JP2008284166A - Pitch type identification apparatus, identifier generator, pitch type identification program, and identifier generation program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a pitch type identification apparatus capable of determining the type of a pitch from a picture of the pitch taken by a single camera. <P>SOLUTION: This pitch type identification apparatus B has: a learning result storage device (identifier storage mans ) 5 prestoring identifiers for identifying the pitch type based on trajectory data, or information on the positions of pitches in respective frame images of the respective pitch pictures of a plurality of pitches, and trajectory data mechanically leaned based on information of pitch types previously associated with the respective pitches; a pitch trajectory analysis device (a pitch trajectory analysis means) 2b for inputting a picture of the pitch, or an identified object, extracting the pitch in each frame image of the image and generating trajectory data; and a pitch identification device (pitch type identification means) 6 for identifying the type of the pitch, or the identified object, based on the trajectory data generated by the pitch trajectory device 2b using the identifier stored in the learning result storage device 2. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to a technique for analyzing a pitch from a pitch video, and more particularly, to a pitch type identifying device, a classifier generating device, a pitch type identifying program, and a classifier generating program for identifying a pitch type. .

  In baseball, various pitches such as straight, curve, and fork are used depending on the pitcher. Whether or not it is hit by a grasshopper is often determined by assembling the pitch, and a viewer with knowledge of baseball views the next ball type in anticipation. In this way, viewers are highly interested in information on the type of sphere, and information on the type of sphere is provided by super display, data broadcasting, the Internet, and the like in relay programs. In addition, if the information on the type of ball is stored as metadata, it can be used for later scene search and analysis of pitching contents.

  And the apparatus which displays the locus | trajectory of pitching conventionally is disclosed (refer patent document 1). This device uses image recognition technology to extract and track a ball area from a video at a field rate in real time, and draws CG (computer graphics) indicating a locus at the obtained position. This makes it possible to provide information useful to the viewer because the drop of the changing ball can be seen at a glance.

  Also, an apparatus for determining a strike zone is disclosed (see Patent Document 2). This apparatus determines, by image processing, whether or not a ball has entered a strike zone from images of a plurality of cameras that photographed the vicinity of a home base.

Further, a device is disclosed in which a golf ball moving from a player is tracked by image processing at a golf driving range to notify characteristics such as a golf flight distance and a track (see Patent Document 3).
Japanese Patent Laying-Open No. 2005-123824 (paragraph numbers 0026 to 0086) JP-A-9-290037 (paragraph numbers 0012 to 0033) Special table 2001-521403 gazette

  However, conventionally, most of the ball types are determined and confirmed by an expert visually by one ball, and a high baseball knowledge is required for the determination. In addition, since human beings make the determination, there is a risk that subjectivity will be mixed.

  Moreover, although the apparatus of patent document 1 can grasp | ascertain the trajectory of a pitch at a glance, it cannot determine what kind of ball it is. Therefore, the discrimination of the ball type is left to the judgment of the commentator or viewer who has seen the trajectory.

  On the other hand, the apparatus of Patent Document 2 can determine whether or not the ball has entered the strike zone, but cannot determine the type of the ball. Further, in this apparatus, it is necessary to install a plurality of cameras, but it is preferable that only one existing broadcasting camera can be operated.

  And the apparatus of patent document 3 requires a plurality of cameras, and cannot be applied to the ball type determination with a monocular camera.

  The present invention is made in order to solve the problems of the prior art, and a pitching type identification device capable of determining the pitch type of the pitch from the video of the pitch shot by one camera, It is an object of the present invention to provide a discriminator generating device, a pitching ball type identifying program, and a discriminator generating program.

  In order to solve the above-mentioned problem, the pitch type identification device according to claim 1 is a pitch type identification device for identifying the pitch type of the pitch from the pitch video, the classifier storage means, the pitch trajectory An analysis unit and a ball type identification unit are provided.

  According to such a configuration, the pitching ball type identifying device stores, in the classifier storage means, trajectory data that is time-series information of the position of the video object of the pitch in the time-series image constituting the pitched video, A discriminator for discriminating the ball type from the trajectory data, machine-learned for a plurality of pitches based on the pitch type information associated with the pitch in advance, is stored in advance. Further, the pitching ball type identifying device inputs a pitching video as a pitch type identification target by the pitching trajectory analyzing means, extracts a pitching video object in a time-series image constituting this video, and Pitch trajectory data is generated, and the ball type discriminating unit identifies the pitch type of the pitch to be identified from the trajectory data generated by the pitch trajectory analyzing unit using the discriminator stored in the discriminator storage unit.

  Thereby, the pitching ball type identifying device can generate trajectory data from the video of the pitching that is the target of the pitching type identification, and can identify the pitching type of the pitching. Here, the throwing ball type identifying device stores the discriminator for identifying the ball type, which is generated by machine learning in advance using the trajectory data of the pitch whose known ball type is known. Using the discriminator, the pitch type of the pitch can be identified from the trajectory data of the pitch to be identified.

  According to a second aspect of the present invention, there is provided the pitched ball type identifying apparatus according to the first aspect, wherein the classifier storage unit normalizes the pitched ball trajectory data according to a predetermined standard from the plurality of pitched track data. Based on the calculated trajectory shape of the pitch, the characteristic amount indicating at least one characteristic of the relative position of the pitch and the speed of the pitch, and information on the pitch type of the pitch previously associated with each pitch A discriminator for discriminating the ball type from the feature quantity, which has been machine-learned, is stored in advance, and the ball type discriminating unit is registered in accordance with the predetermined reference from the trajectory data generated by the throwing trajectory analyzing unit. A pitch type of the pitch to be identified from the feature quantity calculated by the feature quantity calculation means using a feature quantity calculation means for calculating the converted feature quantity and a classifier stored in the classifier storage means Identify That was configured to have an identification means.

  Thereby, the pitch type identification device generates trajectory data from the video of the pitch to be identified as the pitch type, and the feature amount indicates at least one of the shape of the pitch trajectory, the relative position of the pitch, and the ball speed Can be calculated to identify the pitch type of the pitch. Here, the ball type from the feature amount indicating at least one of the shape of the pitch trajectory, the relative position of the pitch, and the ball speed, which is generated by machine learning in advance using the pitch trajectory data of which the ball type is known. By storing a discriminator for discriminating in advance, the pitching ball type identifying device can identify the pitch type of the pitch from the characteristic amount of the pitch to be discriminated using this discriminator. . Note that the relative position of the pitch is the relative position of the pitch with respect to the position of the pitch obtained by analyzing a plurality of pitch positions such as an approximate curve of the trajectory by a predetermined method. Indicates the position or the relative position of the pitch with respect to a person or an object in the video.

  The discriminator generating device according to claim 3 is a discriminator generating device for generating a discriminator for identifying the pitch type of the pitch used in the pitched ball type identifying device according to claim 1. The pitch trajectory analyzing means and the ball type learning means are provided.

  According to such a configuration, the discriminator generation device inputs a plurality of pitch videos by the pitch trajectory analysis means, extracts the pitch video objects in the time-series images constituting each video, and Trajectory data that is time-series information of the position of the video object of the pitch in the image is generated, and the trajectory data generated by the pitch trajectory analyzing means is associated with each pitch in advance by the ball type learning means. Based on the pitch type information of the pitch, a discriminator for identifying the pitch type from the trajectory data is generated by machine learning.

  As a result, the discriminator generation device can generate a discriminator for identifying the ball type from the trajectory data by machine learning based on the trajectory data of a plurality of pitches in which the ball type information is associated in advance. it can.

  According to a fourth aspect of the present invention, at least one characteristic of the shape of the pitch of the pitch, the relative position of the pitch and the speed of the ball used in the pitch type identification device of the second aspect. A discriminator generating device for generating a discriminator for discriminating a ball type from a feature amount indicating a pitch trajectory analysis unit and a ball type learning unit, wherein the ball type learning unit includes a feature amount calculation unit. And learning means.

  According to such a configuration, the discriminator generation device inputs a plurality of pitch videos by the pitch trajectory analysis means, extracts the pitch video objects in the time-series images constituting each video, and Trajectory data that is time-series information of the position of the video object of the pitch in the image is generated, and the trajectory data generated by the pitch trajectory analyzing means is associated with each pitch in advance by the ball type learning means. Based on the pitch type information of the pitch, a discriminator for identifying the pitch type from the trajectory data is generated by machine learning. Here, the discriminator generation device calculates the feature value normalized by a predetermined reference for each pitch from the trajectory data generated by the pitch trajectory analysis unit by the feature amount calculation unit of the ball type learning unit. Then, a learning unit uses a machine learning to identify a discriminator for identifying the ball type from the feature amount based on the plurality of pitch feature amounts calculated by the feature amount calculating unit and the corresponding ball type information. Generate.

