JP2019175063A - Object identification device - Google Patents

Abstract

To accurately identify a target object even when a plurality of target objects photographed by a camera are overlapped and even when a plurality of jockeys wearing the same color hat or the same color cloth.SOLUTION: An object identification device comprises an object identification unit for identifying a target object for each identification element, for eventually identifying the object by aggregating identification results for each identification element of the target object. The object identification device determines a weight to aggregate the identification results for each identification element according to identification scores of a plurality of candidate objects.SELECTED DRAWING: Figure 1

Description

  The present invention relates to an object identification device that captures an image including a moving object with a camera and identifies the object in the image based on the captured image.

  Conventionally, there has been proposed a racehorse position information analysis device that estimates and identifies the position of a race target in a race based on only images from a camera without using a sensor (see, for example, Patent Document 1). .

International Publication No. 2007/037350 Pamphlet

  The user can know the position of each race target in the race by specifying the position of the target using such a device, plotting the specified result on a map, etc., and distributing the result to a terminal such as a smartphone. In addition, in competitions with gambling elements such as horse racing and bicycle racing, many users want to analyze the development of past races and the movement of the race target, and the analysis and display results are used for analysis such as race development. It can also be useful. Among such technologies, especially the technology that uses only cameras to perform position estimation and target identification does not require the installation of sensors to the race target or the establishment of a communication environment. Since it is simple, the merit of performing position estimation and target identification using only a camera is great.

  Patent Document 1 relates to an invention that analyzes and displays the position and movement trajectory of a specific racehorse in a race by analyzing a video from a camera provided around the racetrack and tracking the racehorse. In order to realize position estimation and identification, the movement of each horse is tracked by measuring the similarity to the target object of the previous frame using a hexagonal template that matches the size of each racehorse. Regarding the measurement of the similarity, the similarity is determined by analyzing the color of the template part. When determining the similarity based on the color, a jockey part or a horse body whose features are relatively easy to appear. It is characterized by identifying the racehorse by weighting the result of color analysis of the part.

  As described in Patent Document 1, tracking using a template for estimating the position is very effective when the object is moving away from other objects. They are not separated and projected on the camera, but often appear in overlapping situations when parallel running or overtaking occurs. In such a situation where the objects overlap with each other, the accuracy of similarity measurement decreases due to the fact that some of the objects become invisible.

  Since the method of Patent Document 1 identifies a racehorse by clarifying the track (identifying the number of the racehorse), there is a problem that the identification accuracy at that time naturally decreases. Furthermore, in horse racing, the clothes of jockeys are determined for each owner, and in addition to the possibility that the horse of the same owner runs in the same race, there are many clothes of similar colors even if the owners are different. Also, the color of the hat is determined based on the frame order. When there are a large number of horses in one race, there are cases where two or three horses of the same color run.

  Although there are some conspicuous colors such as white hair and eyelashes with respect to the horse body, racehorses of similar colors often run, and identification based on the similarity of colors makes it possible to see the target object completely. In some cases, erroneous recognition is unavoidable. In particular, the analysis based on color information also has a problem that it tends to be influenced by environmental differences such as differences in color appearance due to differences in weather on the day.

  In the present invention, in order to identify a moving object, when the result of a plurality of identification elements is combined to determine the identification result, the weight for each identification element is set based on the characteristics of the candidate object. An object of the present invention is to realize highly accurate target identification by dynamically changing each candidate object.

  (1) An object identification device according to the present invention includes an object part extraction unit that extracts a part used for identification of a target object from an image, an object position estimation unit that estimates the position of the target object, and each identification element for the target object. The object identification unit that finally identifies the object by counting the identification results for each identification element of the target object, and each candidate object used by the object part extraction unit and the object identification unit An object feature learning unit that learns the features of a part in advance, an identification parameter determination unit that dynamically determines a weight when the object identification unit aggregates identification results for each identification element of the target object, and the object identification Display the final object identification result It has a result display unit.

  (2) In the object identification device according to (1), the identification parameter determination unit may weight the identification elements that can easily identify candidate objects among the identification elements that have been successfully extracted.

  (3) In the object identification device according to (1) or (2), in order for the identification parameter determination unit to heavily weight elements that can easily identify candidate objects, the identification elements of all candidate objects By combining three parameters: the number of clusters when features are clustered, the variance of the representative value of each cluster, and the element reliability of the extracted elements, the identification elements that make it easy to identify candidate objects are greatly weighted. May be.

  (4) In the object identification device according to any one of (1) to (3), when the object identification unit calculates a final identification score, a candidate object that indicates a score that satisfies a certain condition is The candidate objects may be narrowed down while the weighting is repeatedly changed by the identification parameter determination unit.

  (5) In the object identification device according to any one of (1) to (4), when the number of candidate objects is large when the identification parameter determination unit sets the weight for object identification. An adjustment may be made so that the difference between the weights becomes small.

  (6) In the object identification device according to any one of (1) to (5), the object part extraction unit extracts a plurality of parts for identification, and the object identification unit succeeds in extraction. By identifying the objects in order from the target object having the largest number of parts, the number of candidate objects may be decreased in order from the target object that is easy to identify, and the identification accuracy may be increased.

  (7) In the object identification device according to any one of (1) to (6), when the object position estimation unit estimates the position, the target objects exist in the vicinity, and the target in the back When the object is partially obscured by the target object in front, by detecting the part that is not obscured and estimating the number and position of the target object based on the detection of multiple parts, Even when there is an overlap, the position estimation may be performed accurately.

  (8) The object identification device according to any one of (1) to (7), wherein the object identification device identifies an object participating in a race, and the identification parameter determination is performed. The object feature learning unit includes not only the static information registered in the prior information database before the race but also the target before the race. The state of the object may be captured as a video, the part of the object may be extracted from the video, and real-time feature information immediately before the race may be learned.

  (9) In the object identification device according to (8), the part of the candidate object is extracted from the video immediately before the race as the real-time characteristic information immediately before the race learned by the object feature learning unit. The color threshold value may be adjusted so that the circumscribed rectangle size of the region of the part extracted when the color range is dynamically changed approaches the size of the part set before the race.

  (10) In the object identification device according to (8), the object feature learning unit extracts a part of the target object from a video immediately before the race, and uses the color of the part as real-time feature information immediately before the race to learn. A mechanism for calculating a histogram in advance may be provided, and the object identification unit may calculate the object identification score by examining the similarity to the color histogram.

  (11) The object identification device according to any one of (1) to (10), wherein the object identification device is an object identification device that identifies an object participating in a race, and the object identification unit As one of the identification elements to be identified in, the relative position of the target object in the race is used, and the distance from the relative position of the target object in the current frame among the multiple candidate objects identified in the past frame is A large identification score may be given to a candidate object having a close relative position.

