as shown in FIG. 16B, it is assumed that the width in the case where the area of the region is obtained for Vin and Vout is 20 pixels.

Here, Win, Wout, and Wedge are weighting factors.
In back contour extraction at the time of addressing, Win = 1, Wout = −1, and Wedge = 1.
Win = 1, Wout = −1 and Wedge = 0.25 in the rear leg contour extraction and front leg contour extraction at the time of address, and Win = 1, Wout = −1 in the back contour extraction at the time of impact. Wedge = 0.2. In the front leg contour extraction at the time of impact, Win = 1, Wout = −1, and Wedge = 1. Note that the values of Win, Wout, and Wedge are not necessarily the above values, but are adjusted according to the color of the pants.

  When the curve correctly extracts the silhouette of the golfer, Vin becomes larger because the difference image exists when the background and the swing are differentiated. Vout becomes “0” when there is a difference between the background and the swing because there is no difference image. The reason why the weighting factor Wout of this part is negative is to increase the evaluation value when the curve is not correctly extracted on the silhouette (when it exists on the human side). In addition, when the curve correctly extracts a golfer's silhouette, Vedge has the largest value.

The edge strength used when obtaining Vedge will be described below.
A grayscale image is obtained from the swing image, and an edge image is obtained by the Sobel method. Since the edge image is acquired not from the RGB image but from the grayscale image, the edge image is obtained by Expression 14 based on FIG.
Note that the conversion from a swing image to a grayscale image is obtained by grayscale image = 0.3 × R + 0.59 × B + 0.11 × B.

Next, formation of the rear leg initial contour line will be described.
First, the control points P1, P2, and P3 set as described above are moved to obtain an evaluation value, and a curve having the smallest evaluation value is determined as the evaluation value.
For example, as shown in FIG. 18, the control point P1 is moved to a position separated by a predetermined pixel (10 pixels vertically and horizontally) in the front-back, left-right, and diagonal directions, and a total of nine evaluation values are obtained at each position. Similarly, the control points P2 and P3 are also moved rearward and diagonally to obtain evaluation values at the respective positions. A total of 9 × 9 × 9 evaluation values are obtained to obtain a combination of control points P1, P2, and P3 that minimize the evaluation value, and a curve is determined as an initial contour line.

Next, formation of the front leg initial contour line will be described.
As the initial contour of the front leg, the convergence result of the rear leg contour, that is, the control point coordinates and the end points which are the optimal shape of the rear leg contour are used. As shown in FIG. 19, the end points and control point coordinates of the rear leg contour are rotated and scaled. Obtained by conversion. The end points are the leftmost end point M2 and the saddle point M4 at the rear end of the belt on the rear leg, and the rightmost end point M3 and the lower knee point M6 at the front end of the belt on the front leg.
As shown in FIG. 20, the scale conversion magnification is a ratio of the length of the straight line E connecting the belt rear end M2 and the saddle point M4 to the length of the straight line F connecting the belt front end M4 and the knee point M6. The rear leg curve is compressed in the direction of the straight line E. Further, in the rotational movement, the difference between the inclinations of the two straight lines E and F is obtained, and the rear leg curve is rotated so as to match the inclination of the straight line F.

Next, formation of the rear leg initial contour line at the time of the top will be described with reference to FIGS.
In order to obtain the initial hind leg contour at the time of top, first, as shown in FIG. 21, the convergence result obtained by converging the hind leg outline at the address by the optimization process described later is used. The rear leg initial contour line at the time of take 8 is obtained by rotational movement and scale conversion as described above. A convergence result is obtained from the rear leg initial contour line at the time of the take 8 as shown in FIG.
Next, the end point and the control point coordinates are rotationally moved and scale-converted from the convergence result at the time of take 8 to obtain the rear leg initial contour line when the arm is horizontal. The convergence result when the arm is horizontal is obtained from the rear leg initial contour line.
Next, the end point and the control point coordinates are rotationally moved and scale-converted from the convergence result to obtain the initial hind leg contour at the top.
In this way, the contour line is sequentially tracked and deformed from the address time to the horizontal time of the take arm, and the rear leg initial contour line at the top time is obtained. Note that the initial contour line may be obtained not only at the time of take 8 and when the arm is horizontal, but also in other swing postures.
When the contour of another image is not used, the initial contour may be obtained by the same method as that for addressing.