  Thereby, the discriminator generation device calculates a feature amount indicating at least one feature of the shape of the pitch of the pitch, the relative position of the pitch, and the speed of the ball normalized by a predetermined reference from the plurality of track data. Thus, based on the feature amount and the information on the ball type previously associated with the pitch, a discriminator for identifying the ball type from the feature amount can be generated by machine learning.

  Furthermore, the pitching type identification program according to claim 5, the trajectory data which is time-series information of the position of the video object of the pitch in a time-series image constituting the video of the pitch, and each of the pitches A discriminator for discriminating the ball type from the trajectory data, which is machine-learned for a plurality of pitches and stored in advance in the discriminator storage device based on the pitch type information associated with the pitch in advance. In order to identify the pitch type of the pitch from the pitch video, the computer is caused to function as a pitch trajectory analysis means and a pitch type identification means.

  According to this configuration, the pitching type identification program inputs a pitching video as a pitch type identification target by the pitching trajectory analysis means, and extracts a pitching video object in a time-series image constituting this video. Then, the trajectory data of the pitch to be identified is generated from the trajectory data generated by the pitch trajectory analyzing means using the discriminator stored in the discriminator storage device by the ball type discriminating means. Identify.

  As a result, the pitching type identification program can generate trajectory data from the video of the pitch to be identified as the pitch type and identify the pitch type of the pitch.

  According to a sixth aspect of the present invention, there is provided a discriminator generating program for generating a discriminator for discriminating a ball type used in the pitched ball type identifying device according to claim 1, a pitch trajectory analyzing means. It was decided to make it function as a ball type learning means.

  According to such a configuration, the discriminator generation program inputs a plurality of pitch videos by the feature amount calculation means, extracts the pitch video objects in the time-series images constituting each video, and Trajectory data that is time-series information of the position of the video object of the pitch in the image is generated, and the trajectory data generated by the pitch trajectory analyzing means is associated with each pitch in advance by the ball type learning means. Based on the pitch type information of the pitch, a discriminator for identifying the pitch type from the trajectory data is generated by machine learning.

  As a result, the discriminator generation program can generate a discriminator for identifying the ball type from the trajectory data by machine learning based on the trajectory data of a plurality of pitches in which the ball type information is associated in advance. it can.

  The pitched ball type identifying device, the classifier generating device, the pitched ball type identifying program, and the classifier generating program according to the present invention have the following excellent effects.

  According to the invention described in claim 1 or claim 5, the pitch type of the pitch can be identified from the video of the pitch to be identified as the pitch type. For this reason, conventionally, it is possible to automatically determine the ball type that has been visually determined by an expert having knowledge about the ball type from the video of the pitch. Further, by adding the obtained information on the type of sphere to the video as metadata, it is possible to automatically generate metadata for video search.

  According to the second aspect of the present invention, it is possible to calculate the normalized feature amount from the video of the pitch to be identified as the ball type and identify the ball type of the pitch. Then, by obtaining the normalized feature value, it is possible to calculate the feature value that eliminates the effects of camera movement during shooting and the difference in the number of frames due to the difference in ball speed. can do.

  According to invention of Claim 3 or Claim 6, the discriminator for identifying a sphere type from locus | trajectory data can be automatically produced | generated by machine learning. According to the invention described in claim 4, it is possible to automatically generate a discriminator for discriminating the sphere type from the normalized feature quantity by machine learning. Then, by obtaining the normalized feature value, it is possible to calculate the feature value that eliminates the effects of camera movement during shooting and the difference in the number of frames due to the difference in ball speed. Possible classifiers can be generated.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. Here, a ball type determination system including a discriminator generation device and a pitching ball type identification device according to the present invention will be described.

[Configuration of ball type determination system]
First, the configuration of the ball type determination system S will be described with reference to FIG. FIG. 1 is a block diagram showing a configuration of a ball type determination system including a discriminator generating device and a pitching ball type discriminating device of the present invention. The ball type determination system S generates a discriminator that identifies a ball type of a pitch to be identified based on a plurality of past pitch information, and identifies the ball type. The ball type determination system S includes video storage devices 1a and 1b, a discriminator generating device A, and a pitching ball type discriminating device B. Further, the discriminator generation device A includes a pitching trajectory analysis device 2a, a trajectory data storage device 3, and a ball type learning device 4, and the pitching ball type identification device B includes a pitching trajectory analysis device 2b and a learning result storage. A device 5 and a ball type identifying device 6 are provided.

  The video storage devices 1a and 1b store a video obtained by shooting a pitch, and are general storage means such as a hard disk. The video accumulated here is read out by the pitching trajectory analysis devices 2a and 2b. Here, the video storage device 1a stores video of a plurality of pitches whose sphere types are known. In addition, the video storage device 1b stores a video of a pitch that is an identification target of a ball type.

  The pitching trajectory analysis devices 2a and 2b extract and track pitching video objects from the pitching video inputted from the video storage devices 1a and 1b, respectively, and generate trajectory data indicating the pitching trajectory. The trajectory data is time-series information of the position of the video object of the pitch in the video frame image. For example, the two-dimensional coordinates of the input video frame image in the horizontal axis and vertical axis directions are arranged in time series. It is a thing. Here, the trajectory data generated by the pitching trajectory analysis device (throwing trajectory analysis means) 2 a is output to the trajectory data storage device 3. It is assumed that the pitch type analyzed here is determined in advance by a specialist or the like. Further, the trajectory data generated by the pitching trajectory analyzing device (throwing trajectory analyzing means) 2 b is output to the ball type identifying device 6. The pitching trajectory analysis apparatus 2a and the pitching trajectory analysis apparatus 2b have the same configuration and functions except that the video input source and the trajectory data output destination are different. The detailed configuration of the pitching trajectory analysis device 2a (2b) will be described later.

  The trajectory data storage device 3 stores the trajectory data input from the pitch trajectory analysis device 2a in association with the ball type data, which is information about the ball type of the trajectory determined in advance. Storage means. The trajectory data and the ball type data accumulated here are read out by the ball type learning device 4.

  The ball type learning device (ball type learning means) 4 performs machine learning on an identifier for identifying the ball type from the track data stored in the track data storage device 3 and the ball type data associated with the track data. Is generated. Here, the ball type learning device 4 calculates a feature amount indicating at least one feature of the shape of the pitch of the pitch, the relative position of the pitch, and the ball speed based on the track data, and the feature amount and the ball type data. Thus, a discriminator for identifying the sphere type based on the feature amount is generated. The learning result that is the discriminator generated here is stored in the learning result storage device 5. The detailed configuration of the ball type learning device 4 will be described later.

  Here, conventionally, when an expert determines a ball type, it is often determined while watching a video of a pitch. Examples of the judgment material for judging the type of ball from the video of the pitch include a change in ball speed, a trajectory, and a pitcher pitch form. However, if the pitcher changes the pitch form according to the ball type, there is a fear that the ball type may be guessed by the grasshopper, so it is difficult to use the pitch form as a judgment material. Therefore, it is possible to determine the ball type only from the pitch or trajectory of the pitch, and the ball type learning device 4 analyzes the trajectory or the speed change from the trajectory data obtained from the video of the pitch, and the ball We decided to generate a classifier that identifies the species.

  A learning result storage device (discriminator storage means, discriminator storage device) 5 stores a discriminator as a learning result of the ball type learning device 4, and is a general storage means such as a hard disk. This learning result is referred to and used by the ball type identification device 6.

  The ball type identifying device (ball type identifying means) 6 uses the discriminator which is the learning result stored in the learning result storage device 5 based on the trajectory data generated by the pitch trajectory analyzing device 2b. It identifies the species. Here, the ball type identifying device 6 calculates a feature amount indicating at least one feature of the shape of the pitch of the pitch, the relative position of the pitch, and the ball speed based on the trajectory data, and based on the feature amount, The ball type is identified using the classifier stored in the learning result storage device 5. The identification result identified here is output to the outside. The detailed configuration of the ball type identification device 6 will be described later.

(Configuration of pitching trajectory analysis device)
Next, the configuration of the pitching trajectory analysis device 2a (2b) will be described with reference to FIG. 2 (refer to FIG. 1 as appropriate). FIG. 2 is a block diagram showing a configuration of a pitching trajectory analysis device of a ball type determination system including a discriminator generating device and a pitching ball type identifying device of the present invention. The pitching trajectory analyzing apparatus 2a (2b) includes an object candidate image generating unit 21, a ball selecting unit 22, an extraction condition storing unit 23, a position predicting unit 24, a search area setting unit 25, and a pitch interpolating unit 26. Prepare.

  The object candidate image generating means 21 cuts out a search area for each field image constituting the video from the video input from the video storage device 1a (1b), and selects a video object (throwing image) candidate to be tracked. The extracted object candidate image is generated. The object candidate image generation unit 21 includes image storage units 211 and 212, difference image generation units 213 and 214, and a candidate image generation unit 215.