  (12) In the object identification device according to (11), the relative position of the target object during the race is between the target object existing at the beginning and the end during the race, and the target object existing at the innermost and outermost positions. May be calculated as a normalized quantity in the space between and coordinate axes may be taken along the course shape.

  (13) In the object identification device according to any one of (1) to (12), when the object identification unit calculates a score using number recognition for a specific part, the recognition result is completely one. Rather than scoring only the numbers that have been assigned, by giving a certain score even when only certain digits match based on previously known conditions and trends, The possibility of calculating a score based on recognition may be reduced.

  (14) In the object identification device according to any one of (1) to (13), a camera and a number used for photographing when the object identification unit calculates a score using number recognition for a specific part. The distance between the parts to be recognized may be calculated, and the identification score may be calculated by assigning a weight so that the identification score is low when the distance is long.

  (15) In the object identification device according to any one of (1) to (14), it is difficult to recognize a character itself when the object identification unit calculates a score using character recognition for a specific part. In this case, the number of characters in the character portion may be read based on the shape and color information of the character portion, and scoring may be performed when identifying the target object based on the number of characters in the character portion.

  (16) The object identification device according to any one of (1) to (15), wherein the object identification device identifies an object participating in a race, and the object position estimation The unit may detect the presence of a plurality of target objects using a shadow of the target object represented by a shadow, a footprint, or a wave created by the target object, and estimate the position of the target object.

  (17) The object identification device according to any one of (1) to (16), wherein the object identification device identifies an object participating in a race, and the object part extraction Has a function to separate the background and foreground area based on the background subtraction method, and based on the start time determined before the race, it senses that no target object exists in the screen immediately before the race. Thus, foreground extraction may be realized in response to changes in lighting conditions and weather by learning the characteristics of the background immediately before the start of the race.

  (18) In the object identification device according to any one of (1) to (17), the object identification device is an object identification device that identifies an object participating in a race, and tracks the object. The object tracking unit further includes an object tracking unit capable of performing the identification result and the past identification score at the tracking destination position when the tracking is successful. Object identification may be performed again at a frequency, and the largest number among the addition of the identification scores from the tracking start frame to the current frame may be output as the identification result.

  (19) In the object identification device according to (13), the object identification device is an object identification device that identifies a racehorse, and according to the conditions and trends that are known in advance, Using the tendency that a part of the racehorse number may be hidden, taking into account the digits that are difficult to see from the camera position and orientation, give a certain score even when only certain digits match Thus, the possibility of calculating a score based on misrecognition of the number may be reduced.

(20) In the object identification device according to any one of (1) to (19), the object identification device is an object identification device that identifies a racehorse, and is extracted by the object part extraction unit. (1) Racehorse bib (2) Jockey's clothes (3) Jockey's hat (4) Horse face part (including mencos, blinkers and shadow rolls worn by the horse) (5) Bandage (6) With 6 items of horse body,
In the object identification unit, (7) for a total of 7 elements including the relative position of the racehorse,
(1) is number recognition and character recognition, (2) to (6) are color information similarities, and (7) is an identification score based on distance, and based on the weight obtained by the identification parameter determination unit Object identification may be performed.

  (21) In the object identification device according to any one of (1) to (20), the object identification device is an object identification device that identifies a racehorse, and the object feature learning unit performs pre-race The color range in which the race bib can be extracted is set from the color information of the race horse race bib that has been identified in the above, and the object part extraction unit circumscribes all regions obtained by extraction based on the color range. The rectangle is calculated, the number of the number is extracted based on the size of the circumscribed rectangle being similar to the size of the number, and the positional relationship of each part in the image based on the position of the extracted number Each part may be extracted based on the above.

  (22) In the object identification device according to (8), the object identification device is an object identification device that identifies a racehorse, and the image of the target object before the race used in the object feature learning unit is: A video of a paddock that can be obtained before the race, a video of a horse running to the starting point before the start of the race, and a video of a horse that is in the gate just before the start may be used.

  According to the present invention, when identifying a target object, a final identification score is calculated by integrating identification scores calculated from a plurality of identification elements for each candidate object. At that time, it is possible to realize highly accurate target identification by dynamically changing the identification element to be emphasized for each of the plurality of candidate objects.

It is a figure which shows the function structure of the object identification device which concerns on embodiment. It is a figure which shows the function structure of the object identification device which concerns on embodiment. It is a figure which shows the example of isolation | separation of a background area | region and a foreground area | region. It is a figure which shows the positional relationship of the extraction site | part in embodiment. It is a figure which shows the example of extraction of a number area based on a color range and a circumscribed rectangle. It is a figure which shows adjustment of the color range in an object feature learning part. It is a figure which shows specification of the image position of a target object. It is a figure which shows the example which can estimate the overlap of a target object based on a hat. It is a figure which shows the example where a number (No. 13) is hidden by a jockey's leg part. It is a figure which shows the example which recognizes the number of characters. It is a figure which shows the example of calculation of the relative position on a straight line. It is a figure which shows the example of calculation of the relative position in a curve. It is a figure which shows the processing flow of an object identification part. It is a figure which shows the calculation example of an identification score. It is a figure which shows the example of calculation of the identification score based on dynamic weight calculation. It is a figure which shows the example of a result display.

  In the following, an embodiment of the object identification device of the present invention will be described by embodying a moving object as a horse of horse racing.

The “target object” is a target object to be identified.
A “candidate object” is an object that is a candidate for identifying a target object. Generally, when a target object is identified, there are a plurality of candidate objects.
An “identification element” is a characteristic part of an object, such as the color of a horse, the color of a jockey's hat, the number of a bib, and the name of a horse described in the bib, and is effective for identifying the object. It can be a material. A candidate object having an identification element very similar to the identification element of the target object is given a high score for each identification element.

  When identifying a target object during a race, if the method of Patent Document 1 causes an overlap between target objects or if there is a target object having similar color information, the identification accuracy decreases. Stated. In response to this problem, the present invention selects a target object candidate, and dynamically changes the weighting of the identification score calculated from the identification element used for identification for each of the plurality of candidate objects, thereby achieving highly accurate target identification. Realize.

  In this method, the part of the target object in the image is extracted at a specific frame in the race. For example, in the case of a racehorse, the part shown here indicates a part that is easy to use when identifying objects such as a jockey's hat, a jockey's clothes, and a racehorse number. Which part is to be extracted is determined in advance for each competition, such as horse racing, boat racing, bicycle racing, car racing, and the like.

