JP2018060360A - Person number estimation method, person number estimation device, person number estimation program and recording medium - Google Patents

Abstract

A number estimation method, a number estimation device, a number estimation program, and a recording medium capable of reducing the labor and cost of individual learning in the number estimation using image feature amounts are provided. A calibration parameter indicating a geometric relationship between a fixed camera and a real space coordinate system is measured at the time of presetting, a background image that is an image photographed by the fixed camera in an unattended state is acquired, and calibration is performed. A person arrangement image that is an image in which a person image prepared in advance is arranged on a background image while referring to a parameter is generated, an image feature amount is calculated from the generated person arrangement image, and the calculated image feature amount and arrangement are calculated. Feature data for learning indicating the relationship with the number of people is stored. At the time of measurement, a captured image captured by a fixed camera is input, an image feature amount is calculated from the input captured image, and a captured image is calculated based on the calculated image feature amount and stored learning feature amount data. Estimate the number of people in the picture. [Selection] Figure 1

Description

  The present invention is intended for security use, safety management use, marketing use, etc., and in a picture taken with a fixed camera, a number estimating method and a number estimating apparatus for estimating the number of persons existing at a place being shot , A number estimation program, and a recording medium.

  In places where many people gather, such as stations, airports, and event facilities, excessive congestion can cause serious accidents such as shogi defeat. For this reason, in addition to taking measures in advance so as not to cause excessive congestion, it is necessary to take on-site measures such as appropriately guiding when congestion occurs. Therefore, there is a safety management need to monitor the congestion situation at each important point in the facility. On the other hand, in commercial facilities, amusement facilities, etc., there is a marketing need to measure the number of people in the facility as data for the purpose of ensuring customer comfort and effective sales promotion, etc., and to use the measurement results. ing. There are various sensors for measuring the number of people, but when a place to be measured is comparatively wide, an image taken by a camera is an effective measuring means.

  Various methods for measuring or estimating the number of persons in an image from an image have already been proposed and used. These methods are roughly classified into the following three methods.

(1) Method for detecting individual persons on an image (person detection type)
(2) Method to calculate the number of people by tracking individual moving objects on the image (tracking type)
(3) A method for estimating the number of persons by integrating feature amounts on an image, results of local image processing, etc. without detecting individual persons (feature type)

Japanese Patent No. 4914870

Velastin A.N.Marana, LF Costa, RA Lotufo, SAVelastin, "On the Efficacy of Texture Analysis for Crowd Monitoring", Proc. International Symposium on Computer Graphics, Image Processing, Vol.6, pp.3521-3524 (1998 ).

  Many conventional techniques are realized by combining any of the above-described methods or the above-described methods. Among them, the methods (1) and (2) above achieve high accuracy in a situation where there are few people, but in a crowded situation, the error becomes large and the processing breaks down. It becomes difficult. Among the above-mentioned usage scenes, in safety management at public facilities such as stations and event facilities, it is essential that a certain level of accuracy be maintained even in crowded situations, so the above (1) and (2) It is difficult to use the method.

  On the other hand, although the method (3) generally does not provide high estimation accuracy in a situation where there are few people, it can be estimated with relatively stable accuracy even in a crowded situation (see, for example, Non-Patent Document 1). . However, in order to actually use the method (3), learning or regression analysis is required in which learning data is acquired for each camera and the relationship between the feature quantity and the number of people is analyzed. The learning data here is data obtained by acquiring a large number of images in which a person is actually captured, measuring the number of persons in the acquired image, and recording the measured data for each image.

  On the other hand, a method of estimating the number of people by only camera calibration without performing individual learning has been proposed (see, for example, Patent Document 1). The camera calibration is a process of measuring calibration parameters indicating a geometric relationship between the camera and a real space coordinate system such as a floor surface of the camera installation location. However, in an overcrowded situation such as a situation where the image is filled with people, the number of people cannot be estimated by the method using camera calibration.