  The video is composed of, for example, 60 field images per second. Therefore, the object candidate image generation means 21 extracts a video object candidate from the field image and binarizes it, thereby generating an image (object candidate image) consisting only of the video object candidate. This object candidate image is an image obtained by extracting a plurality of video objects similar to the video object to be tracked. For example, the object candidate image is an image obtained by roughly extracting a video object to be tracked, such as a video object with movement.

  The image storage units 211 and 212 are memories for performing various kinds of signal / image processing, and for example, record video signals as digital data in units of one field. The image storage unit 211 stores field images (odd field and even field) at an intermediate position (current) among three consecutive input field images. The current field image stored here is output to the image storage unit 212, the difference image generation units 213 and 214, and the ball selection unit 22. The image storage unit 212 stores field images (odd field and even field) at the first position (past) among three consecutive field images. The past field image stored here is output to the difference image generation unit 214.

  The difference image generation unit 213 subtracts the difference image 1 by subtracting the luminance of the newly input field image (future field image) from the luminance of the delayed field image (current field image) from the image storage unit 211. Is to be generated. Here, the difference image generation unit 213 generates the difference image 1 for the search area input from the search area setting unit 25 described later. The difference image 1 generated here is output to the candidate image generation unit 215.

  The difference image generation unit 214 generates the difference image 2 by subtracting the luminance of the current field image from the image storage unit 211 from the luminance of the delayed field image (past field image) from the image storage unit 212. Is. Here, the difference image generation unit 214 generates the difference image 2 for the search area input from the search area setting unit 25 described later. The difference image 2 generated here is output to the candidate image generation unit 215.

  The candidate image generation unit 215 compares the difference image 1 and the difference image 2 with a predetermined luminance threshold value for all the pixels in the search region, and the pixel value of the difference image satisfies the predetermined condition. A binarized image is generated by discriminating the value as “1 (white)” and “0 (black)” in other cases. This predetermined condition will be described later. Thus, the candidate image generation unit 215 can extract an area where the pixel value is “1 (white)” as a candidate video object. The binarized image generated here is output to the ball selection means 22 as an object candidate image obtained by extracting video object candidates.

  The ball selection unit 22 selects a video object to be extracted (tracked) from the object candidate images generated by the object candidate image generation unit 21 based on the extraction conditions stored in the extraction condition storage unit 23. Then, the position of the video object and the feature quantity characterizing the video object are extracted. Here, the ball selection unit 22 includes a labeling unit 221, a feature amount analysis unit 222, a filter processing unit 223, and an object selection unit 224.

  The labeling unit 221 assigns numbers (labels) to regions that are candidates for video objects in the object candidate images (binarized images) generated by the object candidate image generation unit 21. That is, the labeling unit 221 assigns one number to a connected area (connected area) having a pixel value of “1 (white)” that is an area of the video object. As a result, video object candidates in the object candidate image are numbered. The object candidate images labeled here are output to the feature amount analysis unit 222.

  For each video object candidate numbered by the labeling unit 221, the feature amount analysis unit 222 calculates the position coordinates of the video object candidates and the feature amounts (parameters) such as the area, luminance, color, and circularity of the video object. The value is calculated. The position coordinates and feature amounts calculated here are output to the filter processing unit 223.

The position indicates, for example, the position of the center of gravity of the video object. The area indicates the number of pixels of the video object, for example. Further, the luminance indicates an average value of the luminance of each pixel in the video object. The color indicates an average value of RGB values of each pixel in the video object. The circularity indicates the degree of circularity of the video object, and has a larger value as it is closer to a circular shape. For example, when the video object has a circular shape such as a ball, the circularity is a value close to 1. The circularity e is expressed by the following equation (1), where S is the area of the video object and L is the perimeter.
e = 4πS / L ² (1)

  The filter processing unit 223 uses the parameter value calculated by the feature amount analysis unit 222 to determine whether or not the video object matches the extraction condition information stored in the extraction condition storage unit 23, thereby performing extraction (tracking). ) The target video object is narrowed down. That is, for each video object candidate, the filter processing unit 223 extracts the extraction conditions (for example, area, luminance, color, and circularity) stored in the extraction condition storage unit 23 and the prediction by the position prediction unit 24 described later. Based on the position, the video object satisfying the extraction condition is extracted by filtering the feature quantity analyzed by the feature quantity analysis unit 222 (position filter, area filter, luminance filter, color filter, and circularity filter). Select as a candidate video object.

  The current field image (odd field and even field) stored in the image storage unit 211 is output to the ball selection means 22 as described above, and the filter processing unit 223 converts the input field image into one frame. Filtering is performed using an image delayed by (2 fields) as a reference image.

  The object selection unit 224 selects, as a ball, a video object candidate closest to the position coordinates of the video object extracted from the immediately preceding frame image among the video objects that have passed all the filters. The extracted position of the video object is stored in the extraction condition storage unit 23 as the current video object position information, and is output to the pitching interpolation unit 26. Here, for the position of the video object, the barycentric coordinates of the video object, the vertex coordinates of the polygon approximation, the control point coordinates of the spline curve, and the like can be used. In addition, when the video object suitable for the extraction condition cannot be selected, the object selection unit 224 notifies the pitch interpolation unit 26 to that effect (extraction failure).

  The extraction condition storage unit 23 stores conditions for selecting a video object to be extracted (tracked), and is a storage unit such as a general hard disk. The extraction condition storage means 23 stores extraction condition information indicating various extraction conditions and position information indicating the position of the video object.

  The extraction condition information is information describing the extraction condition of the video object to be extracted, and describes, for example, at least one extraction condition of area, luminance, color, and circularity. This extraction condition information includes a filter (area filter, luminance filter, color filter) for the filter processing unit 223 of the ball selection unit 22 to select a video object to be extracted from the object candidate image generated by the object candidate image generation unit 21. Filter and circularity filter).

  The extraction condition information stores a predetermined initial value and a threshold value indicating an allowable range as conditions for the area filter, the luminance filter, the color filter, and the circularity filter. Thereby, each filter of the ball selecting means 22 can exclude a video object having a value (feature) outside the threshold from candidates of video objects to be extracted.

  The position information stored in the extraction condition storage means 23 is information indicating the position of the video object being tracked. This position information is, for example, the barycentric coordinates of the video object. The barycentric coordinates are calculated by the feature amount analysis unit 222 of the ball selection unit 22. This position information is referred to when the pitch interpolation means 26 described later interpolates the position of the pitch that failed to be extracted, and there are a plurality of video objects that match the extraction condition information in the filter processing by the filter processing unit 223. In this case, the object selection unit 224 also uses the video object closest to the coordinates indicated by the position information as an extraction condition for determining the video object to be extracted.

  The position predicting unit 24 predicts the position of the video object in the field image to be input next based on the position of the video object selected by the ball selecting unit 22 (such as barycentric coordinates), and searches for the predicted position information. This is output to the area setting means 25. The position predicting unit 24 includes a linear predicting unit 241, a curve predicting unit 242, and a switching unit 243.

  The linear prediction unit 241 assumes that the trajectory of the video object extracted from a predetermined number of continuous field images is a straight line, and calculates a motion vector based on the position information obtained by the object selection unit 224 of the ball selection unit 22. This is used to predict the position of the video object in the field image to be input next. The position information predicted here is output to the search region setting means 25 in accordance with the switching of the switching unit 243 described later.

  For example, the linear prediction unit 241 predicts the position of the video object in the next field image (next frame) by applying a Kalman filter to the barycentric coordinates. The Kalman filter changes from moment to moment by applying a recurrence formula that performs “filtering” to estimate the current state based on the observed quantity observed in time series and “prediction” to estimate the future state. The state to be performed is estimated.

  The curve prediction unit 242 assumes that the trajectory of the video object extracted from a predetermined number of continuous field images is a quadratic curve obtained by the least square method, and the position obtained by the object selection unit 224 of the ball selection unit 22 Based on the information, the position of the video object in the field image to be input next is predicted. The position information predicted here is output to the search region setting means 25 in accordance with the switching of the switching unit 243 described later.

The curve prediction unit 242 updates a quadratic curve (y = ax ² + bx + c) used for curve prediction every time position information of a video object is input from the object selection unit 224 of the ball selection unit 22 and 2 The coefficients a and b of the next curve are monitored, whether or not the signs of the coefficients a and b have changed and whether or not the values of the coefficients a and b have exceeded a predetermined value are switched. Output to the unit 243.