  Based on the position of the extracted part, the position of the object is estimated. After that, for each target object for which position estimation has been performed, the identification result is scored based on the part of each target object and the travel position at that time, and integrated to calculate a final identification score. In general, a race competition often has a plurality of pieces of identifiable information represented by numbers and colors so that the audience at the venue can identify the object from a distance, and uses the characteristics. In this case, in the present invention, unlike Patent Document 1, the identification accuracy is improved by combining not only color information but also the results of number recognition and character recognition represented by bib numbers.

  At this time, in order to solve the problem of Patent Document 1 in which the identification accuracy is lowered when there is a target object having similar color information, which identification element is to be emphasized Based on the characteristics relating to the target object that can be obtained in advance, the weighting is dynamically changed for each of the plurality of candidate objects at that time, thereby realizing highly accurate target identification.

  In addition, the “advance information” mentioned here is not limited to text information such as “the hat color is red and the horse body is chestnut hair”, which is known in advance. “Preliminary information” also includes information obtained by analyzing the image of a racehorse when running to the start point immediately before, the image of a horse when it is in the gate, etc. In addition, although the problem that analysis based on color information tends to be influenced by environmental differences such as the difference in color appearance due to the difference in weather on the day, the above-mentioned problem was solved. A mechanism may be provided that makes it easy to perform extraction by dynamically changing the extraction color range when extracting a specific part based on prior information obtained from the video.

  The present invention is an object identification device that can identify and display a target based on an image taken from a camera that takes a moving image or a still image, for example, installed in a venue represented by a racetrack or a circuit. It is. In the case of a camera that shoots still images, a camera that continuously shoots a plurality of still images may be used. It is assumed that the camera will be installed on a racetrack stand or patrol tower a little away from the course. Since the area projected by one camera is not large, if you consider the use of identifying objects throughout the course, you need to use multiple cameras, such as placing cameras every few hundred meters. is there. However, since it is not technically difficult to connect the results obtained by a plurality of cameras, and the processing performed on the video obtained from each camera is independent, it is limited to one camera here. Will be described. In the following, after describing the functional configuration of the present apparatus, the details of the function of each part will be described in the embodiment in which the target object (racing horse) is identified from the camera actually installed on the racetrack. To do.

Hereinafter, an example of an embodiment of the present invention will be described.
FIG. 1 is a diagram illustrating a functional configuration of an object identification device 1 according to the present embodiment.
The present embodiment is an apparatus that estimates the position of a target only from information of a camera installed to shoot a moving image or a still image, and identifies the target. In the case of a moving image, the number of images (frame rate) to be captured in a predetermined unit time (for example, 1 second, 1 minute, etc.) is determined depending on the speed at which the moving object is captured. do it. First, for the video 20 obtained from the camera, the object part extraction unit 120 extracts a target part. At this time, there is a case where a part is extracted by referring to a specific feature. The feature used for this extraction is determined by the object feature learning unit 110 based on the video information 10 before the race, and the like. It is stored in the information database 50. This pre-information database 50 is represented not only by the real-time features analyzed from the pre-race video information 10 and the like, but also by information registered in advance and the colors of the parts available via the network. It is assumed that features are also registered.

  Next, based on the extracted part information, the object position estimation unit 130 determines the object position considering the overlap of the target objects based on the extracted target part. For each object whose position has been estimated in this manner, the object identification unit 140 identifies each object based on number recognition, color, and analysis of the relative position of the horse. At the time of object identification, the identification parameter determination unit 150 dynamically changes and determines elements to be emphasized for each of a plurality of candidate objects based on information recorded in the prior information database 50. Further, the object identification unit 140 stores the result of object identification and the position of the identification target in the analysis result database 60. The result display unit 160 displays the last identified result.

Although the processing up to this point may be executed for each frame each time, it is desirable to perform object tracking because it is wasteful in terms of calculation cost to perform duplicate analysis and identification of objects each time. A functional configuration when the object tracking unit 170 is included is shown in FIG. Information such as past horse positions and identification results for tracking and rectangular areas in images necessary for tracking are stored in the analysis result database 60. If the tracking is successful, a process of taking over the past identification result at the tracking success position is performed.
Below, the case where the position estimation and identification of a racehorse are performed in a racetrack is made into an example, and the detail of each part is demonstrated in an Example.

[Object part extraction part]
The object part extraction unit 120 extracts the part of the target object (race horse) from the image taken by the camera. As a target part in this embodiment, “<1> racehorse race bib <2> jockey clothes <3> jockey hat <4> horse face (including mencos, blinkers, and shadow rolls worn by horses) < A total of six items, “5> Bandage wound around leg <6> Horse body color”, were extracted. In general, in racing competitions, there are often a plurality of parts useful for identification so that a spectator watching from a distance can identify the target. It is assumed that the extraction site is manually set in advance in consideration of calculation time and accuracy.

  In the present embodiment, when extracting a part, for the purpose of avoiding erroneous extraction and limiting the calculation area, the background area and the foreground area where the racehorse is present are divided in advance as shown in FIG. No calculation was made for. As the background separation method, an existing background separation method represented by the background difference method can be used. At this time, since the background color obtained from the camera changes every moment, it is desirable to learn the background characteristics in a scene where the racehorse is not running immediately before the race.

  This may be performed immediately before based on the starting time that can be known in advance, or may be provided with a mechanism that confirms that the horse is not running and learns the background again. Actually, it is difficult to perform perfect background region separation due to noise generated in the image taken from the camera, fluctuation of the grass on the course due to the wind, and the influence of raindrops in the rain. Therefore, the image after background separation may be provided with a function of removing noise such as small chunks. As a process for performing fine noise removal, there is a method of cutting an outline by performing a morphological process on an area obtained by the background difference method.

  Further, when the background area is separated, there is a problem that the influence of the shadow generated when the target object passes cannot be ignored, and the shadow is also determined as the foreground area. Therefore, based on the weather conditions immediately before the race obtained from the Internet, etc., in addition to estimating which side of the target object the shadow will occur, in advance when the target object passes in front of the camera immediately before the previous race or race The shadow color may be learned, and the shadow area may be separated based on the shadow color information and the position where it occurs.

  Part extraction is performed on only the foreground portion of the image from which the background region is separated. The simplest method for extracting a part is a method for extracting a target part from the relative positional relationship between parts separated as a foreground region. In this embodiment, it is known where each part of the racehorse appears in the overall shape of the target object. For example, if the camera is installed directly beside the course, as shown in FIG. 4, the jockey's hat and clothes appear above the target object, the bib appears near the center, and the bandage attached to the foot appears below.

  However, in the case where the target objects are overlapped with each other, there are many cases where the partial extraction cannot be successfully performed only at the position in the image. As a method for supplementing such a situation, it is conceivable that learning is performed based on images of respective parts in advance, and part extraction is performed by applying an object extraction algorithm or the like using deep learning. However, when deep learning is used, the calculation time tends to be long, so in the present embodiment, for the large foreground area where overlap is likely to occur, the part extraction is mainly performed with color information. did.