  In other words, if you want to measure or estimate the number of people in an overcrowded situation, collect learning data for each camera, and the relationship between the feature value or the image itself and the number of people in the image Requires separate learning or regression analysis. Therefore, the trouble and cost for collecting learning data and the trouble and cost for performing learning or regression analysis are hindered, and there has been a problem that the practical application has not progressed. In particular, the labor and cost of collecting learning data are a problem.

  The present invention has been made in view of the circumstances as described above, and a person estimation method, a person estimation apparatus, and a person estimation program that can reduce the labor and cost of individual learning in person estimation using image feature amounts. And a recording medium.

  In order to achieve the above object, a method for estimating the number of persons according to the present invention includes a calibration unit, a background image acquisition unit, a person arrangement image generation unit, a learning feature amount storage unit, an image input unit, an image feature amount calculation unit, Calibration method for measuring a calibration parameter indicating a geometric relationship between a fixed camera and a real space coordinate system at the time of presetting. A background image acquisition step in which the background image acquisition unit acquires a background image that is an image taken by the fixed camera in an unattended state, and the person arrangement image generation unit refers to the calibration parameter. However, a person arrangement image that generates a person arrangement image that is an image in which a person image prepared in advance is arranged on the background image And the learning feature amount storage unit calculates an image feature amount from the generated person layout image, and stores learning feature amount data indicating a relationship between the calculated image feature amount and the number of persons arranged. A learning feature amount storage step, an image input step in which the image input unit inputs a captured image captured by the fixed camera, and an image feature amount calculation unit, from the input captured image, at the time of measurement. An image feature amount calculating step for calculating a feature amount, and the number-of-person estimation unit adds the image feature amount to the captured image based on the calculated image feature amount and the learning feature amount data stored by the learning feature amount storage unit. And a number estimation step for estimating the number of persons in the image.

  In the number estimating method according to the present invention, in the number estimating step, the number estimating unit performs a neighborhood search on the feature amount of the learning feature amount data with respect to the image feature amount calculated at the time of measurement, and the search result is obtained. The number of persons may be estimated based on the above.

  Further, in the number estimating method of the present invention, the number estimating device includes a learning unit, and the learning unit indicates the relationship between the image feature amount and the number of people using the learning feature amount data at the time of presetting. Further, a learning step of learning a regression equation is performed, and in the number of people estimation step, the number of people estimation unit inputs the image feature amount calculated at the time of measurement into the regression equation learned by the learning unit, The number of people may be estimated.

  Further, in the person estimation image generation step of the present invention, in the person arrangement image generation step, the person arrangement image generation unit genders the person image to be arranged based on a profile relating to a person appearing in a place photographed by the fixed camera, You may make it select according to an age and at least one of clothes.

  In the number estimating device according to the present invention, the number estimating device includes a person image detection / recording unit, and the person image detection / recording unit detects a person in a photographed image photographed by the fixed camera. A person image detection recording unit for recording as the person image is further performed. In the person arrangement image generation step, the person arrangement image generation unit converts the person image recorded in the person image detection recording unit into the person arrangement image. It may be arranged.

  In addition, the number estimating device of the present invention was photographed by the fixed camera in the unattended state with a calibration unit that measures a calibration parameter indicating a geometric relationship between the fixed camera and the real space coordinate system at the time of presetting. A background image acquisition unit that acquires a background image that is an image, and a person arrangement image generation unit that generates a person arrangement image that is an image in which a person image prepared in advance is arranged on the background image with reference to the calibration parameter And a learning feature amount storage unit that calculates an image feature amount from the generated person arrangement image, stores learning feature amount data indicating a relationship between the calculated image feature amount and the number of arranged people, and at the time of measurement, An image input unit that inputs a captured image captured by the fixed camera, and an image feature amount calculation that calculates the image feature amount from the input captured image And a number estimation unit that estimates the number of persons in the captured image based on the calculated image feature amount and the learning feature amount data stored by the learning feature amount storage unit. .

  In order to achieve the above object, the number of persons estimation program of the present invention is a program for causing a computer to execute each step of the number of persons estimation method of the present invention.