Based on the number of position coordinates of the video object extracted in the past or the quadratic curve equation, the switching unit 243 detects the position information of the linear prediction by the linear prediction unit 241 and the position of the curve prediction by the curve prediction unit 242. The output of the information to the search area setting means 25 is switched. Here, for example, the switching unit 243 selects the curve prediction when the position coordinates extracted in the past of the video object are equal to or greater than a predetermined number, and selects the linear prediction when the position coordinates are smaller than the predetermined number. The predetermined number is three, for example, and five or more are preferable when the curve prediction unit 242 obtains and predicts a particularly accurate curve. Further, when the curve prediction unit 242 is executing the curve prediction, when the signs of the coefficients a and b of the quadratic curve equation (y = ax ² + bx + c) change, or the values of the coefficients a and b are predetermined. When the value is exceeded, the switching unit 243 switches from curve prediction to linear prediction.

  The search area setting means 25 uses the position information of the video object in the next field (next frame) predicted by the position prediction means 24 to set a search area for a video object within a predetermined range in the field image. is there. The search area setting means 25 outputs the set position and size of the search area to the object candidate image generation means 21 as search area information. Here, the size of the range of the search area and the position information of the search area before the video object is extracted are used as input values (not shown) such as a mouse and a keyboard not shown as the initial value of the search area. ) To the search area setting means 25. When the position information predicted from the position prediction unit 24 is not input, the search region setting unit 25 notifies the object candidate image generation unit 21 of the search region indicated by the initial value, and the position information is received from the position prediction unit 24. When input, a new search area is set based on the position information and information on the size of the search area range indicated by the initial value, and the object candidate image generation means 21 is notified.

  In response to the notification of the extraction failure from the object selection unit 224, the pitch interpolation unit 26, based on the position information of the video object of the past field image stored in the extraction condition storage unit 23, the field image that has failed to be extracted. The position information of the video object is obtained by interpolation. The position information of the video object from the object selection unit 224 and the position information of the video object obtained here are information indicating the trajectory of the pitch extracted as the video object (trajectory data). Alternatively, it is output to the ball type identification device 6).

  Note that, for example, in the pitch video for the left grasshopper, the field image when the ball passes through the grasshopper area may not be extracted. Trajectory data can be generated.

  As described above, the pitching trajectory analysis device 2a (2b) can extract the video object of the pitching from the input field image and output the position coordinates as trajectory data. The pitch trajectory analyzing apparatus 2a (2b) can extract the pitch position from the input field image in real time. Further, the pitching trajectory analyzing apparatus 2a (2b) can acquire the pitching position (image coordinates) until the ball away from the pitcher's hand reaches the catcher's mitt (or hits the grasshopper's bat).

  Note that the trajectory data is time-series position coordinate data of one pitch video object. For example, the position information of the pitch video object in the field image is two-dimensional coordinates in the horizontal and vertical axes, These two-dimensional coordinates may be arranged in time series.

(Configuration of ball type learning device)
Next, the configuration of the ball type learning device 4 will be described with reference to FIG. The ball type learning device 4 includes a feature amount calculating unit 41 and a learning unit 42.

  The feature amount calculating means 41 calculates the feature amount indicating the feature of the trajectory of the pitch for each trajectory data based on the trajectory data input from the trajectory data storage device 3 based on a predetermined standard. is there. The feature amount calculated here is output to the learning means 42.

  Here, since the trajectory data includes information such as changes in ball speed and the degree of trajectory bending, it may be used as a feature value as it is, but baseball pitch video is used for shooting with a fixed camera. In many cases, the angle of view changes for each ball. In addition, the number of extracted balls varies depending on the ball speed, and the number of data differs. Therefore, the feature quantity calculation means 41 normalizes and obtains feature quantities indicating features such as pitching speed, trajectory shape, and relative position of the pitching. In this way, the feature amount calculation means 41, for example, from the trajectory data of the video imaged by one camera without taking the three-dimensional position of the pitch by photographing with two cameras as in the prior art, It is possible to calculate the feature amount excluding the influence of the movement and the ball speed.

  Hereinafter, an example of a feature amount calculated by the feature amount calculation unit 41 and a calculation method thereof will be described. Here, the feature amount calculation means 41 approximates the number of frames (fields) from throwing to catching (hereinafter referred to as the number of extractions), all sections of the pitching trajectory, and a predetermined section as a feature amount. The slope of the straight line (hereinafter referred to as the first derivative), the curvature of the approximate curve (hereinafter referred to as the second derivative) when approximating the curve (secondary curve) over the entire pitch trajectory and the predetermined section, home base Two-dimensional coordinates in the field image of the catching position with the position of the origin as the origin (hereinafter referred to as catching position), and the amount of change in the horizontal axis direction and the change in the vertical axis direction of adjacent two-dimensional coordinates in the trajectory data The amount (hereinafter referred to as motion vector) was calculated. Note that these feature amounts are examples, and the feature amount calculation unit 41 may calculate other feature amounts indicating the characteristics of the pitching trajectory.

<Number of extractions>
Here, it is assumed that trajectory data (x, y) as shown in the following equation (2) is input to the feature amount calculation means 41. Then, the feature amount calculation unit 41 sets the number of data n as the number of extractions.
(X, y) = (x ₁ , y ₁ ), (x ₂ , y ₂ ),... (X _n , y _n ) (2)

  In general, in baseball, the time from throwing to catching is about 0.5 seconds. When the measurement rate of the image coordinates of the ball is a frame interval (1/30 second), approximately 15 coordinates can be acquired. In addition, the measurement rate of the pitch trajectory analysis apparatus 2a is the field rate (1/60 seconds) at the maximum. If the number of extractions is small, the ball is caught quickly, and if the number of extractions is large, it takes time to catch the ball. That is, the number of extractions indicates the speed of the ball.

<First derivative>
The feature quantity calculating means 41 generates an approximate line with respect to the trajectory, and sets the inclination of the approximate line as the first derivative. For example, as shown in FIG. 3, the feature amount calculating unit 41, a plurality of balls M indicated by the locus data, M, to generate the approximate line L ₁ with respect ... locus of. FIG. 3 is a schematic diagram schematically showing a pitching trajectory and approximate straight lines and approximate curves generated by the feature amount calculating means.

When placing the approximation line L ₁ as shown in the following expression (3), the first derivative a ₁ can be calculated by the following equation (4). A first derivative a ₁ indicating the inclination of the approximate straight line indicates the height of the ball. Here, the feature amount calculation means 41 calculates the primary differential coefficient for each of the pitching sections and the first, middle, and last three sections obtained by dividing all the sections into three.
y = a ₁ x + b ₁ (3)

<Secondary derivative>
The feature quantity calculation means 41 generates an approximate curve with respect to the trajectory, and uses the curvature of the approximate curve as a second derivative. For example, as shown in FIG. 3, the feature amount calculating unit 41, a plurality of balls M indicated by the locus data, M, to generate an approximated curve L ₂ relative ... locus of. When this approximate curve L ₂ is placed as in the following equation (5), the residual sum of squares S of the approximate curve L ₂ and the coordinates of the ball can be calculated by the following equation (6).
y = a ₂ x ² + b ₂ x + c ₂ (5)

Then, the feature quantity calculating means 41 obtains the coefficients a ₂ , b ₂ , and c _{2 by} partially differentiating the coefficients a ₂ , b ₂ , and c ₂ and solving the simultaneous equations with a value of 0. The secondary differential coefficient a ₂ indicating the curvature of the approximate curve L ₂ indicates the degree of curvature of the locus. Here, the feature quantity calculating means 41 calculates secondary differential coefficients for all pitching sections and the first, middle, and last three sections obtained by dividing all the sections into three.

<Catching position>
The catching position can be estimated from the image coordinates of the ball extracted last, but it is difficult to determine the catching position only from the image coordinates of the ball when the camera is not fixed and photographed. For this reason, the feature amount calculating means 41 finally detects the home base from the extracted image of the ball, and determines the relative position between the barycentric coordinates and the image coordinates of the ball as the catching position. For example, as shown in FIG. 4, when the image coordinates of the home base H are (x _h , y _h ) and the image coordinates of the last extracted ball M are (x _b , y _b ), the feature amount calculating means 41 designates a vector B _h (x _h −x _b , y _h −y _b ) from the home base H to the ball M as a catching position. FIG. 4 is a schematic diagram schematically showing the image of the ball extracted last and the catching position generated by the feature amount calculating means.

  Thereby, the feature amount calculation means 41 can acquire the catching position while eliminating the influence of the camera operation. Since the falling ball such as a fork has a low catching position, this catching position is a feature amount indicating whether or not the falling ball is a falling ball. Since the occlusion is small in the vicinity of the home base H, the feature amount calculating means 41 acquires the center of gravity position by detecting the region of the home base H by image processing from the video image corresponding to the trajectory data. Can do.

