JP2013246490A - Estimation device, estimation method and computer program - Google Patents

Abstract

An object of the present invention is to reduce errors in the amount of real space movement caused by occlusion.
An estimation apparatus that uses a captured image to calculate the number of persons in the area where the number of persons in the area where the subject is observed is measured, and the amount of movement of the person on the image from the captured image And a corrected estimated value acquiring unit that acquires a value smaller than the actual space moving amount estimated from the moving amount on the image as an estimation result of the actual space moving amount. The acquisition unit calculates a real space movement amount estimation unit that estimates the real space movement amount from the movement amount on the image, and calculates a correction coefficient that corrects an error of the real space movement amount estimated by the real space movement amount estimation unit. And a real space movement amount correction unit that corrects the real space movement amount estimated by the real space movement amount estimation unit using the correction coefficient.
[Selection] Figure 2

Description

  The present invention relates to a technique for estimating the number of people passing by.

In order to measure the number of passing people using a camera image, it is necessary to detect that a person has appeared on the image and that a person has passed. Various image processing techniques for realizing these detections have been proposed, and can be roughly divided into two approaches.
The first approach is based on a person tracking technique (tracking technique) in which an individual person existing on an image is detected and its position is captured and tracked (see Non-Patent Document 1). However, the person tracking technology has a problem that it becomes difficult to detect and track individual persons due to occurrence of occlusion. Such a problem occurs, for example, when a person tries to be applied to a place where a lot of people are crowded, such as a concourse of a large station, and people overlap each other on an image.

  The second approach is a method in which individual persons are not detected but the whole group of persons is regarded as a group like a fluid (see Non-Patent Document 2). In this method, the number of passing people is estimated by calculating the number of people in the region at the moment and the movement amount (movement speed) of the entire group within a certain time. Such an approach that does not require the detection of individual persons can be expected to operate stably even when it is congested.

An outline of a specific algorithm of the existing method described in Non-Patent Document 2 will be briefly described. In this method, while acquiring images in a time series using a calibrated fixed camera (a camera in which so-called calibration is performed so that it can be specified where each pixel of the camera is shooting in real space). The following values are calculated.
1. Number of people in a given area in the instantaneous image (number of people in the area)
2. Amount of movement of a group in real space between real time adjacent frame images (or between two frame images in a short time interval such as 0.1 second) (real space movement amount)

Note that camera calibration consists of internal parameters that describe the characteristics of the optical system of the camera (focal length, aspect ratio, distortion parameters, etc.) and external parameters that specify the position and angle of the camera in real space. Is to get.
Calibration is performed by photographing a subject whose geometric properties (shape or size) are known, such as a subject such as a dot pattern or chessboard pattern, or a specific three-dimensional structure, and performing a predetermined geometric calculation. Can be realized. Various specific methods for realizing calibration have already been proposed, and they are generalized such as being released in the form of open software such as OpenCV.

Various methods already proposed can be applied to calculate the number of people in the area. For example, the number of people in the area may be calculated using the technique described in Non-Patent Document 3.
Various methods that have already been proposed can be applied to the calculation of the real space movement amount. For example, the real space movement amount may be calculated using the technique described in Non-Patent Document 2.
Then, the number of passing people is estimated using the number of people in the area and the real space movement amount. In this method, when a fluid with a certain density flows through a known flow path (such as a pipe with a known cross-sectional area) at a certain speed, the flow rate can be calculated if there is density, speed, and geometric information of the flow path. The same principle is used.

Tatsuya Osawa, 5 others, “Person Tracking Using 3D Model for Object and Environment Based on MCMC Method”, IEICE Transactions. D, Information and Systems, The Institute of Electronics, Information and Communication Engineers, August 2008 1st, J91-D, No.8 p.2137-2147 Isamu Igarashi, 3 others, “Measurement of the number of people passing from video using collective movement based on voting regarding local movement”, IS3-03, 2011 Image Sensing Symposium, 2011 Hiroyuki Arai and three others, “Estimating the number of people from images based on geometric models”, The Institute of Image Information and Television Engineers Technical Report, June 24, 2008, 32, 26, p.33-36

  In the method of Non-Patent Document 2 described above, the entire group of persons is regarded as a fluid, and the number of people passing through is estimated by calculating the number of people in the area at the moment and the real space movement amount of the entire group. However, since the feature point of the lower body such as a person's foot is not detected by the occlusion, only the feature point of the upper body such as the person's head is detected, which causes an error in estimating the real space movement amount. There was a thing.