  For example, in the extraction of a racehorse race bib, the color of the race bib is determined by the race grade and is known in advance. Therefore, when a color range indicating the color of the bib is given in advance in a color space such as RGB or HSV, the color range is extracted from the foreground area, and the circumscribed rectangle is calculated for the extracted area. By extracting the number close to the size of the bib, it is possible to extract the bib as shown in FIG.

  At this time, the estimation accuracy of the race bib area may be increased by performing corner extraction or the like. As another extraction method, a method of preparing a template of each shape in advance and extracting a target region by template matching is conceivable. Since template matching is a technique that is relatively easy to increase the amount of calculation, narrow down the area where there is a high possibility that the part exists based on the relative position of the image number from the bib area obtained from the color information described above. Template matching may be used for the area.

  As a means for extracting a jockey part, a jockey part may be extracted based on a representative feature amount such as a HOG feature amount effective for person extraction. The object part extraction unit 120 does not necessarily need to extract all the specified parts. For example, the size of the circumscribed rectangle extracted by the color information described above, the matching rate of template matching, or the likelihood derived by deep learning Is outside the target value range, the result may be a part extraction failure.

  In the present embodiment, it is assumed that an image obtained from a camera installed on a racetrack is targeted. Usually, an image taken with a camera does not always accurately reflect the position in the real space, such as distortion caused by the camera lens, so it is desirable to perform calibration. Therefore, calibration parameters may be measured in advance based on a known camera calibration technique or information on the camera to be used, and the calibration may be performed before performing the object part extraction process. This camera calibration does not necessarily have to be performed if the influence of distortion is small, but it is conceivable that the accuracy of position estimation is improved because the distortion is reduced by performing the camera calibration.

[Object feature learning unit]
It has been described that it is an effective means to extract a part using a color in the object part extraction unit 120. Regarding the colors of racehorse race bibs and jockey hats, although it is possible to assume the color as static information in advance, such as “black letters on a white background” and “yellow hat”, the actual race In this case, the luminance value of the color to be extracted also changes due to environmental changes represented by the weather of the day and the inclination of the sun that changes depending on the time of day. Therefore, it is desirable that the color range to be extracted in advance is finely adjusted according to the race conditions. The object feature learning unit 110 learns the characteristics of an object represented by color information or the like for extracting a part (or used for object identification described later) from the video of the target object immediately before the race.

  As a specific means, the video of the target object immediately before the race is used. In the case of horse racing, a paddock is performed before the race, and it is possible to obtain an image of a horse that runs to the start point immediately before the start, or an image of a horse that is in the gate just before the start. For example, if a paddock or the like is used, another camera is required, but if it is a horse image that runs to the start point immediately before the start, it can be realized with the same camera.

  At this time, the part extraction is performed using the part extraction method described in the object part extraction unit 120. However, since the object feature learning unit 110 does not need to aim at execution of processing in real time unlike during a race, calculation is performed with high accuracy as represented by deep learning for one frame in which the target object is reflected. It is good also as performing the part extraction using the extraction algorithm with large cost. A feature used as a comparison target when performing object identification is calculated in advance from the extracted part and registered in the prior information database 50.

  For example, when the object part extraction unit 120 or the object identification unit 140 performs analysis using the similarity of the histogram, the object feature learning unit 110 extracts the target part from the video immediately before the race, and the histogram of the part is obtained. Have it ready. When the race actually starts, the prepared histogram is compared with the histogram of the target object created by the object part extraction unit 120 to identify the object.

  In addition, when the target part is cut out using the color range, when the bib can be extracted as shown in FIG. 6, the extracted color range is dynamically changed with respect to the area, and the size of the area that can be extracted is confirmed. To do. At that time, by determining the color range when the circumscribed rectangle size is closest to the assumed bib size as the color range used for the extraction, the color range used for the extraction according to the environment of the race is dynamically changed. It becomes possible to decide.

Similarly, if the object identification unit 140 is assumed to use a feature quantity based on the shape to measure the similarity of the jockey's clothes and the like, the object feature learning unit 110 may use the jockey of the target object. The feature amount of the clothing part is acquired in advance and used for comparison at the time of identification.
The value of the color boundary when defining the color range is referred to as “color threshold value”.

[Object position estimation unit]
The position of each target object is estimated based on the foreground region obtained by the object part extraction unit 120 and the extracted part information.
First, the number of heads is determined for each foreground region (the range of the thick dotted line) separated as shown in FIG. At this time, the overlap is extracted based on the size of the foreground area and the number of specific extraction parts. For example, even in a scene where horses overlap each other as shown in FIG. 8, a specific part (for example, a jockey's hat) is often visible. Therefore, it detects that there are two racehorses because there are a plurality of specific parts. At this time, the position of the object is determined based on the extracted position information of the part. As a position to be determined at this time, for example, a representative point such as a nose of a racehorse is determined in advance, and the position is set as a racehorse position. In the object position estimation unit 130, the position of the object is registered in a form in which the extracted part is linked to the object at that position, and is passed to the object identification unit 140 in the subsequent stage.

  In addition, the shadow region is separated by the background separation of the object part extraction unit 120, but conversely, the determination accuracy of the horse overlap may be improved based on the shadow created by the racehorse. However, this method contributes to improving accuracy only when the weather is fine and the shadow is more easily separated. As another example of using other information, it is conceivable to perform position estimation based on a horse footprint on a dirt course. This concept can be applied to other competitions, for example, in a boat race, in which the position is specified with reference to the shape of the waves generated by the ship.

[Object identification part]
The object identification unit 140 is a part that identifies a target object whose part extraction has been completed. Identification is a race in which horses with horse numbers 1 to 18 are running for each target object, and the target object corresponds to what number among 1 to 18 It is shown to clarify. Therefore, the element identification score is calculated for a plurality of identification elements, and the final identification score is calculated by adding a weighted value to each element identification score.

  The elements for calculating this element identification score are: [1] Number written on the racehorse number [2] Letter indicating the horse name [3] Color of the jockey hat [4] Color of the jockey clothes [5] Horse color [6] Menco and blinker colors worn by racehorses [7] Colors of bandages around racehorses [8] Past relative position information. At that time, if [1], number recognition, [2] character recognition, [3]-[7] color analysis, [8] distance based analysis, etc. The method shall be determined in advance for each part. In the case of identification of racehorses, number recognition and character recognition are performed because there is a race bib. However, for example, when identifying a target object having no numbers or characters on the surface, number recognition or character recognition is necessarily incorporated. There is no need.