  In order to achieve the above object, the recording medium of the present invention is a computer-readable recording medium recording a number estimation program for causing a computer to execute each step of the number estimation method of the present invention.

  According to the present invention, it is possible to reduce the labor and cost of individual learning in estimating the number of people using image feature amounts.

It is a functional block diagram which shows the basic functional structure of the number estimation apparatus which concerns on embodiment. It is a schematic diagram for demonstrating the production | generation method which produces | generates a person arrangement | positioning image in the person estimation apparatus which concerns on embodiment. It is a flowchart which shows the flow of the production | generation process and estimation process in the person estimation apparatus which concerns on embodiment. It is a functional block diagram which shows another example of the basic functional structure of the number estimation apparatus which concerns on embodiment.

  Hereinafter, an embodiment of a number estimating device according to the present invention will be specifically described with reference to the drawings.

  In the present embodiment, a case will be described in which image data (video data) captured by a fixed camera is input to a computer, and image processing is performed on the input image data on the computer. However, the present embodiment is limited to this embodiment. Not. In addition, in the present embodiment, real-time processing that acquires image data from a fixed camera one by one and performs image processing one by one, and image data that is once recorded on a recording medium such as a hard disk are obtained. Any of the off-line type processes that read the image and perform image processing can be applied.

  FIG. 1 is a functional block diagram showing a basic device configuration of a number estimating device 10 according to the present embodiment.

  First, a configuration for performing processing at the time of presetting will be described with reference to the drawings.

  As shown in FIG. 1, a person image storage unit 12, an image input unit 14, a background image acquisition unit 16, and a calibration unit are configured as a configuration for performing processing at the time of presetting, as shown in FIG. 18, a person arrangement image generation unit 20, an image feature amount calculation unit 22, a learning feature amount storage unit 24, and a learning feature amount data storage unit 26.

  The person image storage unit 12 stores a plurality of person images in which a person is drawn in advance. The person image here is a rectangular image in which the whole body of one standing person is drawn as shown in FIG. 2 as an example. In the present embodiment, the person image storage unit 12 stores a plurality of person images 44 in which the whole bodies of various persons are drawn.

  The image input unit 14 acquires a frame image from a fixed camera, a moving image file captured by the fixed camera, a still image sequence captured by the fixed camera, and the like at the time of presetting.

  The background image acquisition unit 16 acquires, from the frame image acquired by the image input unit 14, a background image that is an unmanned image that is captured by the fixed camera and does not show a person. The background image 50 can be most easily acquired by manually capturing an image with the fixed camera while no person is shown. Alternatively, the background image 50 may be an image obtained by calculating an average value in the time direction for each pixel using a plurality of frame images photographed by the fixed camera at different times. Various methods for acquiring the background image 50 have already been proposed, and the method is not limited here.

  After the installation of the fixed camera, the calibration unit 18 acquires a calibration parameter describing the geometric arrangement of the fixed camera as a pre-process of the measurement process. The calibration parameters here are parameters indicating the position of the fixed camera in the floor coordinate system, the shooting direction, the focal length (magnification ratio), and the like. As such camera calibration, a method using an image obtained by photographing a known grid pattern, dot pattern or the like on the floor surface is widely known, but other methods using an image obtained by placing a known structure are used. (For example, see the following reference 1).

[Reference 1] Japanese Patent No. 5331047

  Various methods are used as the camera calibration method. Here, it is only necessary to acquire the above-described parameters, and it is possible to specify what line each pixel of the frame image corresponds to in real space. It is only necessary to acquire information that can be obtained, and the specific method is not limited.

  The person arrangement image generation unit 20 generates a person arrangement image 52 in which the person image 44 is arranged on the background image 50 using the background image 50, the calibration parameters, and the person image 44 prepared in advance.

  In the following description, it is necessary to appropriately execute the following two geometric calculations (4) and (5) on the assumption that the calibration parameters are known.