  In view of the above circumstances, an object of the present invention is to provide a technique for reducing an error in estimating a real space movement amount caused by occlusion.

  According to an aspect of the present invention, an area number measurement unit that measures the number of persons in an area where a subject is observed using a captured image, and an image that calculates a movement amount of the person on the image from the captured image An estimation apparatus comprising: an upper movement amount calculation unit; and a corrected estimated value acquisition unit that acquires a value smaller than the actual space movement amount estimated from the upper image movement amount as an estimation result of the actual space movement amount.

  One aspect of the present invention is the above estimation apparatus, wherein the correction estimated value acquisition unit estimates a real space movement amount from the movement amount on the image, and the real space movement amount estimation. A correction coefficient calculation unit that calculates a correction coefficient that corrects an error in the real space movement amount estimated by the unit, and a real space that corrects the real space movement amount estimated by the real space movement amount estimation unit using the correction coefficient A movement amount correction unit.

  One aspect of the present invention is the above-described estimation device, wherein the correction estimated value acquisition unit calculates a voting coefficient for each height using the number of people in the region and the captured image. A voting coefficient calculation unit, a weighting voting unit that weights the movement amount on the image using the voting coefficient for each height and weights the real space movement amount, and the real space movement amount is estimated by the weighted voting. A real space movement amount voting estimation unit.

  One aspect of the present invention is the above-described estimation device, in which an imaging region is determined using the real space movement amount acquired by the correction estimated value acquisition unit and the number of people in the region of the captured image. A passing number estimating unit for estimating the number of passing persons is further provided.

  According to one aspect of the present invention, an area number measurement step for measuring the number of persons in an area where a subject is observed using a captured image, and an image for calculating a movement amount of the person on the image from the captured image. The estimation method includes an upper movement amount calculation step and a corrected estimated value acquisition step of acquiring a value smaller than the actual space movement amount estimated from the upper image movement amount as an estimation result of the actual space movement amount.

  According to one aspect of the present invention, an area number measurement step for measuring the number of persons in an area where a subject is observed using a captured image, and an image for calculating a movement amount of the person on the image from the captured image. An upper moving amount calculating step, and a corrected estimated value acquiring step for acquiring a value smaller than the actual space moving amount estimated from the moving amount on the image as an estimation result of the actual space moving amount; It is a computer program.

  According to the present invention, it is possible to reduce the error of the real space movement amount estimation caused by occlusion.

It is a figure which shows the system configuration | structure of the passage number estimation system in this invention. It is a schematic block diagram showing the functional composition of a first embodiment (passage number estimating device 100a) of passing number estimating device 100. It is a figure which shows the specific example of the moving amount on an image. It is a figure which shows the specific example of the process which estimates the real space movement amount based on the movement amount on an image. It is a figure which shows the specific example of the process which estimates the real space movement amount. It is the figure showing the specific example which estimates the real space movement amount according to the number of votes obtained. It is a figure showing the specific example of a rectangular range. It is a figure showing the concept of the expected value of the height of a person's body. It is a figure showing expected value H1 (R) of the height of a body when R persons exist in rectangular range 306. It is a flowchart which shows the flow of a process of the passage number estimation method in 1st embodiment of this invention. It is the figure which showed the specific example of the process which estimates real space movement amount using the frame image image | photographed at the time of a quiet time. It is the figure which showed the specific example of the process which estimates real space movement amount using the frame image image | photographed at the time of congestion. It is a schematic block diagram showing the functional composition of a second embodiment (passage number estimating device 100b) of passing number estimating device 100. It is a figure showing the specific example of the voting coefficient P (h) for every height. It is a figure which shows the specific example of a process of the real space movement amount vote estimation part. It is a flowchart which shows the flow of a process of the passage number estimation method in 2nd embodiment of this invention.