  After the initial contour is formed, the contour is optimized and converged to the optimal shape. The optimization process uses a quasi-Newton method to obtain control point coordinates that minimize the contour evaluation value.

First, the principle of the quasi-Newton method will be described.
First, consider the minimization problem of the function f (x).
Quadratic model approximated by X = _{X k} Taylor expansion of f (x) in the vicinity of (up to second order) becomes Equation 15.

数式１６は、図２３の直線Ｇを表している。
ｆ（ｘ）は２次式なのでｆ’（ｘ）＝０のときｘが最小になる。即ち、数式１７で表される。

上式を解いて求めたｘをｘ_ｋ＋１として更新していくと、数式１８に示すようにｆ’＝０に収束していく。

In Equation 15, H (x _k ) is a Hessian matrix (matrix _{obtained by} differentiating f twice), and a positive definite symmetric matrix (for all n-dimensional vectors x in which the n × n matrix A is not x = 0, When the inequality x ^T Ax> = 0, the function f becomes a downward convex function. That is, f ′ = 0 is minimum.
When f (x) is differentiated, Equation 16 is obtained.

Equation 16 represents the straight line G in FIG.
Since f (x) is a quadratic expression, x is minimized when f ′ (x) = 0. That is, it is expressed by Expression 17.

When x obtained by solving the above equation is updated as x _{k + 1} , it converges to f ′ = 0 as shown in Equation 18.

In Equation 18, if the function to be optimized is a quadratic function, 1 is before the inverse of the Hessian matrix. However, when it is applied to an actual problem, it is more efficient to search α sequentially and return the best result. Therefore, using Equation 19, x is updated as x _{k + 1} and converged to the minimum value of f (x).
At this time, the Hessian inverse matrix (x _{k + 1} ) ⁻¹ becomes Equation 20.

In this embodiment, as shown in FIG. 24, X of f (x) is a vector value of the control point P. For example, when there are two control points P1 and P2, x = (x1, y1, x2). , Y2).
As described above, the control points P1 and P2 are moved within a certain range to obtain respective evaluation values, and x when the evaluation value is minimum is set to _{xk + 1} . Specifically, α in Equation 19 is changed to obtain x _{k + 1} , a Bezier curve is created for each x _{k + 1} , and an evaluation value is obtained for each Bezier curve. X _{k + 1} that minimizes the evaluation value is selected, and x _{k + 1} is substituted for x in Equation 19 again. In this way, the calculation of Expression 19 is repeated to obtain x having a high evaluation value.

The calculation of the optimization process ends in any of the following (1) to (3).
(1) When the number of calculations exceeds 500 (2) When the evaluation value in the previous calculation is hardly different from the evaluation value in this calculation (the absolute value of the evaluation value change is the argument eps = 0.001) Less than)
(3) When the differential value (slope) of the evaluation function is eps or less (when the function shape is almost flat)
When the calculation of the optimization process is completed, x becomes the coordinate position of the control points P1 and P2 having the smallest evaluation value.