<Motion vector>
The feature amount calculation means 41 obtains the amount of change in the horizontal axis direction and the amount of change in the vertical axis direction of the image coordinates of adjacent balls in the trajectory data. For example, as shown in FIG. 5, indicated by the trajectory data, ball _M A adjacent, image coordinates of _{M B,} respectively _{(x A,} y _A), if a _{(x B, y} B), the feature amount calculating means 41 calculates the ball _M a, the motion vector _{B M (x B -x a,} y B -y a) between _{M B} a. FIG. 5 is a schematic diagram schematically showing a pitching trajectory and a motion vector calculated by the feature amount calculating means.

  Here, even if the pitcher throws from the exact same position, since the camera for shooting is not fixed, the position on the image of the trajectory for each ball fluctuates. Is calculated as a feature amount, and the feature amount is free from the influence of panning and tilting of the camera. Since the number of extracted trajectory data differs for each pitch, the number of motion vectors that can be calculated also changes for each pitch. Therefore, for example, the feature amount calculation means 41 calculates a predetermined number of motion vectors of a, b in the middle, and c immediately before catching from throwing, and calculates (a + b + c) motion vectors in total. Also good.

Note that the feature amount calculating means 41 may calculate the distance from the image coordinates of the ball with respect to the approximate straight line of the trajectory instead of the motion vector. For example, as shown in FIG. 6, the feature amount calculation means 41, from the approximate straight line L ₁ with respect to the trajectory, it is also possible to calculate the distance D to the center of gravity of the ball M. Furthermore, in order to align the number of distances for each pitch, a predetermined number of distances may be calculated immediately before throwing out, in the middle, and immediately before catching, as in the case of the motion vector. FIG. 6 is a schematic diagram schematically showing a pitching trajectory and a distance calculated by the feature amount calculating means.

  The learning unit 42 learns a discriminator for identifying the sphere type based on the feature amount from the feature amount calculated by the feature amount calculation unit 41 and the sphere type data of the locus stored in the locus data storage device 3. To do. Here, the classifier generated by machine learning is output to the learning result storage device 5 as a learning result.

  Here, it is assumed that the learning means 42 learns by a supervised random forest (Leo Breiman, Random Forests, Machine Learning, 45, 5-32, 2001). Hereinafter, with reference to FIG. 7 (refer to FIG. 1 as appropriate), a method in which the learning unit 42 learns with a random forest as supervised learning will be described. FIG. 7 is an explanatory diagram for explaining a method in which the learning means learns with a random forest, (a) is a schematic diagram schematically showing an outline of processing of the random forest, and (b) is a schematic diagram of a decision tree. It is a schematic diagram shown in FIG.

  As shown in FIG. 7 (a), the learning means 42 allows a predetermined number of random data (in FIG. 7 (a)) by allowing duplication from various trajectory data t, t,... Stored in the trajectory data storage device 3. (4 pieces) of trajectory data t, t,... (Bootstrap sample) are extracted to generate n nodes T, T,..., And trajectory data included in each node T calculated by the feature amount calculation means 41. n decision trees are generated based on the feature quantities t, t,. The learning unit 42 generates nodes T, T,... By randomly extracting the locus data t, t,... From the locus data t, t,. , Deviation of the trajectory data t, t,... Included in each node T can be prevented.

At this time, the learning means 42 branches the trajectory data t, t,... Of each node T (or subset T ₁ , T ₂ ,...) Into two using the Gini index. Here, learning means 42, locus data t by using the Gini index, t, ... to the method branches to two, the trajectory data t of the node T, t, ... into two subsets _T 1, _{T 2} An example of branching will be described.

First, the learning means 42 randomly selects a predetermined number of feature quantities that are candidates for feature quantities to be used for branching from the feature quantities calculated by the feature quantity calculation means 41. For example, when there are 37 types of feature amounts V _{1 to} V ₃₇ calculated by the feature amount calculation unit 41, the learning unit 42 selects four feature amounts V ₁₀ , V ₂₁ , V ₂ , and V ₂₄ at random. Suppose that Then, the learning means 42 has a feature amount (V ₁₀ , V ₂₁ , V ₂ or V ₂₄ ) for branching the trajectory data t, t,... Into _two subsets T ₁ , T ₂ , and the feature amount ( The threshold value x of V ₁₀ , V ₂₁ , V ₂ or V ₂₄ ) is determined based on the Gini index.

Here, the Gini index gini (T) of a certain node T before branching is obtained by the following equation (7). Note that n is the number of classes (sphere types), and p _j is the relative frequency at the node T of the trajectory data t of the class j (sphere type j).

The M pieces of trajectory data t, t,... Included in the node T are a subset T _{1 of} M ₁ pieces of trajectory data t, t,..., And M ₂ pieces of trajectory data t, t,. Gini index _{gini split} node T after being branched into a set _{T 2} (T) is represented by the following equation (8).

The Gini index gini _split (T) is such that the distribution of the relative frequency of each class is biased, that is, the trajectory data t, t,... Of a specific sphere type is biased in each of the subsets T ₁ and T _2. The smaller the value, the smaller the value. For this reason, the learning unit 42 determines, for each feature quantity V ₁₀ , V ₂₁ , V ₂ , V ₂₄ , from the maximum value to the minimum value of each feature quantity (V ₁₀ , V ₂₁ , V ₂ or V ₂₄ ). The threshold value of one feature quantity (V ₁₀ , V ₂₁ , V _2, or V ₂₄ ) that calculates the Gini index gini _split (T) by changing the threshold value x in the range and gives the minimum Gini index gini _split (T) x is determined. Similarly, the learning means 42 is also a set of trajectory data t, t,... For each of the subsets T ₁ and T _2. Branches are repeated until a decision tree is generated.

  In this way, in the random forest, the type of feature quantity to be used is selected at random every time it branches, so that n decision trees that are not similar to each other and have many branches are generated. The learning unit 42 stores n decision trees as learning results and stores them in the learning result storage device 5.

  Here, although the learning means 42 has explained the method of learning by a random forest, it may be learned by other supervised learning such as bagging (Bagging; Bootstrap aggregating), support vector machine, or Adaboost.

  A support vector machine is an algorithm that linearly separates data in a high-dimensional space using a kernel function. For example, when the learning unit 42 learns using a support vector machine, the margin between two types of classes is maximized, and a boundary (hyperplane) passing through the middle of the margin is obtained to separate the data.

  Further, when the learning means 42 learns by bagging, for example, it can be performed as follows. That is, the learning means 42 restores and extracts n data sets consisting of a predetermined number of trajectory data at random from the various trajectory data stored in the trajectory data storage device 3 by the booth trap method, Using each data set as a training sample, n discriminators are generated based on the feature amount of the trajectory data included in each data set calculated by the feature amount calculating means 41. The learning means 42 stores this discriminator in the learning result storage device 5. Then, the discriminating means 62 of the ball type discriminating device 6 to be described later discriminates using each discriminator stored in the learning result storage device 5, and the final discriminating result is obtained by the average or majority of the discriminating results of each discriminator. It is good to do.

  Furthermore, the learning means 42 may learn by unsupervised learning. Hereinafter, a case where the learning unit 42 learns using principal component analysis as an example of unsupervised learning will be described with reference to FIG. 8 (refer to FIG. 1 as appropriate). Note that the learning means 42 is not limited to principal component analysis, and can be learned by other unsupervised learning such as the K-means method. FIG. 8 is a scatter diagram showing the distribution of principal component feature points obtained by principal component analysis of the feature value of the learning means. FIG. 8 shows an example in which the learning unit 42 obtains two principal component scores for the first principal component and the second principal component.

  Here, the feature quantity obtained by the feature quantity calculation means 41 is high-order, and the processing load for identification is large. Therefore, the learning unit 42 performs principal component analysis on the feature amount input from the feature amount calculation unit 41 and compresses the dimension. At this time, the learning means 42 obtains eigenvalues and eigenvectors of each principal component, and selects a predetermined number of principal components having high eigenvalues. Then, the learning means 42 calculates the inner product of the feature amount and the eigenvector of each selected principal component to obtain a principal component score.

  As shown in FIG. 8, the main component scores tend to be distributed for each sphere type. For example, in FIG. 8, the principal component score obtained from the feature quantity of the trajectory of the straight sphere type is indicated by a circle, the principal component score of a curve is indicated by a square mark, and the principal component score of a fork is indicated by a triangle mark. Then, in the scatter diagram of FIG. 8, straight principal component scores are distributed in region R1, curved principal component scores are distributed in region R2, and fork principal component scores are distributed in region R3. .

  The learning means 42 stores this principal component score distribution in the learning result storage device 5. Then, a principal component score is also calculated for the feature quantity input from the feature quantity calculation means 61 by the discrimination means 62 of the ball type identification apparatus 6 described later, and the coordinates of the principal component score are stored in the learning result storage device 5. It is possible to identify the sphere type by determining which sphere type of the distribution is closest. For example, when the principal component score r shown in FIG. 8 is calculated by the discriminating means 62, the principal component score r is closest to the curve, so the discriminating means 62 determines the trajectory as a curve.