Hereinafter, an embodiment of the present invention will be described with reference to the drawings.
FIG. 1 is a diagram showing a system configuration of a passing person estimation system according to the present invention. The passing person estimation system of the present invention includes a fixed camera 10 and a passing person estimation device 100.
The fixed camera 10 is installed obliquely downward like a security camera, for example, and images the imaging region 20. The imaging region 20 that the fixed camera 10 captures is determined by prior camera calibration.

The passing person estimation device 100 is configured using an information processing device. The passing number estimating device 100 estimates the number of people who have passed through the shooting area 20 by performing image processing based on the frame image shot by the fixed camera 10.
The shooting area 20 is an area shot by the fixed camera 10.
Hereinafter, a specific configuration example (first embodiment and second embodiment) of the passing person estimation device 100 will be described.

[First embodiment]
FIG. 2 is a schematic block diagram showing the functional configuration of the first embodiment (passage number estimating device 100a) of the passing number estimating device 100. The passing person estimation device 100a includes a CPU (Central Processing Unit), a memory, an auxiliary storage device, and the like connected by a bus, and executes a passing person estimation program. By executing the passing number estimating program, the passing number estimating device 100a is configured to include the frame image input unit 101, the in-region number measuring unit 102, the on-image movement amount calculating unit 103, the camera parameter storage unit 104, the passing number estimating unit 108, and the correction estimation. It functions as an apparatus including the value acquisition unit 109. The correction estimated value acquisition unit 109 functions as a real space movement amount estimation unit 105, a correction coefficient calculation unit 106, and a real space movement amount correction unit 107. Note that all or part of the functions of the passing person estimation device 100a may be realized using hardware such as an application specific integrated circuit (ASIC), a programmable logic device (PLD), or a field programmable gate array (FPGA). good. The passing number estimation program may be recorded on a computer-readable recording medium. The computer-readable recording medium is, for example, a portable medium such as a flexible disk, a magneto-optical disk, a ROM, a CD-ROM, or a storage device such as a hard disk built in the computer system. Further, the passing number estimation program may be transmitted / received via a telecommunication line.

The frame image input unit 101 inputs frame images taken by the fixed camera 10 in chronological order.
The in-area number counting unit 102 measures the number of persons in a predetermined area in the frame image input by the frame image input unit 101. The predetermined area is an area where a person is observed in the frame image.

The on-image movement amount calculation unit 103 calculates the on-image movement amount of a person in the frame image input by the frame image input unit 101 in chronological order. The amount of movement on the image represents the amount of movement of the person on the image between two frame images that are close in time.
The camera parameter storage unit 104 is configured using a storage device such as a magnetic hard disk device or a semiconductor storage device. The camera parameter storage unit 104 stores camera parameters (for example, internal parameters and external parameters) set by calibration of the fixed camera 10. The internal parameter is a parameter related to the configuration of the camera. For example, the internal parameter represents a focal length, an aspect ratio, a distortion parameter, and the like. External parameters are parameters relating to the position (including height) and orientation of the fixed camera 10 in real space.

The corrected estimated value acquisition unit 109 acquires a value smaller than the actual space movement amount estimated by the conventional method based on the on-image movement amount as an estimation result of the actual space movement amount. Hereinafter, a specific configuration example of the corrected estimated value acquisition unit 109 will be described.
The real space movement amount estimation unit 105 estimates the real space movement amount using the camera parameters stored in the camera parameter storage unit 104 and the image movement amount calculated by the image movement amount calculation unit 103. The real space movement amount represents the movement amount of the person in the real space between two frame images that are close in time. The real space movement amount estimation method will be described with reference to FIGS.

  FIG. 3 is a diagram illustrating a specific example of the movement amount on the image. The frame image 300-1 is an image captured when a group of people is moving in real space at the same speed. The frame image 300-2 is an image captured after a certain time (for example, 5 seconds, 10 seconds, etc.) has elapsed from the frame image 300-1. The image 300-3 represents the movement of the group of persons on the image as an optical flow by associating the feature points of the group of persons detected in the frame images 300-1 and 300-2. It is an image. In the image 300-3, the length of each arrow (optical flow) represents the amount of movement on the image, and the direction of the arrow represents the movement direction.