FIG. 25 is a flowchart for obtaining the coordinate position of the control point having the smallest evaluation value by the quasi-Newton method.
In step S1, the initial value of the control point is x (0), and the evaluation value is f (x (0)).
Let k = 0, H (X0) ⁻¹ = I (unit matrix).
In step S2, f ′ (x (k)) is calculated, and it is determined whether | f ′ (x (k)) | <ε. Alternatively, it is determined whether k> 500 or | f (x (k)) − f (x (k−1)) | <ε (k> 0). If the conditional expression is satisfied, the flowchart is terminated.
In step S3, Formula 19 is calculated. α is in steps of 0.01 from 0 to 1, and 101 types of x (k + 1) are obtained. 101 types of Bezier curves are created, and 101 evaluation values are obtained for each of the Bezier curves.
In step S4, x (k + 1) that minimizes f (x (k + 1)) from 101 types of x (k + 1) is determined. That is, α that minimizes the evaluation value is determined, and x (k + 1) to be used when returning to step S3 from Equation 19 is determined.
In step S5, f ′ (x (k + 1)) is calculated. f ′ (x (k + 1)) is obtained by Expression 21 and is a differential vector of the function f (x). Note that f (x) is a scalar, but f ′ (x) is a vector.
In step S6, the Hessian inverse matrix of Equation 20 is calculated.
In step S7, k = k + 1 is set, and the process returns to step S2.
The flowchart of FIG. 25 is repeated until the calculation of the optimization process is completed in step S2, and X at the time of completion becomes the coordinate position of the control points P1 and P2.

Note that the Bezier curve in step S3 can be obtained from the coordinate positions of the control points and feature points. x (k + 1) indicates the coordinate positions of the control points P1 and P2, and in the example of FIG. 26, the coordinate positions of the control points P1 and P2, and the coordinates of the rear neck point M1 and the belt leftmost point M2 that are characteristic points. A cubic Bezier curve G is obtained from the four points of the position, and is used as an optimum contour line.
A cubic Bezier curve can be expressed as Equation 22. When a Bezier curve is plotted on an image, the value of t is changed in increments of 0.01, and the pixel information on the curve is determined by rounding off the decimal point. In Equation 21, B ₀ indicates the coordinate position of the rear neck point M1, B ₃ indicates the coordinate position of the belt leftmost point M2, and B ₁ and B ₂ indicate the coordinate positions of the control points P1 and P2.

27A to 27D and FIG. 28 show the first embodiment.
In the first embodiment, the back contour line at the time of addressing is extracted.
The still image at the time of addressing the golfer shown in FIG. 27 (A) is stored in a computer, and the difference image shown in FIG. 27 (B) and the edge shown in FIG. 27 (C) are stored by a program built in the computer. An image is formed. Using the above-described extraction of the feature points M1 and M2, an initial contour line is formed, the acquired initial contour line is optimized, and the optimized contour line 101 shown in FIG. A golfer's silhouette is obtained (shown in red line).
FIG. 28 shows the process of creating and converging five contour lines in order to minimize the evaluation value V when the initial contour line is formed and when the optimization process after the initial contour line is formed. ing.

29 and 30 show a second embodiment.
In the second embodiment, the rear legs of the address image shown in FIG. 29 and the contour lines of the front legs shown in FIG. 30 are extracted.
FIG. 29A includes an optimized rear leg contour 102. As shown in FIG. 29B, five contours are created so that the evaluation value V is minimized, and the process of convergence is shown. Show.
FIG. 30A shows the progress of convergence by creating an outline 6 times so as to minimize the evaluation value V as shown in FIG. ing.

31 and 32 show a third embodiment.
In the third example, the rear leg outline 104 of the top image is extracted.
The golfer shown in FIG. 31A is photographed, but the still image at the top is stored in the computer, and the difference image shown in FIG. 31B and the edge shown in FIG. An image is formed. The initial contour line is formed by using the extraction of the feature points M1 and M2 described above, the obtained initial contour line is optimized, and the golfer having the optimized contour line 104 shown in FIG. Get a silhouette.
FIG. 32 shows a process of creating and converging six contour lines in order to minimize the evaluation value V when the initial contour line is formed and when the optimization process after the initial contour line is formed. ing.
Note that the initial value at this time uses the result of convergence in the horizontal direction of the take arm.