(4) Image coordinates (x of the corresponding pixel on the image when real space coordinates (X, Y, Z) (where X, Y are coordinates on the floor and Z is coordinates in the height direction) are designated. , Y) (5) Calculation of the coordinates (X, Y) on the floor when the image coordinates (x, y) and the height Z from the floor in real space are specified

  However, since these specific calculation methods are already widely known as shown in the above-mentioned reference 1 and the like, detailed description thereof is omitted here.

  Hereinafter, an example of a specific method for generating the person arrangement image 52 will be described.

  First, from the image coordinates (x, y) of the four corners of the image or the rectangular four corners that specify the processing range in the image, the floor surface coordinates (X, Y) at the corresponding floor surface height using the calibration parameters. Ask for. Then, a quadrangle having the obtained four points on the floor as vertices is set as a person arrangement range 40 in which a person is arranged on the floor.

  As a method of arranging a person at a random position within a predetermined area on the floor surface, for example, a square (length W in the X direction, length W in the Y direction) that includes the above-described square on the floor surface. Position of the square), for example, two random numbers from -1.0 to +1.0 are generated, and the coordinates within the square are determined according to the generated random number values. Can be generated randomly. If the generated floor surface coordinates are in the rectangle indicating the person arrangement range described above, the floor surface coordinates are adopted, and if not included in the rectangle, the floor surface coordinates are used. It may be rejected, a random number is generated again, and repeated until a floor surface coordinate within a quadrilateral representing the person arrangement range 40 is generated.

  In this way, for example, when three person images 44 are arranged, three floor surface coordinates included in a rectangle indicating the person arrangement range 40 are calculated.

  Next, a rectangle 42 corresponding to the size of the person (for example, a rectangle having a size similar to the actual size of the person, such as a height of 1.65 m and a width of 0.4 m) 42 is placed in the real space at the position of the generated floor coordinate. Place on top. At this time, as shown in FIG. 2, it is desirable to arrange the rectangle 42 at an angle facing the fixed camera 48. Then, the person image 44 stored in advance in the person image storage unit 12 is fitted into the arranged rectangle 42. At this time, the calibration parameters are used to associate the three-dimensional coordinates of X, Y, and Z on the rectangle 42 with the pixels of the person image 44, and based on this, the person image 44 on the rectangle 42 is converted into an image. Project above. At this time, the pixel value of the person image 44 is transferred onto the background image 50 while calculating x and y on the image from X, Y, and Z on the rectangle 42, thereby projecting the person image onto the image. Thereby, as shown in FIG. 2, a person arrangement image 52 in which a person image 44 is arranged on a rectangle 42 can be obtained.

  At this time, the background portion (portion other than the person) of the person image 44 is assumed to be the background (for example, transparent) and is not transferred onto the background image 50. Like that. Further, regarding the order of transfer of the plurality of person images 44 arranged at a plurality of positions to the background image 50, the distance from the floor coordinate immediately below the fixed camera 48 is calculated for the plurality of person images 44, and the distance is calculated. By drawing from a distant place, it is possible to generate a person arrangement image 52 with the correct context.

  In this way, the number of persons to be arranged, that is, the number of person images 44 arranged on the background image 50 is changed to 0, 1, 2, 3, 10, 100,. An image 52 is generated.

  In addition, as described above, the case where the range in which the rectangle on the background image 50 (the processing range specified by the four corners or the rectangle of the image) is projected on the floor surface is set as the range in which the person is arranged has been described. Absent. For example, when the movement range of a person in real space is limited by structures such as pillars, the range corresponding to structures such as walls and pillars is excluded from the movement range of persons. By determining the person arrangement range on the floor so that the movement range of the person in the space is accurately reflected in the person arrangement image 52, the person arrangement image 52 that matches the structure and characteristics of the place can be generated. .

  Such arrangement is limited by designating an end point (such as a column corner) where the floor and the wall or the lower end of the wall contact on the background image 50 on the background image 50 and obtaining the floor surface coordinates. The positions of pillars, walls, etc. can be projected on the plane coordinates. The person arrangement range on the floor may be limited using this.