  As is apparent from the image 300-3, the amount of movement on the image differs depending on the location of the person in the real space. The amount of movement on the image observed for a person who is close to the fixed camera 10 in real space is larger than the amount of movement on the image observed for a person far from the fixed camera 10. Thus, even if the movement amount in the real space is the same, the movement amount on the image differs depending on the place where the person is in the real space.

  FIG. 4 is a diagram illustrating a specific example of processing for estimating the real space movement amount based on the movement amount on the image. A frame image 301-1 in FIG. 4A is a specific example of an optical flow detected in two inter-frame images. At this time, the movement amount in the real space represented by the optical flow of the frame image 301-1 is unknown. Furthermore, it is also unknown which optical point of the person the characteristic flow of the frame image 301-1 is. For this reason, the optical flow of the frame image 301-1 may be an optical flow in the head portion of the person or an optical flow in the foot portion of the person.

However, as described above, the amount of movement in the real space represented by the length of the optical flow (the amount of movement on the image) varies depending on the distance that the feature point is away from the fixed camera 10 in the real space. In general, the head portion of a person is closer to the fixed camera 10 than the foot portion of the person. Therefore, the amount of movement in real space estimated from the optical flow is assumed to be the optical flow of the person's head and the optical flow of the person's foot. It depends on the case.
A frame image 301-2 in FIG. 4B is an image representing a state when the feature point represented by the optical flow of the frame image 301-1 is assumed to be the head of a person. A frame image 301-3 in FIG. 4C is an image representing a state when the feature point represented by the optical flow of the frame image 301-1 is assumed to be a human foot. Even for the same optical flow, the movement amount in the real space estimated when the human foot is assumed is larger than the movement amount in the real space estimated when the human head is assumed.

  FIG. 4D is a diagram illustrating a specific example of the real space movement amount estimated from the movement amount on the image. The horizontal axis represents the coordinates of the movement direction of the person in real space. An arrow 313 indicates the movement of the person represented by the optical flow when it is assumed that the optical flow in FIG. 4A is an optical flow at the person's head. That is, the arrow 313 represents the state of movement shown in FIG. An arrow 314 indicates the movement of the person represented by the optical flow when it is assumed that the optical flow in FIG. That is, the arrow 314 represents the movement state shown in FIG. As described above, even in the case of one optical flow on the image, if the feature portion corresponding to the optical flow is different, the movement in the real space represented by the optical flow is also different.

  FIG. 5 is a diagram illustrating a specific example of processing for estimating the real space movement amount. FIG. 5A is a diagram illustrating a specific example of the real space movement amount estimated from one optical flow. H represents the height of a person's body. rH represents a real space movement amount estimated when a person's head is used as a feature point. r0 represents a real space movement amount estimated when a person's foot is a feature point.

  FIG. 5B is a diagram illustrating the concept of voting performed when the real space movement amount is estimated. As is clear from FIG. 5A, the part closest to the camera among the parts of the person photographed by the fixed camera 10 in the real space is the head of the person. Therefore, it can be assumed that rH when the person's head is assumed as the feature point is the minimum value of the real space movement amount. Further, the part farthest from the fixed camera 10 in the real space is a person's foot (toe). Therefore, it is possible to assume that r0 when a person's foot is assumed to be a feature point is the maximum value of the real space movement amount. V represents the estimated range of the real space movement amount of the person observed by the fixed camera 10. The real space movement amount estimation unit 105 performs voting in the real space movement amount estimation range, and estimates the real space movement amount. Voting is a process of adding one vote for each predetermined width in the estimated range of the real space movement amount estimated from each optical flow.

  FIG. 5C is a diagram illustrating a specific example of the optical flow obtained for the three feature points of the person. An optical flow 400-3 represented by a broken line is an optical flow obtained using a person's head as a feature point. An optical flow 400-2 represented by an alternate long and short dash line is an optical flow obtained using a central portion (around the arm) in the height direction of a person as a feature point. An optical flow 400-1 represented by a two-dot chain line is an optical flow obtained using a person's foot as a feature point. In actual processing, it is unclear which part of the person each of the three optical flows is an optical flow obtained as a characteristic. For ease of understanding, this description clarifies which part of the person each optical flow is obtained as a feature point.