  Further, the learning means 42 associates with the respective trajectory data, the type of grasshopper in the pitch (right or left pitcher pitch), pitcher type (right pitch or left pitcher pitcher), pitch By inputting information indicating the type (underslow or overslow) of the direction from the outside, it is possible to generate a discriminator separately for each pitch that matches the grasshopper type, pitcher type and throw type. preferable. At this time, the discriminating means 62 of the ball type discriminating device 6 described later also inputs information on the type of grasshopper, pitcher type and throwing type in the pitch to be discriminated from the outside, and uses the corresponding discriminator. It will be identified.

(Configuration of ball type identification device)
Further, the configuration of the ball type identification device 6 will be described with reference to FIG. The ball type identification device 6 includes a feature amount calculation unit 61 and an identification unit 62.

  The feature amount calculating means 61 calculates a predetermined feature amount indicating at least one of the shape of the trajectory of the pitch, the relative position of the pitch, and the ball speed based on the trajectory data input from the pitch trajectory analysis device 2b. It is. The feature amount calculated here is output to the identification unit 62. The feature quantity calculation means 61 differs from the feature quantity calculation means 41 of the ball type learning device 4 only in the input source of the trajectory data and the output destination of the feature quantity. Are the same.

  The identifying unit 62 identifies the pitch type of the pitch from the feature amount input from the feature amount calculating unit 61 based on the learning result stored in the learning result storage device 5. The identification result obtained here is output to the outside.

  Here, an identification method of the identification unit 62 when the learning unit 42 of the ball type learning device 4 performs machine learning using a random forest will be described. First, the identification unit 62 reads n decision trees stored in the learning result storage device 5, and for each decision tree, based on the feature amount input from the feature amount calculation unit 61, the trajectory data of the pitch Is included in which subset of the lowest level of the decision tree. Here, the learning means 42 generates the decision tree by repeating the branch until all the subsets that are branched and generated in the lowermost layer become a set of trajectory data of one kind of sphere type. Each subset is associated with one type of sphere.

  Therefore, the identification unit 62 uses the sphere types associated with the lowermost subset of each decision tree including the trajectory data as the identification result, and integrates these n identification results by majority vote to obtain a final result. It was decided that it would be output as a correct identification result. Further, here, the ball types are classified into three systems, that is, a straight sphere, a changing sphere that bends, and a falling sphere that falls, based on the characteristics, and the identification means 62 determines the system by majority decision from the identification result of each decision tree. It was decided to output the result of identification together with the information of the ball type. Here, straights and chutes are classified into straight balls, curves, sliders, changing spheres that turn cut balls, forks, sinkers, and changing spheres that fall down change-ups.

  In this example, the identification unit 62 never sets a large number of identification results of each decision tree as a final identification result. However, the ratio of the identification results of each decision tree (for example, straight 70%, curve 30%). ) May be the final identification result. Furthermore, since the ball type identifying device 6 includes a trajectory image generating means (not shown), the extracted pitch position is determined based on the trajectory data input from the pitch trajectory analyzing device 2b, for example, the identification result of each decision tree. A trajectory image indicated by a color circle in which colors (for example, straight is red and curve is yellow) that are pre-assigned to each ball type according to the ratio of the sphere type may be generated and output as an identification result to the outside. Good.

  The configuration of the ball type determination system S has been described above, but the present invention is not limited to this. For example, here, the case where the baseball ball type is determined has been described as an example. However, the discriminator generating device and the throwing ball type discriminating device of the present invention are not limited to baseball, but include softball, soccer, golf, etc. It can be applied to various ball games. When applied to a soccer player, for example, it is possible to determine the degree of bending of a free kick, and when applied to golf, it can be used to determine, for example, a slice or a hook. In addition, it can be incorporated into an attraction or the like that determines the type of pitch of his / her pitch at an amusement facility or the like.

  Here, the pitching trajectory analysis device 2b analyzes the video stored in the video storage device 1b. However, for example, the video captured in real time may be input and analyzed. At this time, the ball type determined by the ball type determination system S on the broadcast station side can be displayed immediately in the relay program. Further, the configuration of the pitching trajectory analysis devices 2a and 2b is not limited to that shown in FIG. 2, and the pitching trajectory analysis devices 2a and 2b may be anything that can analyze the pitching trajectory from the input video of the pitching.

  Furthermore, by adding the information on the sphere type obtained by the sphere type determination system S to the video as metadata, it is possible to automatically generate metadata for video search. Furthermore, the obtained information on the type of sphere can be used as information distributed via data broadcasting or the Internet.

  The discriminator generating device A and the pitched ball type discriminating device B can also realize each means as a function program in a computer, and the function program is combined to identify the discriminator generating program and the pitched ball type. It is also possible to operate as a program.

[Operation of the ball type determination system]
Next, with reference to FIGS. 9 to 13 (refer to FIG. 1 as appropriate), the operation of the ball type determination system S including the discriminator generation device A and the pitching ball type identification device B in the present invention will be described. FIG. 9 is a flowchart showing the discriminator generating operation of the discriminator generating apparatus according to the present invention. FIG. 10 is a flowchart showing the trajectory data generation operation of the pitching trajectory analysis apparatus. FIG. 11 is an explanatory diagram for explaining a conditional expression of a difference image, where (a) shows three continuous field images, and (b) shows two difference images. FIG. 12 is an explanatory diagram for explaining the position prediction process. FIG. 13 is a flowchart showing the pitching ball type identifying operation of the pitching ball type identifying device according to the present invention.

(Classifier generation operation)
First, with reference to FIG. 9, the discriminator generation operation in which the ball type determination system S generates a discriminator for identifying the ball type will be described. The ball type determination system S inputs a video of one pitch stored in the video storage device 1a by the pitch trajectory analysis device 2a, and generates trajectory data by a trajectory data generation operation described later (step S11). Then, the ball type determining system S associates the ball type data input from the outside with the track data generated in step S11 and stores it in the track data storage device 3 (step S12).

  Further, the ball type determination system S determines whether or not the pitch trajectory analysis device 2a has finished generating the trajectory data for all the videos stored in the video storage device 1a (step S13). If not completed (No in step S13), the ball type determination system S returns to step S11, and the pitching trajectory analysis device 2a performs trajectory data on the next video stored in the video storage device 1a. The operation after the operation of generating the trajectory data by the generation operation is performed.

  On the other hand, when the processing is completed (Yes in step S13), the ball type determination system S inputs the trajectory data stored in the trajectory data storage device 3 in step S12 by the feature amount calculation unit 41 of the ball type learning device 4. Then, a feature amount is calculated for each trajectory data (step S14).

  Subsequently, the ball type determination system S uses the learning unit 42 of the ball type learning device 4 to convert the feature value of the trajectory data calculated in step S14 and the ball type data associated with the trajectory data in step S12. Based on the feature amount of the trajectory data, a discriminator for identifying the sphere type is generated by machine learning, stored in the learning result storage device 5 (step S15), and the operation ends.

(Track data generation operation)
Next, referring to FIGS. 10 to 12 (refer to FIGS. 1 and 2 as appropriate), the pitching trajectory analysis device 2a (2b) of the ball type determination system S extracts the video object of the pitching from the video, and the trajectory data. The trajectory data generation operation (step S11 in FIG. 9) for generating the will be described.

  First, a search area is input by an input means such as a mouse or a keyboard (not shown), and the pitching trajectory analysis apparatus 2a (2b) uses a field image in the field image constituting the video input by the search area setting means 25. A search area for searching for a video object is set in a part of the range (step S21). It should be noted that the position of this search area is appropriately updated following the movement of the ball in step S32 described later.

  Subsequently, the pitching trajectory analysis device 2a (2b) uses the current input image stored in the image storage unit 211 by the difference image generation unit 213 of the object candidate image generation unit 21 for the search region set in step S21. The difference image 1 with the luminance difference from the future input image (latest input image) as a pixel value is generated, and the current input image stored in the image storage unit 211 by the difference image generation unit 214 Then, the difference image 2 is generated with the luminance difference from the past input image stored in the image storage unit 212 as a pixel value (step S22). Here, three consecutive input images (past, present, and future) are input to the object candidate image generating means 21, and the current field image is stored in the image storage unit 211, and the past field image is stored in the image storage unit. 212.