  In the present embodiment, the case where the person arrangement image 52 is generated using the two-dimensional person image 44 has been described, but the present invention is not limited to this. For example, instead of the two-dimensional person image 44, a three-dimensional person model (a three-dimensional object for computer graphics (CG)) may be arranged and an image obtained by rendering may be used. Rendering here refers to, for example, generating a two-dimensional image obtained when a three-dimensional human model is illuminated from a predetermined direction.

  Next, the image feature amount calculation unit 22 performs predetermined image processing (feature amount calculation processing) on each of the generated person layout images 52. Here, as the image feature amount, the congestion situation (total number of people in the place, density) of the person may be a numerical value or a multi-dimensional vector that is reflected directly or indirectly. Various texture features have been proposed, but the specific calculation method is not specified. As an example, the “fractal dimension number” proposed in Reference Document 2 (and related research) below may be used as the image feature amount.

[Reference 2] Miyagawa, Arai et al. “Adjusting Camera Parter Using Genetic Algorithm for Estimating the Number of People from Video”, Intelligence and Information: Journal of Japan Society for Fuzzy Theory and Intelligent Informatics 21 (1), pp.41 -55, 2009-2-15.

  The learning feature quantity storage unit 24 uses data (learning feature quantity data) obtained by pairing the image feature quantity calculated for the person layout image 52 and the number of people arranged for learning feature quantities such as a hard disk and media. The data is stored in the data storage unit 26.

  Next, a configuration for performing processing during measurement will be described with reference to the drawings.

  Further, as shown in FIG. 1, the number estimating device 10 according to the present embodiment includes an image input unit 28, an image feature amount calculating unit 30, and a number estimating unit 32 as a configuration for performing processing at the time of measurement. I have.

  At the time of measurement, the image input unit 28 acquires a frame image from a fixed camera, a moving image file captured by the fixed camera, a still image sequence captured by the fixed camera, and the like.

  The image feature amount calculation unit 30 calculates an image feature amount for the image acquired by the image input unit 28, similarly to the image feature amount calculation unit 22 described above.

  The number-of-persons estimation unit 32 includes the image feature amount calculated by the image feature amount calculation unit 30 and the image of the learning feature amount data that is stored in the learning feature amount data storage unit 26 and created in advance. Based on the feature quantity, the number of likely persons is estimated, and an estimation result 34 is output.

  In the present embodiment, as the learning feature value data, data in which the image feature value is associated with the number of persons to be arranged is stored. Therefore, an image feature value that is close to the image feature value of the image at the time of measurement is obtained by proximity search. The learning feature amount data is searched for, and the number of people in the searched neighborhood data is set as the estimated number. At this time, there are a method of creating a tree structure in advance and a method of creating a hashing table in order to speed up the neighborhood search. These processes may be used as appropriate.

  The number estimating device 10 according to the present embodiment is configured by a computer device including, for example, a CPU (Central Processing Unit), a RAM (Random Access Memory), and a ROM (Read Only Memory) that stores various programs. Moreover, the computer which comprises the number-of-people estimation apparatus 10 may be provided with memory | storage parts, such as a hard disk drive and a non-volatile memory. In the present embodiment, the CPU reads and executes a program stored in a storage unit such as a ROM or a hard disk, whereby the hardware resources and the program cooperate to realize the above-described function.

  Here, the processing procedure of the generation process performed at the time of presetting described above and the estimation process performed at the time of measurement will be described with reference to the flowchart shown in FIG.

  In this embodiment, the generation process and estimation process programs are stored in the storage unit in advance, but the present invention is not limited thereto. For example, the program for the generation process and the estimation process may be received from an external device and executed. Further, the generation process and the estimation process may be executed by reading a program for the generation process and the estimation process recorded on a recording medium such as a CD-ROM with a CD-ROM drive or the like.

  First, the generation process performed at the time of presetting will be described.