  FIG. 5D is a diagram illustrating a result of voting performed based on the three optical flows illustrated in FIG. In FIG. 5D, the vertical axis represents the arrangement of optical flows, and the horizontal axis represents the real space movement amount M. 401-1 represents the result of voting performed based on the optical flow 400-1. 401-2 represents the result of voting performed based on the optical flow 400-2. 401-3 represents the result of the voting performed based on the optical flow 400-3.

  FIG. 6 is a diagram showing a specific example in which the real space movement amount is estimated according to the number of votes obtained. The vertical axis represents the number of votes T. The horizontal axis represents the real space movement amount M. The real space movement amount estimation unit 105 acquires an estimation range of the real space movement amount for each optical flow obtained from the frame image. Thereafter, the real space movement amount estimation unit 105 votes equally for the estimation range for each optical flow. The real space movement amount estimation unit 105 acquires the real space movement amount Mmax having the highest number of votes as an estimation result.

  The correction coefficient calculation unit 106 calculates the correction coefficient C using the camera parameters stored in the camera parameter storage unit 104 and the number of people in the predetermined area measured by the in-area number measurement unit 102. A specific method for calculating the correction coefficient C will be described with reference to FIGS.

  FIG. 7 is a diagram illustrating a specific example of a rectangular range. Hereinafter, the rectangular range will be described with reference to FIG. Tz represents the height from the floor surface to the fixed camera 10. G represents a frame of an image of the imaging region 20 that is captured by the fixed camera 10. G1 represents the target pixel of the image G. The target pixel G <b> 1 can specifically calculate which region is being imaged by calibration of the fixed camera 10.

  P in FIG. 7 represents a space captured by the target pixel G1 with a straight line. Strictly speaking, the straight line P has a conical shape that spreads farther away, but is represented by a straight line for simplification. Reference numerals 305-1 to 305-4 represent a person approximated by a rectangle having a standard size (W is a width and H is a body height). The rectangles 305-1 to 305-4 are approximated so that the rectangles 305-1 to 305-4 are viewed from the fixed camera 10 so as to be viewed as a rectangle having a width W and a body height H facing the fixed camera 10. .

  A rectangular area 306 on the floor surface represents a range in which a part of the body of the person (rectangle) appears in the target pixel G1 when the person stands on the floor surface. A rectangular area 306 is an area having a width W and a length L. The length L is calculated from the inclination of P with respect to the floor surface and the height H of the rectangles 305-1 to 305-4.

  Based on the number of people in the rectangular range 306 when a subject is observed at the target pixel G1, and the camera parameters stored in the camera parameter storage unit 104, the person density in the shooting region 20 (σ people per square meter) Can be calculated. Since the area of the rectangular range 306 is W · L, the expected value of the number of people in the rectangular range 306 is R = σ · W · L.

FIG. 8 is a diagram illustrating a concept of an expected value of the height of a person's body. The expected value of the height of the body represents the height from the floor surface of the body part of the person located closest to the fixed camera 10 that appears in the target pixel G1. The expected value of the height of the body changes according to the number R of persons existing in the rectangular range 306. Therefore, the expected value of the height of the body is expressed as H1 (R).
In FIG. 8, the vertical axis represents the height, and the horizontal axis represents the coordinate in the length direction (L) of the rectangular range 306. FIG. 8 shows a situation where one person is observed in the rectangular area 306. H represents the height of a person's body. The straight lines P1 and P2 are straight lines connecting the viewpoint of the fixed camera 10 and the target pixel G1. When the person is present at the position X, the part of the person shown in the target pixel G1 is a part located at the height H1 from the floor.

  L1 is the position closest to the fixed camera 10 within the rectangular range 306. When the person is located at L1, the head of the person is shown in the target pixel G1. L2 is a position farthest from the fixed camera 10 within the rectangular range 306. When the person is positioned at L2, the foot of the person appears in the target pixel G1. When the number of people observed by the fixed camera 10 is one, the probability that there is a person located closest to the camera is even in the range from L1 to L2. Therefore, the expected value H1 (1) = H / 2 of the height of the body.