Subsequently, in the pitch trajectory analysis device 2a (2b), the candidate image generation unit 215 of the object candidate image generation unit 21 generates an object candidate image from the difference image 1 and the difference image 2 generated in step S22 (step S22). S23). Here, the candidate image generation unit 215 is a region where the luminance value Image1 (x, y) of the difference image 1 and the luminance value Image2 (x, y) of the difference image 2 satisfy the following expressions (9) and (10). Therefore, this region is determined as a candidate for the object. Note that (x, y) indicates pixel coordinates in the difference image 1 and the difference image 2, and T is a luminance threshold value.
Image1 (x, y)> T (9)
Image2 (x, y) <-T (10)

  In other words, in this embodiment, the video object to be extracted (tracked) is a ball, and the luminance of the ball is generally higher than the luminance of the background image, so that the equations (9) and (10) are satisfied simultaneously. Is the conditional expression at this time. The reason why the video object (ball) moving at high speed can be extracted when the difference image 1 and the difference image 2 satisfy the expressions (9) and (10) as described above with reference to FIG. I will explain.

  As shown in FIG. 11A, three consecutive field images G1, G2, and G3 representing the past, present, and future have one ball (M1, M2) at the past, present, and future positions. , M3). The balls M1, M2, and M3 are moving at a high speed that does not overlap the field images, and the brightness of the balls is higher than the brightness of the background image. Further, as shown in FIG. 11B, the difference image 1 (G′1) is generated using a difference obtained by subtracting the luminance of the field image G3 from the luminance of the field image G2 as a pixel value. Further, the difference image 2 (G′2) is generated by using a difference obtained by subtracting the luminance of the field image G2 from the luminance of the field image G1 as a pixel value.

  The region M′2 in the difference image G′1 is configured with a luminance difference value obtained by subtracting the luminance of the region where no object exists from the luminance of the ball M2 as a pixel value, and therefore the pixel value has a plus sign (for example, white) It becomes. Further, the region M′3 in the difference image G′1 is configured with a luminance difference value obtained by subtracting the luminance of the ball M3 from the luminance of the region where the object does not exist as a pixel value. Black). Furthermore, the other areas including the area M′1 in the difference image G′1 are configured with the luminance difference value obtained by subtracting the luminances of the areas where no object is present as the pixel value, and thus the pixel value is 0.

  On the other hand, the region M′1 in the difference image G′2 is configured with a luminance difference value obtained by subtracting the luminance of the region where no object is present from the luminance of the ball M1 as a pixel value, and therefore the pixel value has a plus sign (for example, White). In addition, since the area M′2 in the difference image G′2 is configured with a luminance difference value obtained by subtracting the luminance of the ball M2 from the luminance of the area where no object exists, the pixel value has a minus sign (for example, Black). Further, since the other areas including the area M′3 in the difference image G′2 are configured with the luminance difference value obtained by subtracting the luminances of the areas where no object exists as the pixel value, the pixel value is 0.

  When the difference image G′1 and the difference image G′2 are compared, the sign of the pixel value in the difference image G′1 is positive (equation (9)), and the sign of the pixel value in the difference image G′2 is The area that is negative (equation (10)) represents the area M′2, that is, the ball M2 (current position). Therefore, the candidate image generation unit 215 can extract a video object (ball) moving at high speed as an object candidate image from the equations (9) and (10).

Note that when the luminance of the target video object is lower than the luminance of the background image, it is a condition that Expression (11) and Expression (12) are satisfied at the same time.
Image1 (x, y) <-T (11)
Image2 (x, y)> T (12)

  In this case, the region where the sign of the pixel value is negative in the difference image G′1 (formula (11)) and the sign of the pixel value is positive in the difference image G′2 (formula (12)) is the video object. Extracted as

  The description of the locus data generation operation in FIG. 10 will be continued. The pitching trajectory analysis device 2a (2b) assigns a number (label) to a region that is a candidate for a video object in the object candidate image generated in step S23 by the labeling unit 221 of the ball selection unit 22 ( Step S24). The subsequent operations are processed in units of video objects based on the numbers assigned to the video object candidates.

  Then, the pitch trajectory analyzing apparatus 2a (2b) stores in the extraction condition storage unit 23 from the object candidate images for each video object numbered in step S24 by the feature amount analysis unit 222 of the ball selection unit 22. Based on the extracted extraction condition and the like, a video object to be extracted (tracked) is selected, and the position and feature amount of the selected video object are analyzed (calculated) (step S25). Here, as the position of the video object, for example, the barycentric coordinates of the video object are used. As the feature amount of the video object, the area, luminance, color, circularity, etc. of the video object are used.

  Further, the pitching trajectory analysis device 2a (2b) selects a video object from the object candidate images based on the number (label) by the filter processing unit 223 of the ball selection unit 22 and performs a filtering process (step) S26). That is, the filter processing unit 223 determines whether or not the “position” of the selected video object matches the range predicted by the position prediction unit 24. Here, when the matching condition based on “position” is met, the filter processing unit 223 determines whether the “area” of the selected video object matches the extraction condition stored in the extraction condition storage unit 23. To do. Here, when the matching condition based on the “area” is met, the filter processing unit 223 determines whether the “color” of the video object matches the extraction condition. When the matching condition based on “color” is met, the filter processing unit 223 determines whether the “luminance” of the video object matches the extraction condition. If the matching condition based on “luminance” is met, the filter processing unit 223 determines whether the “circularity” of the video object matches the extraction condition.

  Subsequently, the pitching trajectory analyzing apparatus 2a (2b) causes the object selecting unit 224 of the ball selecting unit 22 to have no video object candidates that meet all the extraction conditions in step S26, that is, video object candidates to be extracted. It is determined whether or not extraction has failed (step S27).

  If the extraction has not failed (No in step S27), the pitching trajectory analysis device 2a (2b) is adapted to all the extraction conditions in step S26 by the object selection unit 224 of the ball selection unit 22. From the video object candidates, the one closest to the position coordinates of the video object extracted from the previous frame image is selected as the video object to be tracked (step S28). As a result, even when there are a plurality of video object candidates that have passed through all the filters, the object selection unit 224 can select one video object. Note that the position information of the video object selected here is stored in the extraction condition storage means 23.

  Further, the pitch trajectory analyzing apparatus 2a (2b) uses the pitch interpolating means 26 to detect the position information of the video object selected in step S28 when there is a field image that failed to be extracted before and after the previous extraction. Based on the position information of the video object of the past field image stored in the extraction condition storage means 23, the position information of the video object of the field image that failed to be extracted is obtained by interpolation (step S29), and step S31. Proceed to

  On the other hand, when the extraction fails in step S27 (Yes in step S27), the pitching trajectory analysis device 2a (2b) determines whether the object selection unit 224 has failed to extract continuously in a predetermined number of frames or more. Is determined (step S30). If the extraction fails continuously for a predetermined number of frames or more (Yes in step S30), the operation ends. On the other hand, if the extraction has not failed continuously for a predetermined number of frames or more (No in step S30), the process proceeds to step S31.

  Then, the pitch trajectory analysis device 2a (2b) outputs the position information of the video object selected in step S28 or the position information obtained in step S30 to the trajectory data storage device 3 by the pitch interpolation means 26 ( Step S31). Further, the pitching trajectory analysis device 2a (2b) estimates the search area of the ball in the field image to be input next by the position predicting unit 24 based on the position of the video object selected in step S28 (step S32). ), The process returns to step S21.

  Here, the pitching trajectory analysis device 2a (2b) can estimate the search area by the position predicting means 24 using the least square method and the Kalman filter. As shown in FIG. 12, balls M1 to M4 at different times are obtained by past object extraction processing, and the position coordinates of the ball M5 (double circle in the figure) in the current field image are extracted. Assuming that

  Then, the position predicting unit 24 is indicated by the arrow V1 by the linear predicting unit 241 by the extracted position coordinates (ball M5) in the current field and the extracted position coordinates (ball trajectory M4) in the previous field (previous frame). Find the motion vector. Then, based on the motion vector V1, the linear prediction unit 241 adds a motion vector V2 similar to the motion vector V1 to the position of the ball M5 with respect to the extracted position coordinates (ball M5) in the current field. As a predicted position of the ball, a linear predicted position coordinate Y1 (triangle mark in the figure) is temporarily determined (linear prediction).

  On the other hand, the position predicting unit 24 uses the curve predicting unit 242 to generate a quadratic curve K1 indicated by a one-dot chain line that passes through the ball trajectory (balls M1 to M5) from the past to the present using the least square method. Ask. Then, the position predicting unit 24 corrects the linear predicted position coordinate Y1 on the quadratic curve K1, and obtains a curve predicted position coordinate Y2 indicated by a square mark (curve prediction). The obtained curve predicted position coordinate Y2 is set as a search position in the next field image (next frame).

However, the position predicting unit 24 performs linear prediction by the linear prediction unit 241 and curve prediction by the curve prediction unit 242 based on the number of ball position coordinates or a quadratic curve equation extracted by the switching unit 243 in the past. Switch. That is, the position predicting unit 24 does not perform the curve prediction when there is not a sufficient number of past ball position coordinates (three but five or more are preferable) to execute the curve prediction. Perform linear prediction only. Further, the position predicting unit 24 changes the sign of the coefficients a and b of the quadratic curve (y = ax ² + bx + c) used for the curve prediction when the curve prediction unit 242 executes the curve prediction, Even when the values of a and b exceed a predetermined value, switching to linear prediction is performed. In this way, the pitching trajectory analysis device 2a (2b) extracts the ball while appropriately updating the area where the ball is searched, and therefore, it is possible to reduce the amount of calculation for extracting and updating the ball.