  In the present embodiment, the generation process program is executed when an instruction to execute the generation process is input.

  In step S101, the image input unit 14 inputs a captured image and acquires a frame image.

  In step S <b> 103, the calibration unit 18 performs camera calibration by acquiring the calibration parameters of the fixed camera 48.

  In step S <b> 105, the background image acquisition unit 16 acquires the background image 50 from the captured image input by the image input unit 14.

  In step S107, the person arrangement image generation unit 20 generates a plurality of person arrangement images 52 using the acquired background image 50 and the person image 44 stored in the person image storage unit 12.

  In step S109, the image feature amount calculation unit 22 calculates an image feature amount from the generated person layout image 52.

  In step S111, the learning feature quantity storage unit 24 stores the calculated image feature quantity in the learning feature quantity data storage unit 26 in association with the number of persons arranged in the person arrangement image 52, thereby generating the main generation. Terminates execution of the processing program.

  Next, the estimation process performed at the time of measurement is demonstrated.

  In the present embodiment, the estimation process program is executed when an instruction to execute the estimation process is input.

  In step S201, the image input unit 28 inputs a captured image captured by the fixed camera 48, and acquires a frame image from the input captured image.

  In step S203, the image feature amount calculation unit 30 calculates the image feature amount of the acquired frame image.

  In step S205, the place where the number of people estimation unit 32 is photographed by the fixed camera 48 from the learning feature amount data stored in the learning feature amount data storage unit 26 and the calculated image feature amount of the frame image. Estimate the number of people who appeared in

  In step S207, the number estimating unit 32 determines whether or not the end condition is satisfied. The termination condition here is, for example, that the processes in steps S203 to S205 have been performed on all the frame images acquired in step S201. If it is determined in step S207 that the termination condition is not satisfied (S207, N), the process proceeds to step S201. If it is determined in step S207 that the termination condition is satisfied (S207, Y), the execution of the program of the estimation process is terminated.

  As described above, in the present embodiment, the calibration parameter indicating the geometric relationship between the fixed camera 48 and the real space coordinate system is measured by the calibration unit 18 at the time of presetting. In addition, the background image acquisition unit 16 acquires a background image 50 that is an image taken by the fixed camera 48 in an unattended state. In addition, the person arrangement image generation unit 20 generates a person arrangement image 52 that is an image in which the person image 44 prepared in advance is arranged on the background image 50 while referring to the calibration parameters. Further, the image feature amount calculation unit 22 calculates an image feature amount from the generated person arrangement image 52, and the learning feature amount storage unit 24 indicates a relationship between the calculated image feature amount and the number of persons arranged. The feature amount data is stored in the learning feature amount data storage unit 26.

  Furthermore, in the present embodiment, a captured image captured by the fixed camera 48 is input by the image input unit 28 during measurement. Further, the image feature amount calculation unit 30 calculates an image feature amount from the input captured image. In addition, the number of persons estimated in the photographed image is estimated by the number estimating unit 32 based on the calculated image feature amount and the learning feature amount data stored in the learning feature amount data storage unit 26. .

  This eliminates the need to acquire learning data for each fixed camera 48 at the time of pre-setting, so that it is possible to execute the number-of-people estimation process in a form that greatly reduces the labor and cost at the time of pre-setting. Become. In particular, learning data, a large amount of images that actually contain people, and the processing to measure and store the number of people in the images are reduced, which greatly reduces labor and cost. Can be reduced.

  In the present embodiment, the learning feature value data storage unit uses, as learning feature value data, a pair of the image feature value calculated for the person placement image 52 and the number of people arranged in the person placement image 52. Although the case where it memorize | stores in 26 was demonstrated, it is not restricted to this.

  For example, as shown in FIG. 4, a learning unit 36 may be provided in the learning feature amount storage unit 24. The learning unit 36 uses the learning feature amount data to estimate the relationship between the number of arranged persons and the image feature amount, generates a regression equation, and obtains parameters of the generated regression equation. Further, the learning feature quantity storage unit 24 causes the learning feature quantity data storage unit 26 to store a regression equation to which the parameter obtained by the learning unit is applied.