  Further, the object part extraction unit 120 does not need to calculate an identification score for a part where the part cannot be extracted, and calculates an identification score only for a part that can be analyzed. If the object tracking unit 170 described later has successfully tracked the object and the tracked object already holds the result of identification in the past, it is not necessary to perform identification again. Alternatively, a mechanism for determining a priority identification result by comparing the tracking result with the identification score of the result newly obtained by the object identification unit 140 may be provided.

  When the position of the object is estimated by the object position estimation unit 130, generally the positions of a plurality of target objects can be estimated. In general, there is a difference in the number of parts that can be extracted for the plurality of target objects. It can be considered that a target object having a large number of parts that can be extracted can be identified more accurately than a target object having a small number of parts that can be extracted. For example, in FIG. 8, it is considered that the front target object can be identified more accurately than the back target object. Therefore, it is preferable that the object identification unit 140 identify the target objects in order of the number of parts that can be extracted. It is clear that the number of candidate objects decreases by one when the identification of one target object is completed.

  Here, first, a method of calculating each element identification score will be described. After that, the final identification score is calculated from each element identification score, and the procedure until the result is fixed to one is shown.

(1) Calculation method of each element identification score (A) Identification based on number recognition Number recognition is performed on the part where the number obtained by the object part extraction unit 120 is described. In this embodiment, based on a deep learning method represented by a convolutional neural network, a model that can be classified in number is created in advance based on a numbered image in advance, and a bib portion is added to the model. The number is recognized by recognizing the image. Since the maximum number of race horses is 18 in the central race, a model that can classify numbers from 1 to 18 is created, and if a number greater than the number of races in the race is recognized, the recognition failure shall be made. I made it. Actually, a model may be created for each head, and the model to be used may be changed based on the number of heads in the race.

  As an identification scoring method for number recognition, if automation is considered, for example, there is a method using an output probability. When the softmax function or the like is used as the output function of the convolutional neural network, it is possible to calculate the value of which number is the number as the probability. From this, for example, when 3 is recognized as 69% and 13 as 31% when passing through a certain bib image, scoring is automatically performed, such as 69 in 3 and 31 in 13. Is possible.

  Also, this scoring may be manually determined from pre-recognized preconditions. For example, when an experiment was conducted prior to filing, in the number recognition of the racehorse number, for example, when there was a “13” horse, “1” in the tens place was shown by the jockey's feet as shown in FIG. It has been found that there are many cases in which it is difficult to see and the number “13” is misrecognized as “3”, or the number “3” is misrecognized as “13”. From this fact, for example, when it is recognized as “3”, it is determined that there is a high possibility that it is either “3” or “13”, and “3” is 100 points, “13” is 70 points, Other numbers, such as 0, may be graded based on the already known misrecognition tendency and scored with emphasis on the recognition result of the last one digit.

  As a result of the number recognition, the recognition accuracy tends to be higher when the number is clearly displayed. Therefore, a mechanism may be provided that performs weighting so that the identification score of a long distance is low based on the distance between the camera number extracted in number recognition and the camera. Further, as a number recognition method, the number may be recognized using a matching method represented by template matching or the like without using deep learning.

(B) Identification Based on Character Recognition Since a horse race number has a horse name as well as a number, the horse can be determined based on the character recognition of the number. As a character recognition method, it is possible to determine a character by a classification method represented by machine learning or the like for each character, as in number recognition. However, as described above, in the present embodiment, it is assumed that the image is taken with a camera from a distant position, so it is difficult to accurately read the character with a camera taken from a long distance.

  However, as for the number of characters in the horse name, it was possible to roughly measure the number of characters based on the color information of the horse number portion of the bib number as shown in FIG. In Japan, the name of a racehorse is defined as two or more characters and no more than nine characters. However, by determining the approximate number of characters and assigning a certain identification score to the racehorses with a similar number of characters, the accuracy of identification is improved. A mechanism may be provided.

(C) Identification Based on Color Information It is also a useful means to analyze the color information of each part and determine what number the racehorse is closest to. For example, in the case of a jockey's hat, the first and second horses are supposed to wear white hats in 18 races. Also, the color of the jockey's clothing is known information determined by the owner, and it is hard to imagine that the color of the racehorse's body will change suddenly. Is known information.

Therefore, for example, if the hat portion can be extracted, the histogram similarity with the hat image of each color prepared in advance (or the luminance value of the assumed hat color) is calculated, and the highest similarity is obtained. The color can be determined by using an identification result that is close. This time, an RGB image was converted to HSV, and histogram similarity was calculated in the HSV color space. As the similarity, a method using a correlation coefficient was used. The formula for obtaining the correlation coefficient is shown below.

The above N represents the total number of bins in the histogram, the histogram of the reference image H _r is to be compared, H _t denotes the histogram of the target image. The color of hat in identification of racehorses a H _r for a eight by eight prepared, a correlation between the target image H _t of the extraction region determined, and the recognition result most things correlation is large. In addition, as the reference image, it is desirable that the object feature learning unit 110 extracts the part of the hat based on the previous video and uses the extracted image as the reference image.

  Further, as a method of analyzing colors other than the similarity of the histogram, a range of luminance values indicating the color of each hat of the candidate object is registered, and based on the number of pixels included in the range, Identification may be performed based on a method of determining the closest hat color, an average of luminance values of images in the target region, or the like. In addition, when judging the color of jockey's clothing, the shape feature values represented by HOG (Histograms of Oriented Gradients) and EOH (Edge of Orientation Histograms) are used to consider the clothing pattern. Classification may be performed based on the degree.

(D) Identification Based on Past Relative Position Information In general, when a race competition is performed, it is difficult to think of a sudden change in the race position of the target object. For example, it is not a frequent occurrence that a racehorse that ran to the end suddenly jumps to the beginning, or a horse that ran outside suddenly starts to run inside the course. In horse racing, races often proceed while keeping the formed horse group. Therefore, in the group formed by a plurality of target objects, information indicating where the target objects are currently located can be useful information for identification in comparison with past position information. From this point of view, the relative position of the target object in the race is one of the elements for identification.

  The method for calculating the relative position is realized by defining horizontal and vertical coordinate axes with respect to the course shape as shown in FIG. At this time, the axis representing front and back is normalized as p-axis, the top is normalized by p = 0, and the last is normalized by p = 1. Similarly, the axis representing the inside and outside was designated as the v axis, the inside was normalized with v = 0, and the outside with v = 1. In the case of a straight line, as shown in FIG. 11, even if the course shape is a curve, the axes are arranged along the course shape as shown in FIG.