  FIG. 9 is a diagram illustrating the expected value H1 (R) of the body height when there are R persons within the rectangular range 306. In FIG. The probability that the person located at X in the rectangular area 306 is located closest to the fixed camera 10 is the probability P that the person located at X is located closer to the fixed camera 10 than all the other (R-1) people. It is the same value as (X) and is expressed as in equation (1).

  When the R person is in the rectangular range, the expected value X1 representing the position of the person closest to the fixed camera 10 is expressed as in Expression (2) based on Expression 1.

  By transforming Equation (2), the expected value X1 = L / (R + 1).

  Of the human body part located at the expected value X1, the height of the part shown in the target pixel G1 is the expected value H1 (R) of the body height when there are R persons within the rectangular range 306. It corresponds to. H1 (R) is a right triangle similar to a right triangle whose base is L and height is H, and corresponds to the height of a right triangle whose base is (L-X1). Therefore, H1 (R) is expressed as in Equation (3).

  Based on the results described above, the correction coefficient C is expressed as in Equation (4).

  The real space movement amount correction unit 107 corrects the real space movement amount using the correction coefficient C. The corrected real space movement amount Dc is expressed as shown in Equation (5).

D in Expression (5) is the real space movement amount estimated by the real space movement amount estimation unit 105.
The passing number estimation unit 108 stores the number of people in the predetermined area measured by the in-area number measurement unit 102, the real space movement amount corrected by the real space movement amount correction unit 107, and the camera parameter storage unit 204. The number of passing people between frame images is estimated using the camera parameters.

FIG. 10 is a flowchart showing the flow of processing of the passing person estimation method in the first embodiment of the present invention.
First, the frame image input unit 101 inputs images in chronological order (step S101). Next, the in-area number counting unit 102 measures the number of persons in the predetermined area from the frame image input by the frame image input unit 101 (step S102).

  The on-image movement amount calculation unit 103 detects a feature point of a person from two temporally adjacent frame images input by the frame image input unit 101. The on-image movement amount calculation unit 103 detects the feature points of the person from the two frame images, associates the feature points of the two frame images, and acquires an optical flow. The on-image movement amount calculation unit 103 calculates the on-image movement amount of the person on the image based on the acquired optical flow (step S103).

  Next, the real space movement amount estimation unit 105 performs voting based on the image movement amount calculated by the image movement amount calculation unit 103 (step S104). The real space movement amount estimation unit 105 estimates the real space movement amount based on the vote result of the movement amount on the image (step S105).

  The correction coefficient calculation unit 106 calculates a correction coefficient using the camera parameters stored in the camera parameter storage unit 104 and the number of people in the predetermined area measured by the in-area number measurement unit 102 (step S106).

The real space movement amount correction unit 107 corrects the real space movement amount estimated by the real space movement amount estimation unit 105 using the correction coefficient calculated by the correction coefficient calculation unit 106 (step S107).
The passing person estimation unit 108 includes the number of people in a predetermined area measured by the in-area number measurement unit 102, the real space movement amount corrected by the real space movement amount correction unit 107, and the camera stored in the camera parameter storage unit 104. The number of people passing between the frame images is estimated using the parameters (step S108).

Next, the effect of the passing number estimating device 100a in the first embodiment will be described using FIG. 11 and FIG.
FIG. 11 is a diagram showing a specific example of processing for estimating the real space movement amount using a frame image taken in a quiet time. FIG. 11A is a specific example of a frame image taken at a quiet time. FIG. 11B is a diagram illustrating a specific example of a voting result performed based on a frame image taken in a quiet time. As shown in FIG. 11A, since there are a small number of people in a predetermined area when it is quiet, people are rarely photographed by overlapping each other. Therefore, it is possible to obtain the amount of movement on the image from various height portions of the person's body. Therefore, voting can be performed based on the movement amount on the image obtained from various height portions, and the real space movement amount can be estimated with high accuracy.