  As a specific regression equation, various methods such as a straight line and approximation to a polynomial can be used, and the method is not limited here.

  In addition, when the above-described regression equation is stored in the learning feature data storage unit 26, the number estimation unit 32 uses the image feature calculated from the captured image captured by the fixed camera 48 at the time of measurement. It is better to calculate the estimated number of people by entering the regression equation.

  In the present embodiment, the case where the person image 44 prepared in advance is used when the person arrangement image 52 is generated has been described. However, the present invention is not limited to this. Depending on the measurement location, the profile of the person appearing at the location photographed by the fixed camera 48 may be different. For example, in offices, commercial facilities, etc., there are places where there is a large bias with respect to the profile (gender, age, clothes, etc.) of the appearing person, such as a place with many male office workers, a place with many students, and a place with many shoppers. In addition, there are places where there is a large bias in the profile of the appearing person depending on the day of the week, the season, the time zone, and the like.

  In such a case, when the profile of the person appearing at each location is different, a plurality of types of person images 44 are stored in the person image storage unit 12 and the person image 44 is selected according to the profile of the person appearing. By doing so, more accurate estimation becomes possible. For example, information indicating at least one of the gender, age, and clothes of a person shown in the person image 44 is added to the person image 44, and at least gender, age, and clothes are matched with the profile of the person who appears. The person image 44 may be selected based on one. In addition, when the person appearing according to the day of the week, the season, or the time zone is different, it can be dealt with by switching the learning feature data for each day of the week, season, and time zone.

  In addition, the number estimation device 10 according to the present embodiment includes a person image detection recording unit that detects a person in a captured image captured by the fixed camera 48 and stores the person image 44 in the person image storage unit 12. May be. The person image detection / recording unit sequentially collects captured images obtained by capturing a person appearing at a place captured by the fixed camera 48, detects a person appearing in the captured image, and stores it as a person image 44 in the person image storage unit 12. Remember.

  Thereby, the person arrangement image 52 can be generated using the person image 44 that matches the lighting condition of the place and reflects the profile of the person appearing in the place. In addition, it is expected to improve the accuracy of estimating the number of people. In addition, as a method for detecting a person in a moving image, a combination of discriminators such as SVM (Support Vector Machine) using HOG (Histograms of Oriented Gradients) features has been generalized. Also good. In addition, it becomes possible to collect human images more stably by combining tracking techniques.

  At this time, when the degree of congestion is high, there is a high possibility that surrounding persons will be mixed as noise in the person image. As a result of the above-described number estimation processing and person detection by HOG, there are few persons in the image. Only when it is assumed, it is necessary to devise such as collecting human images. In addition, by storing in combination with the date, time zone, and the upper part of the day of the week when collecting human images, it can be applied to changes due to the above-mentioned time zone, season, etc.

  In the present embodiment, the case where the image input unit 14 for the generation process at the time of presetting and the image input unit 28 for the estimation process at the time of measurement are provided separately has been described, but the present invention is not limited thereto. . For example, an image input unit that is commonly used in the generation process at the time of presetting and the estimation process at the time of measurement may be provided.

  In the present embodiment, the case has been described in which the image feature amount calculation unit 22 for the generation process at the time of presetting and the image feature amount calculation unit 30 for the estimation process at the time of measurement are provided separately. Not limited to this. For example, you may provide the image feature-value calculation part used in common by the production | generation process at the time of prior setting, and the estimation process at the time of measurement.

  Although the present embodiment has been described in detail with reference to the drawings, the specific configuration is not limited to the present embodiment, and includes design and the like within a scope not departing from the gist of the present invention.

  Further, in the present embodiment, the operation of the components of the functions shown in FIG. 1 is constructed as a program, and is installed and executed on a computer used as the number estimating device 10, but is not limited thereto, and is distributed via a network. You may let them.