When the object identification is completed, the object identification unit 140 records the relative position of the object in the analysis result database 60 based on the position of the target object and the final identification result. Assuming that this recording has been performed in the past frame, if the target object to be identified is currently located at p = 0.5 and v = 0.5, p = 0.5 and v = The identification score is calculated in a form that increases the identification score of the number located near 0.5. As a formula, when the position of the target object to be identified is (T _p , T _v ) and the relative position in the past frame of the candidate object whose score is to be calculated is indicated by (C _p , C _v ), The distance d between the position (T _p , T _v ) and the relative position (C _p , C _v ) of the candidate object in the past frame is calculated as follows.

_{d = k 1 | T p -C} p | + k 2 | T v -C v |

However, k ₁ and k ₂ are coefficients that determine whether to place importance on the result of front and back or inside and outside, and may be determined manually, but the actual distance from p = 0 to p = 1, and v = 0 It may be determined in a form proportional to the actual distance up to v = 1 (provided that k ₁ + k ₂ = 1). At this time, the closer the relative position is, the closer the distance d is to 0, and the farther the relative position is, the larger the distance d is. Therefore, the distances d for all the object numbers are calculated, and the score of the one having the smallest distance d is 100. The score of the one with the maximum distance d is normalized and scored so that the score becomes zero.

上記のように計算される。 (2) Determination method of identification result Here, the process in the case of finalizing a final identification result from each element identification score is shown. The present invention is characterized in that the target object is identified by dynamically changing the weight according to the candidate object.
The calculation method of the final identification score in this method is that when a score for a candidate object of a number i having a certain element identification score (element number j) is indicated by p _ij ,

Calculated as above.

Here, s _i represents the final identification score of the candidate object of number i, J represents the number of elements for which the element identification score can be calculated (parts that failed to extract the part are not included), and j represents the calculation of the element identification score The number of the successful part is shown. For example, the number of the race bib is j = 0, the color of the jockey's hat is j = 1, and the like. W _j represents a weight determined for each part. The calculation method of w _j is not described here because the identification parameter determination unit 150 performs dynamic weight determination.

The calculation of s _i shown here is preferably performed a plurality of times to reduce the number of candidate objects when the candidate objects cannot be narrowed down to one by a single implementation. The processing flow at this time is shown in FIG.

When the final identification score is calculated for each candidate object when performing multiple times, the maximum value among all s _i is set as s _max . For this _smax , a score recorded at a certain ratio or higher is set as a new candidate object, and the process is repeated until the number of candidate objects finally becomes one or the number of candidate objects does not change. During this time, as the candidate objects are narrowed down, the optimum parameter for identifying the narrowed candidate objects is output from the identification parameter determination unit 150.

FIG. 14 shows the calculation when the final identification score is calculated on the assumption that the weight w _j is calculated. However, for the sake of simplicity, the element identification score is set to four items: a bib number, a jockey hat, a horse color, and a relative position. Next _{s max} in this case is _161, defined those scores more than 50% of the value of _{s max} to the next candidate object. Therefore, the horse numbers j = 3, 4, 5, 6, and 8 become the next candidate objects, the weight w _j is calculated again, and the score is calculated based on the weight w _j . Thereafter, this is repeated to determine one final candidate.

[Identification parameter determination unit]
The identification parameter determination unit 150 has a mechanism for calculating a weight for determining the weight w _j for calculating the identification score in the object identification unit 140 for each candidate object.
In general, adding a plurality of element identification scores to calculate a final identification score is useful for performing highly reliable object identification. However, when candidate objects can be narrowed down, it is inappropriate to use the same weighting every time, and it is desirable to use parameters that are easy to identify between candidate objects.

  For example, if there are three candidate horses, and one of the candidates has a white horse and the other two have a black horse under such conditions, the information on the horse will be those candidates. This can be useful information for identification. However, when the horse body is a color close to black for all three candidates, it is desirable to refer to information on other parts and not to see the horse information as much as possible. Therefore, the identification parameter determination unit 150 provides weights that are dynamically changed in accordance with candidate objects obtained from the object identification unit 140.

As a technique for calculating the weight w _j of an identification element (number j), in this embodiment, the number of clusters C _j , the variance V _j of the representative value of each cluster, and the identification score obtained from the element are used. The element reliability R _j indicating how much reliability can be obtained is calculated by multiplying as follows.

w _j = (C _j / N) × V _j × R _j

Here, N represents the number of candidate objects handled this time.
Here, each parameter will be described based on an example. For example, the candidate objects given here are the six heads of No. 6, No. 16 and No. 17, and the elements that have succeeded in calculating the element identification score are the No. number (j = 0), the hat part of the jockey ( j = 1), horse color (j = 2), and past relative position (j = 3). FIG. 15 shows the information and identification score of each part in this example. When explanation is added to this identification score, the bib number was identified with a high score (100 and 50) because the shapes of the numbers 6 and 16 are similar. In the 18 races, the hats were No. 6 (red), No. 16 (peach), and No. 17 (peach), and all of the scores were relatively high due to relatively similar color information. Regarding the horse body, the 6th horse was a white horse, and the other horses had a color close to black. As for the relative positions, all of No. 6, No. 16, and No. 17 had been running relatively close in the past, and all showed high scores.

At this time, the cluster number C _j is the number of clusters when clustering is performed for a certain feature. In general, when looking at a certain part, it can be said that a part where candidates can be narrowed down at a time is an excellent part for identification. For example, in the case of FIG. 15, the number number can be classified by itself so that C _j = 3. Similarly, since the hat is composed of 6 types (red), 16 (peach), and 17 (peach), C _j = 2. In this way, it is decided that the number of types that can be clearly divided will be determined based on information such as the number of runners of the race that can be obtained as advance information via a network or the like.

Similarly, jockey's clothing is tied to the owner, so clustering may be performed based on the name of the owner, but other ideas include, for example, preparing images of jockey's clothing in advance. The number of clusters that can be obtained when the similarity is calculated and repeated in the order of decreasing similarity until the similarity falls below a certain threshold may be defined as C _j .

Also, regarding the past relative positions, the number of clusters formed when the distance is repeated until the distance falls below a certain threshold is determined as C _j while considering the same cluster in order from the closest distance. However, when C _j = 1, the parameter is not suitable for identification, and the weight is set to 0. In the example of FIG. 15, since the relative positions of all the horses are similar, C ₃ = 1 and weight calculation is not performed.

Next, the variance V _j of the representative value of each cluster is calculated. This is based on the fact that even if the number of clusters C _j is large, erroneous recognition may occur if the features of the respective clusters are not significantly different, and it cannot be said that it is an excellent part for identification. Yes. Here, for the element identification score p _ij calculated in advance, the average value of each cluster is calculated, and the variance with respect to the average value is calculated. The variance V _j for each identification element thus obtained is normalized by 1 at the maximum.