  Through the above operation, the pitching trajectory analysis device 2a (2b) sequentially extracts and tracks video objects (balls) to be tracked from field images input in time series as video, and generates video object trajectory data. be able to.

(Classifier generation operation)
Further, with reference to FIG. 13, the pitch type identifying operation for identifying the pitch type of the pitch using the classifier generated by the classifier generating operation shown in FIG. 11 will be described. The ball type determination system S inputs the video of one pitch stored in the video storage device 1b by the pitch trajectory analysis device 2b, and generates trajectory data by the trajectory data generation operation shown in FIG. 10 (step S41). . Then, the ball type determination system S calculates the feature amount of the trajectory data generated in step S41 by the feature amount calculation unit 61 of the ball type identification device 6 (step S42).

  Subsequently, the ball type determination system S uses the discriminator generated by the discriminator generating operation shown in FIG. 9 by the discriminating means 62 of the ball type discriminating device 6 and stored in the learning result storage device 5, step S 42. Based on the feature amount of the trajectory data calculated in step S43, the pitch type of the trajectory data is identified (step S43).

  Furthermore, the ball type determination system S determines whether or not the pitch trajectory analysis device 2b has finished generating the trajectory data for all the videos stored in the video storage device 1b (step S44). If not completed (No in step S44), the ball type determination system S returns to step S41, and the pitching trajectory analysis device 2b performs trajectory data for the next video stored in the video storage device 1b. The operation after the operation of generating the trajectory data by the generation operation is performed. On the other hand, when the operation is completed (Yes in step S44), the operation is ended.

It is the block diagram which showed the structure of the ball type determination system provided with the discriminator production | generation apparatus and throwing ball type identification device of this invention. It is the block diagram which showed the structure of the pitch trajectory analysis apparatus of a ball | bowl type determination system provided with the discriminator production | generation apparatus and pitching ball type identification device of this invention. It is the schematic diagram which showed typically the locus | trajectory of a pitch and the approximate straight line and the approximate curve which are produced | generated by the feature-value calculation means of the discriminator production | generation apparatus of this invention, and a throwing ball type identification apparatus. It is the schematic diagram which showed typically the image from which the image of the ball | bowl was extracted finally, and the catching position produced | generated by the feature-value calculation means of the discriminator production | generation apparatus of this invention, and a throwing ball type identification apparatus. It is the schematic diagram which showed typically the locus | trajectory of a pitch, and the motion vector calculated by the feature-value calculation means of the discriminator production | generation apparatus of this invention, and a pitching ball type identification device. It is the schematic diagram which showed typically the locus | trajectory of a pitch, and the distance calculated by the feature-value calculation means of the discriminator production | generation apparatus of this invention, and a pitching ball type identification device. Explanatory drawing for demonstrating the method in which the learning means of the discriminator production | generation apparatus of this invention learns by a random forest, (a) is a schematic diagram which shows typically the outline | summary of the process of a random forest, (b) is decision It is a schematic diagram which shows a tree typically. It is a scatter diagram which shows distribution of the principal component special point obtained by the principal component analysis of the feature-value of the learning means of the discriminator generation device of this invention. It is the flowchart which showed the discriminator production | generation operation | movement of the discriminator production | generation apparatus of this invention. It is the flowchart which showed the locus | trajectory data generation operation | movement of a pitching locus | trajectory analyzer. It is explanatory drawing for demonstrating the conditional expression of a difference image, (a) is three continuous field images, (b) is two difference images. It is explanatory drawing for demonstrating a position prediction process. It is the flowchart which showed the pitching type identification operation | movement of the pitching type identification device of this invention.

Explanation of symbols

S Ball type determination system A Ball type learning device B Throwing ball type identification device 2a Throwing trajectory analysis device (throwing trajectory analysis means)
2b Throwing trajectory analysis device (throwing trajectory analysis means)
4 Ball type learning device (ball type learning means)
41 feature quantity calculation means 42 learning means 5 learning result storage device (discriminator storage means, discriminator storage device)
6 Ball type identification device (ball type identification means)
61 feature amount calculation means 62 identification means

Claims (6)

A pitching type identifying device for identifying the pitch type of the pitch from the pitch video,
The trajectory data, which is the time-series information of the position of the video object of the pitch in the time-series image constituting the video of the pitch, and the information on the type of the pitch of the pitch previously associated with the pitch Discriminator storage means for storing in advance a discriminator for discriminating the ball type from the trajectory data, machine-learned for a plurality of pitches
A pitching trajectory analysis means for inputting a video of a pitch to be identified as a pitch type, extracting a video object of the pitch in a time-series image constituting the video, and generating trajectory data of the pitch;
Using a discriminator stored in the discriminator storage means, a ball type identifying means for identifying the ball type of the pitch to be identified from the trajectory data generated by the pitch trajectory analyzing means;
A pitching ball type identifying device comprising: A feature amount indicating at least one of the shape of the trajectory of the pitch, the relative position of the pitch, and the ball speed calculated by the discriminator storage means normalized by a predetermined reference from the trajectory data of the plurality of pitches A pre-stored discriminator for identifying the ball type from the feature amount, machine-learned based on the pitch type information of the pitch previously associated with each pitch,
The ball type identifying means is
Feature amount calculating means for calculating the feature amount normalized by the predetermined reference from the trajectory data generated by the pitch trajectory analyzing means;
Using a discriminator stored in the discriminator storage unit, an identifying unit for identifying a pitch type of the pitch to be identified from the feature amount calculated by the feature amount calculating unit;
The pitching ball type identifying device according to claim 1, wherein: A discriminator generating device for generating a discriminator for identifying a pitch type of a pitch used in the pitched ball type identifying device according to claim 1,
This is time series information of the position of the pitch video object in the image by inputting a plurality of pitch videos and extracting the pitch video object in the time series image constituting each video. Pitching trajectory analysis means for generating trajectory data;
Based on the trajectory data generated by the pitch trajectory analyzing means and information on the pitch type of the pitch previously associated with each pitch, a discriminator for identifying the ball type from the trajectory data is provided. A ball type learning means generated by learning;
A discriminator generation device comprising: A discriminator for discriminating a ball type from a feature amount indicating at least one feature of a shape of a pitch of a pitch, a relative position of the pitch, and a ball speed, which is used in the pitch type identification device according to claim 2. A discriminator generating device for generating,
This is time series information of the position of the pitch video object in the image by inputting a plurality of pitch videos and extracting the pitch video object in the time series image constituting each video. Pitching trajectory analysis means for generating trajectory data;
Based on the trajectory data generated by the pitch trajectory analyzing means and information on the pitch type of the pitch previously associated with each pitch, a discriminator for identifying the ball type from the trajectory data is provided. A ball type learning means generated by learning;
With
The ball type learning means
Feature amount calculating means for calculating the feature amount normalized by a predetermined standard for each of the pitches from the trajectory data generated by the pitch trajectory analyzing means;
Learning that generates, by machine learning, a discriminator for identifying the ball type from the feature value based on the feature value of the plurality of pitches calculated by the feature value calculating means and the information on the corresponding ball type Means,
A discriminator generation device comprising: The trajectory data, which is the time-series information of the position of the video object of the pitch in the time-series image constituting the video of the pitch, and the information on the type of the pitch of the pitch previously associated with the pitch The pitch type of the pitch is identified from the video of the pitch using a discriminator for discriminating the ball type from the trajectory data, which is machine-learned based on a plurality of pitches and stored in advance in the discriminator storage device Computer for
A pitching trajectory analysis means for inputting a video of a pitch to be identified as a pitch type, extracting a video object of the pitch in a time-series image constituting the video, and generating trajectory data of the pitch,
Using a discriminator stored in the discriminator storage device, a ball type identifying unit for identifying the ball type of the pitch to be identified from the trajectory data generated by the pitch trajectory analyzing unit;
Throwing ball type identification program characterized by functioning as A computer for generating a discriminator for identifying a pitch type of a pitch used in the pitch type discriminating apparatus according to claim 1.
This is time series information of the position of the pitch video object in the image by inputting a plurality of pitch videos and extracting the pitch video object in the time series image constituting each video. Pitching trajectory analysis means for generating trajectory data;
Based on the trajectory data generated by the pitch trajectory analyzing means and information on the pitch type of the pitch previously associated with each pitch, a discriminator for identifying the ball type from the trajectory data is provided. Sphere learning means generated by learning,
A discriminator generation program characterized by functioning as:

Images