  Further, the constructed program may be stored in a portable storage medium such as a hard disk, a flexible disk, or a CD-ROM, and installed in a computer or distributed.

DESCRIPTION OF SYMBOLS 10 Person estimation apparatus 12 Person image memory | storage part 14, 28 Image input part 16 Background acquisition part 18 Calibration part 20 Person arrangement | positioning image generation part 22, 30 Image feature-value calculation part 24 Learning feature-value preservation | save part 32 Number-of-person estimation part 34 Estimation Result 36 Learning Department

Claims (8)

A number estimation method in a number estimation device including a calibration unit, a background image acquisition unit, a person arrangement image generation unit, a learning feature amount storage unit, an image input unit, an image feature amount calculation unit, and a number estimation unit,
At the time of presetting, the calibration unit measures a calibration parameter indicating a geometric relationship between the fixed camera and the real space coordinate system; and
The background image acquisition unit acquires a background image that is an image taken by the fixed camera in an unattended state; and
A person arrangement image generating step in which the person arrangement image generation unit generates a person arrangement image that is an image in which a person image prepared in advance is arranged on the background image with reference to the calibration parameter;
The learning feature amount storage unit calculates an image feature amount from the generated person arrangement image, and stores learning feature amount data indicating a relationship between the calculated image feature amount and the number of people arranged. A save step;
An image input step in which the image input unit inputs a captured image captured by the fixed camera at the time of measurement;
An image feature amount calculating unit that calculates the image feature amount from the input captured image;
Step of estimating the number of persons in which the number of persons estimating section estimates the number of persons in the captured image based on the calculated image feature quantity and the learning feature quantity data stored by the learning feature quantity storage section. When,
How to estimate the number of people. 2. In the number estimating step, the number estimating unit performs a neighborhood search on the feature amount of the learning feature amount data for the image feature amount calculated at the time of measurement, and estimates the number of people based on a search result. Number of people estimation method of mention. The number-of-people estimation device includes a learning unit,
At the time of presetting, the learning unit further performs a learning step of learning a regression equation indicating a relationship between the image feature amount and the number of people using the learning feature amount data,
The said number estimation part estimates the said number of persons by inputting the image feature-value calculated at the time of a measurement into the said regression equation learned by the said learning part in the said number of persons estimation step. Number of people estimation method. In the person arrangement image generation step, the person arrangement image generation unit selects the person image to be arranged by at least one of gender, age, and clothes based on a profile relating to a person appearing in a place photographed by the fixed camera. The number-of-people estimation method according to any one of claims 1 to 3. The number-of-people estimation device includes a person image detection recording unit,
The person image detection recording unit further performs a person image detection recording unit that detects a person in a captured image captured by the fixed camera and records the person image as the person image,
5. The number of persons according to claim 1, wherein, in the person arrangement image generation step, the person arrangement image generation unit arranges the person image recorded in the person image detection recording unit in the person arrangement image. Estimation method. A calibration unit that measures a calibration parameter indicating a geometric relationship between the fixed camera and the real space coordinate system at the time of presetting;
A background image acquisition unit that acquires a background image that is an image taken by the fixed camera in an unattended state;
A person arrangement image generation unit that generates a person arrangement image that is an image in which a person image prepared in advance is arranged on the background image while referring to the calibration parameter;
A learning feature quantity storage unit that calculates an image feature quantity from the generated person arrangement image and stores learning feature quantity data indicating a relationship between the calculated image feature quantity and the number of arranged persons;
An image input unit that inputs a captured image captured by the fixed camera at the time of measurement;
An image feature amount calculation unit for calculating the image feature amount from the input captured image;
Based on the calculated image feature amount and the learning feature amount data stored by the learning feature amount storage unit, a number estimation unit that estimates the number of persons in the captured image;
Number estimation device including   A number estimation program for causing a computer to execute each step of the number estimation method according to any one of claims 1 to 6. A computer-readable recording medium recording a number estimation program for causing a computer to execute each step of the number estimation method according to claim 1.