  As described above, the golf swing measurement system of the present invention has a silhouette extraction function for data of a swing image taken in color by a digital camera, an analog camera, or a mobile phone camera at any place such as outdoors or indoors. When the image processing program is loaded into a computer, a silhouette fitted to the golfer's silhouette can be extracted based on a still screen selected from an arbitrarily selected address, top, impact, finish, or the like.

In addition, by creating an optimized silhouette using the above-mentioned edge image, even if the golfer's background changes due to sunshine conditions, etc., the golfer's silhouette can be extracted without being affected by the change. It can be performed with high accuracy.
In this way, not only the difference image between the background image and the swing image but also the edge image can be used together to extract the silhouette even if the background changes slightly. As a result, the golf swing measurement device was installed. Even if it is not a dedicated place, swing measurement can be performed using a swing image taken outdoors.

  FIG. 1 is a schematic view showing a configuration of a golf swing measurement system of the present invention.   These are figures which show the feature point extracted at the time of golf swing measurement.   It is drawing which shows the search range for calculating | requiring the feature point of a back neck, and the reference point of a search range.   It is drawing which shows the grayscale image obtained from FIG.   4A is a diagram showing an image obtained by binarizing the image of FIG. 4, FIG. 5B is a diagram in which (A) is contracted and then expanded, and FIG. 5C is an explanatory diagram showing feature points of the back neck.   It is drawing around a belt area.   It is drawing which shows the feature point of a belt area | region.   (A) is a drawing for explaining a search region at the belt region end, and (B) is a conceptual diagram showing a minus region and a plus region at the belt region end.   It is drawing explaining a saddle point and a toe point.   (A) is a drawing showing a binarized image, (B) is a drawing showing an image obtained by contracting and then expanding (A), and (C) is an image for obtaining a saddle point and a toe point from (B). It is drawing to explain.   It is drawing explaining the search range and reference | standard line for calculating | requiring the intersection of a foot | mat | matte.   It is drawing explaining the intersection of a foot and a mat.   It is explanatory drawing of formation of the line of a back outline, (A) is an address image, (B) is a case of an impact image.   It is explanatory drawing of formation of the line of a back leg outline.   It is explanatory drawing of formation of a back initial outline.   (A) is explanatory drawing of the edge strength average on a model curve, (B) is explanatory drawing of the background difference value average in the area | region inside an outline model, and the background difference value average in the area | region outside an outline model.   It is explanatory drawing which calculates | requires edge strength.   It is explanatory drawing of how to move a control point.   It is explanatory drawing of scale conversion and rotation movement of formation of a front leg initial outline.   It is explanatory drawing of formation of a front leg initial outline.   It is explanatory drawing of the rear leg initial stage outline at the time of a top.   It is explanatory drawing of the rear leg initial stage outline at the time of a top.   It is explanatory drawing of an optimization process.   It is explanatory drawing of a control point.   It is a flowchart which shows an optimization process.   It is explanatory drawing of a Bezier curve.   The back contour line forming process at the time of address in the first embodiment is shown, (A) to (C) are photographs showing images of the respective processes, and (D) is a photograph in which the result of the silhouette is superimposed on the original swing image. is there.   It is a table | surface which shows the progress which formed the outline so that it might become the minimum with respect to the evaluation value of 1st Example.   (A) is a photograph in which the rear leg contour line at the address of the second embodiment is superimposed on the original swing image, and (B) is a table showing the process of forming the contour line so as to be the minimum with respect to the evaluation value. is there.   (A) is a photograph in which the front leg contour line at the time of addressing in the second embodiment is superimposed on the original swing image, and (B) is a table showing the process of forming the contour line so as to be the minimum with respect to the evaluation value. .   The top leg hind legs contour forming process of the third embodiment is shown, (A) to (C) are photographs showing images of the respective processes, and (D) is a photograph in which the result of the silhouette is superimposed on the original swing image. It is.   It is a table | surface which shows the progress which formed the outline so that it might become the minimum with respect to the evaluation value of 3rd Example.   It is a table | surface which shows extraction of the knee point in FIG. 11 and FIG.