  This calculation will be described based on the hat portion of FIG. 15. The hat portion is divided into No. 6 (cluster 1), No. 16 and No. 17 (cluster 2). Since the average score of the hat portion of cluster 1 is 100 and the average score of the hat portion of cluster 2 is 60, the variance of 100 and 60 is calculated.

  15 will be described. The horse part is divided into No. 6 (cluster 1), No. 16 and No. 17 (cluster 2). Since the score average of the horse part of cluster 1 is 10 and the score average of the horse part of cluster 2 is 95, the variance of 10 and 95 is calculated.

The element reliability R _j is a score indicating how much the identification score obtained from the element is reliable, and a larger value is entered as the reliability is higher. Of course, in general, when extracting and recognizing a part, there is a difference between a part where extraction is likely to be successful and a part where the identification score is reliable. Since this R _j is based on the ease of extracting the part and the reliability, it may be a static one set in advance.

Since the hat is very small in the image and there are cases where the extraction fails, the value of R _j is reduced in FIG. On the other hand, as an idea for automating R _j , in the case of part extraction, template matching with a template prepared in advance is performed on a part that has been extracted, and a matching rate is obtained and the value is set to R _j. If the relative position of the target is used as an element, the situation is considered to change as the frame is earlier than the current frame, so it decreases based on the time difference from the frame with the relative position compared. It is conceivable to set R _j so that

Thus, w _j is dynamically calculated for each of a plurality of candidate objects. However, when the number of candidate objects is large, this dynamic weighting may not work well because a specific weight becomes too large. For example, when 18 candidate objects are handled, the number number is C _j = 18, but there is only one horse with a characteristic hair color, and the other is a horse with a similar hair color. Will be C _j = 2 and the weight of the horse will be very small. Therefore, in cases where there are many candidate objects, the weight of a fixed number such as a bib number will be too large every time, and if a recognition error occurs in the bib number part, the wrong result will be selected regardless of the score of other elements It becomes easy to be. In this embodiment, candidate objects can be narrowed down little by little, so it is not desirable to perform identification that greatly depends on one weight from the beginning.

Therefore, with respect to this dynamic weighting, it is desirable to make an adjustment so that the difference between the weights becomes smaller while the number of candidate objects is large. In order to realize this automatically, the adjustment is performed so that the weights after adjustment are gathered around the average value of the weights as the number of candidate objects is larger as follows.

w _j ′ = w _ave + (w _j −w _ave ) / (N−1)

However, w _ave indicates the average value of all weights, and N is the number of candidate objects. When this adjustment is performed, w _j ′ is used as the final weight for object identification in the object identification unit 140.
and _{w k} the weights of different sites and w _j, the weights after adjustment of _{w k} and _{w k} ', from the above equation,

w _j ′ −w _k ′ = (w _j −w _k ) / (N−1)

Therefore, the difference between the weights is reduced to 1 / (N−1).

  By this calculation, the final weight is determined as shown in FIG. In the example of FIG. 15, when all weights are set to 1, the final identification score is 310 for horse number 6, 300 for horse number 16, and 260 for horse number 17, so horse number 6 is selected. . However, in the weighting calculated in the present invention, it is emphasized that the weight of the horse No. 6 is greatly different from the target object by weighting the horse part greatly, and the final identification score is No. 16 Is selected.

[Result display area]
The result display unit 160 displays the position estimation and identification results so far. Here, the result is displayed on a plane map simulating the shape of the course as shown in FIG. At this time, not only the position but also the track of the racehorse may be displayed based on the tracking result, and the marker indicating each horse may be displayed in the color of the frame. As a means of obtaining the position on the planar map that overlooks the course from the top from the position on the image, correspondence information on which position on the planar map the pixel position in the image corresponds to and a conversion formula are prepared in advance. Just keep it.

  This correspondence information may be prepared manually, or in the case of a linear shape, a transformation that determines the position by estimating the homography matrix between the quadrilateral indicating the course shape and the quadrilateral indicating the position on the plane map There is also an example of obtaining an expression. The position of the racehorse corresponding to the position on the plane map may be the head part of the racehorse or the foot part, but which part is used is determined in advance.

[Object tracking part]
The apparatus of the present invention is more likely to perform high-speed and high-accuracy recognition if an object tracking algorithm is incorporated.
The object tracking unit 170 tracks the current frame object based on the past frame information. As a tracking method, there is a method in which the moving distance from the previous frame is obtained and the closest candidate object is made the same object, but in addition to that, the entire target object in the image estimated by the object position estimation unit 130 or a part It is conceivable to set a rectangular area including “” and track based on the characteristics of the rectangular area.

  As the latter method, for example, it is possible to apply an existing tracking method such as MediaFlow or KCF. At this time, for example, if it is a racehorse, it will often go straight forward during the race, and it will rarely change direction. The result may be passed to the object position estimation unit 130 as a failure.

  Further, when the tracking is successful, the identification number determined by the object identification unit 140 may be taken over so that the candidate object is not identified again during the tracking success. Alternatively, an implementation may be considered in which the object is identified again once every few frames during tracking and the identification score is calculated again. In particular, in the case where object identification is performed once every few frames even during the latter tracking, the identification score is added as needed during tracking, and the highest number among the identification scores from the tracking start frame to the present is added. May be provided as a result of identification.

  As mentioned above, although embodiment of this invention was described, this invention is not restricted to embodiment mentioned above. Further, the effects described in the present embodiment are merely a list of the most preferable effects resulting from the present invention, and the effects of the present invention are not limited to those described in the present embodiment.

  In the present embodiment, horse racing is taken as an example, but the present invention can also be applied to other races such as bicycle races, boat races, and car races. In addition, the present invention is not limited to a race, and is applied to a field in which an arbitrary moving object (for example, a student playing at an athletic meet, a player playing in a competition such as soccer) is photographed with a camera and identified. be able to.

  Each process in the object identification device of the present invention is realized by software. When realized by software, a program constituting the software is installed in an information processing apparatus (computer). These programs may be recorded on a removable medium such as a CD-ROM and distributed to the user, or may be distributed by being downloaded to the user's computer via a network. Furthermore, these programs may be provided to the user's computer as a Web service via a network without being downloaded.