Explanation of symbols

DESCRIPTION OF SYMBOLS 11 Golfer 14 Camera 16 Computer M1-M6 Feature point 100-104 Contour line converged by optimization processing

Claims (10)

Computation processing means for capturing at least one color image of a golfer who holds and swings a golf club,
The arithmetic processing means includes:
Extracting a plurality of feature points that can identify a golfer's swinging posture from the golfer's swing image, obtaining position coordinates of the feature points, and connecting the extracted position coordinates of the plurality of feature points to each body part The image processing function is provided to extract the silhouette of the required part of the golfer's swing posture by forming an initial contour line corresponding to the shape of the golf ball, optimizing the initial contour line and converging to the optimal shape. A unique golf swing measurement system. A swing image and a background image are obtained from the color image captured by the arithmetic processing means,
A difference value obtained by subtracting the swing image and the background image, a grayscale image is obtained from the swing image, an edge image is obtained by a Sobel method based on the grayscale image,
The golf according to claim 1, wherein an evaluation value is created using the difference image and the edge image, and a line connecting the feature points is changed from an approximate curve made of a Bezier curve so that the evaluation value is minimized. Swing measurement system.   The golf swing measurement system according to claim 2, wherein the intensity of each pixel on the edge image is defined as an edge strength, and the evaluation value is set using the difference value and the edge strength. The golfer's color image during the swing is taken by an analog or digital dedicated camera, or a camera attached to an electronic device including a mobile phone,
4. The method according to claim 1, further comprising: at least one still image acquired from a continuous moving image from an address during swinging indoors or outdoors to a finish, or at least one still image between an address and a finish. The golf swing measurement system according to claim 1.   The position of the feature point extracted from the color image includes the rear neck, both end positions on the back side and the front side of the waist position, the toe position of the shoe, the heel position, and the lower knee position. The golf swing measurement system according to item 1.   Extraction of the respective feature points at the back side and the front side both ends of the back neck and the waist position is based on a difference value obtained by subtracting a swing image and a background image around the body part including the feature points and a golfer's clothing color area. A difference value obtained by subtracting the acquired color information and the swing image is acquired, a grayscale image is acquired from information of the acquired two difference values, and the feature point is obtained from the binarized image obtained by binarizing the grayscale image. The golf swing measurement system according to claim 5, wherein the golf swing measurement system is specified and extracted.   The feature points of the toe position and the heel position of the shoe are extracted by obtaining a difference value obtained by subtracting a swing image around the shoe including the feature point and a background image, obtaining a grayscale image from the obtained information, and obtaining the grayscale image. The golf swing measurement system according to claim 5, wherein the feature point is specified and extracted from a binarized image obtained by binarizing the image.   Extraction of the feature point at the position below the knee is performed by obtaining an evaluation value composed of a difference value between a swing image around the body part having the feature point and a background image and the edge strength, and specifying a portion where the evaluation value is maximized The golf swing measurement system according to claim 5, wherein the golf swing measurement system is extracted. For the formation of the initial contour line, around the line connecting the feature points, a swing image and a background image are subtracted to obtain an evaluation value consisting of the difference value and the edge strength,
A plurality of points existing between the feature points as control points,
The golf swing measurement system according to any one of claims 3 to 8, wherein the golf swing measurement system is formed by an approximate curve generated by moving the control point so that the evaluation value is minimized.   The optimization process of the initial contour line obtains an evaluation value composed of a difference value between the swing image and the background image of the initial contour line and the edge intensity, and a curve having the minimum evaluation value is obtained by a quasi-Newton method The golf swing measurement system according to claim 3, wherein the golf swing measurement system is converged to an optimal shape using

Images