DESCRIPTION OF SYMBOLS 1 Object identification apparatus 2 Object identification apparatus 10 Camera image | video 20 Camera image | video 50 Prior information database 60 Analysis result database 110 Object feature learning part 120 Object site | part extraction part 130 Object position estimation part 140 Object identification part 150 Identification parameter determination part 160 Result display part

Claims (22)

An object part extraction unit for extracting a part used for identification of a target object from an image;
An object position estimation unit for estimating the position of the target object;
An object identification unit that performs identification for each identification element for the target object, and finally identifies the object by aggregating the identification results for each identification element of the target object;
An object feature learning unit that learns in advance the feature of each candidate object used by the object part extraction unit and the object identification unit;
An identification parameter determination unit that dynamically determines a weight when the object identification unit aggregates identification results for each identification element of the target object;
A result display unit for displaying a final object identification result of the object identification unit;
An object identification device characterized by comprising: The object identification device according to claim 1, wherein the identification parameter determination unit weights an identification element that can easily identify candidate objects among identification elements that have been successfully extracted. In order for the identification parameter determination unit to heavily weight elements that can easily identify candidate objects,
The number of clusters when clustering the features of the identification elements of all candidate objects,
The variance of the representative value of each cluster,
The element reliability of the extracted element,
The object identification device according to claim 2, wherein a combination of the three parameters gives a large weight to identification elements that facilitate identification of candidate objects. When the object identification unit calculates a final identification score, a candidate object that shows a score that satisfies a certain condition is determined as the next candidate object, and the identification parameter determination unit narrows down the candidate objects while repeatedly changing the weighting. The object identification device according to claim 1, wherein The identification parameter determination unit, when setting weights for object identification, makes adjustments so as to reduce the difference between the weights when the number of candidate objects is large. 5. The object identification device according to any one of 4. The object part extraction unit extracts a plurality of parts for identification,
The object identification unit reduces the number of candidate objects in order from the target object that is easy to identify by identifying the object in order from the target object having the largest number of successfully extracted parts, and increases the identification accuracy. The object identification device according to claim 1, wherein the object identification device is a device. When the position estimation is performed by the object position estimation unit, when the target objects exist in the vicinity and the target object in the back is partially obscured by the target object in front, the part that is not obscured The position is accurately estimated even when there is an overlap by estimating the number and position of the target objects based on detection of a plurality of the parts. The object identification device according to any one of the above. An object identification device for identifying an object participating in a race,
A pre-information database that the identification parameter determination unit refers to determine a weight;
The object feature learning unit captures not only the static information registered in the prior information database before the race but also the state of the target object before the race as a video, and extracts the part of the object from the video. The object identification device according to claim 1, wherein real-time feature information immediately before the race is learned. As the real-time feature information immediately before the race learned by the object feature learning unit, the part of the candidate object is extracted from the video immediately before the race and the color range is dynamically changed with respect to the part. 9. The object identification apparatus according to claim 8, wherein the color threshold is adjusted so that the circumscribed rectangle size of the region of the part approaches the size of the part set before the race. The object feature learning unit includes a mechanism for extracting a part of the target object from the video immediately before the race and pre-calculating a color histogram of the part as real-time feature information immediately before the race to learn,
9. The object identification device according to claim 8, wherein the object identification unit calculates an object identification score by examining a similarity to the color histogram. An object identification device for identifying an object participating in a race,
The relative position of the target object in the current frame among the plurality of candidate objects identified in the past frame using the relative position of the target object in the race as one of identification elements to be identified by the object identification unit 11. The object identification device according to claim 1, wherein a large identification score is assigned to a candidate object having a relative position that is close to a distance from the object object. The relative position of the target object during the race is calculated as a normalized value in the space between the target object existing at the beginning and the end of the race and the target object existing at the innermost and outermost points. The object identification apparatus according to claim 11, wherein coordinate axes are taken along the course shape. When calculating a score using number recognition for a specific part in the object identification unit, a score or condition that is known in advance is not used for scoring only numbers that are completely consistent as a recognition result. 13. The possibility of calculating a score based on misrecognition of a number is reduced by giving a constant score even when only a specific digit is matched based on The object identification device according to any one of the above. When calculating the score using the number recognition for the specific part in the object identification unit, the distance between the camera used for shooting and the part that performs the number recognition is calculated, so that the identification score when the distance is long is lowered. The object identification device according to claim 1, wherein an identification score is calculated by weighting the object. When calculating the score using character recognition for a specific part in the object identification unit, if it is difficult to recognize the character itself, the number of characters in the character part is read based on the shape and color information of the character part. The object identification device according to claim 1, wherein scoring is performed when the target object is identified based on the number of characters. An object identification device for identifying an object participating in a race,
The object position estimation unit detects the presence of a plurality of target objects by using a trace of a target object represented by a shadow, footprint, or wave created by the target object, and estimates the position of the target object. The object identification device according to any one of claims 1 to 15. An object identification device for identifying an object participating in a race,
The object part extraction unit has a function of separating the background and the foreground area based on the background difference method, and there is no target object in the screen immediately before the race based on the start time determined before the race. The foreground extraction is realized in response to changes in lighting conditions and weather by learning the background characteristics immediately before the start of the race. Object identification device. An object identification device for identifying an object participating in a race,
An object tracking unit capable of tracking the object;
When the tracking is successful, the object tracking unit can transfer the identification result and the past identification score to the position of the tracking destination, identify the object again at a constant frequency during the tracking, and start the current tracking from the tracking start frame. 18. The object identification device according to claim 1, wherein the largest number among the identification scores up to the frame is output as an identification result. An object identification device for identifying a racehorse, using a tendency that a part of a racehorse number is hidden by a jockey's foot as a condition or tendency that is known in advance, and a camera Taking into account digits that are difficult to see from the position and orientation of the, it is possible to reduce the possibility of calculating a score based on misrecognition of numbers by giving a certain score even when only certain digits match The object identification device according to claim 13. An object identification device for identifying a racehorse, in which the object part extraction unit (1) racehorse number (2) jockey clothes (3) jockey hat (4) horse face Parts (including mencos, blinkers, and shadow rolls worn by horses) (5) Bandage (6) Over 6 items of horse body,
In the object identification unit, (7) for a total of 7 elements including the relative position of the racehorse,
(1) is number recognition and character recognition, (2) to (6) are color information similarities, and (7) is an identification score based on distance, and based on the weight obtained by the identification parameter determination unit The object identification device according to claim 1, wherein the object identification is performed. An object identification device for identifying a racehorse, wherein the object feature learning unit sets a color range in which the number can be extracted from the color information of the number of the racehorse known before the race, and the object part extraction The unit calculates the circumscribed rectangle for all the areas obtained by extraction based on the color range, and extracts the bib portion based on the size of the circumscribed rectangle being similar to the bib size. 21. The object identification device according to claim 1, wherein each part is extracted based on a positional relationship of each part in the image with reference to the position of the extracted bib portion. An object identification device for identifying a racehorse, which is a video of a target object before a race used in the object feature learning unit, and a video of a paddock that can be obtained before a race, and runs to a starting point before a race starts. The object identification device according to claim 8, wherein the image of the horse at the time and the image of a horse that is housed in the gate immediately before the start.

Images