FIG. 12 is a diagram illustrating a specific example of a process of estimating the real space movement amount using a frame image taken at the time of congestion. FIG. 12A is a specific example of a frame image taken at the time of congestion. FIG. 12B is a diagram showing a specific example of a voting result performed based on a frame image taken at the time of congestion. In the frame image taken at the time of congestion, occlusion occurs frequently, and people overlap each other in many parts. For this reason, the lower body of the person is hardly photographed, and a large amount of the upper body of the person is photographed. Therefore, voting will be performed based on the amount of movement on the image obtained from the feature points of the upper body. In general, the estimation result of the real space movement amount based on the movement amount on the image obtained from the feature point of the upper body becomes a value larger than the actual real space movement amount. For this reason, the accuracy of the estimation result of the real space movement amount decreases.
For such a problem, the corrected estimated value acquisition unit 109 of the passing person estimation device 100a uses a value smaller than the real space movement amount estimated by the conventional method based on the movement amount on the image as the real space movement amount. Obtained as an estimation result. Therefore, the real space movement amount can be accurately estimated from the frame image in which occlusion occurs. More specifically, the real space movement amount is corrected by multiplying the real space movement amount by a correction coefficient having a value smaller than 1. Therefore, it is possible to acquire the corrected real space movement amount as a highly accurate estimation result.

[Second Embodiment]
FIG. 13 is a schematic block diagram illustrating a functional configuration of the second embodiment (passage number estimation device 100b) of the passage number estimation device 100. The passing person estimation device 100b includes a CPU, a memory, an auxiliary storage device, and the like connected by a bus, and executes a passing person estimation program. By executing the passing number estimating program, the passing number estimating apparatus 100b is configured to include the frame image input unit 201, the in-region number measuring unit 202, the on-image movement amount calculating unit 203, the camera parameter storage unit 204, the passing number estimating unit 208, and the correction estimation. It functions as an apparatus including the value acquisition unit 209. The corrected estimated value acquisition unit 209 functions as a height-by-height voting coefficient calculation unit 205, a weighting voting unit 206, and a real space movement amount voting estimation unit 207. Note that all or part of the functions of the passing number estimating device 100b may be realized using hardware such as ASIC, PLD, or FPGA. The passing number estimation program may be recorded on a computer-readable recording medium. The computer-readable recording medium is, for example, a portable medium such as a flexible disk, a magneto-optical disk, a ROM, a CD-ROM, or a storage device such as a hard disk built in the computer system. Further, the passing number estimation program may be transmitted / received via a telecommunication line.

The frame image input unit 201 inputs frame images taken by the fixed camera 10 in chronological order.
The in-area number counting unit 202 measures the number of persons in a predetermined area in the frame image input by the frame image input unit 201. The predetermined area is an area where a person is observed in the frame image.

  The on-image movement amount calculation unit 203 calculates the on-image movement amount of the person in the frame image input by the frame image input unit 201 in chronological order. The amount of movement on the image represents the amount of movement of the person on the image between two frame images that are close in time.

The camera parameter storage unit 204 is configured using a storage device such as a magnetic hard disk device or a semiconductor storage device. The camera parameter storage unit 204 stores camera parameters (for example, internal parameters and external parameters) set by calibration of the fixed camera 10.
The voting coefficient calculation unit 205 for each height uses the camera parameters stored in the camera parameter storage unit 204 and the number of people in the predetermined area measured by the in-area number measuring unit 202 to determine the voting coefficient P ( h) is calculated. A specific method for calculating the voting coefficient P (h) for each height will be described.

  As shown in FIG. 8, when the person is at the position X, the height (here, h) of the portion of the person shown in the target pixel G1 is obtained based on the ratio of the right triangle as described above. . The height h of the person's body corresponding to the position of X is expressed as in Equation (6).

  A voting coefficient P (h) for each height is obtained based on the equations (1) and (6). P (h) represents the probability when the height of the portion of the person shown in the target pixel G1 is h. Therefore, P (h) is a value between 0 and 1.

  FIG. 14 is a diagram illustrating a specific example of the voting coefficient P (h) for each height. The vertical axis represents the voting coefficient P (h) for each height, and the horizontal axis represents the real space movement amount M. In FIG. 14, m1 represents a real space movement amount when it is assumed that the optical flow to be processed is an optical flow obtained at the head of a person. m2 represents a real space movement amount when it is assumed that the optical flow to be processed is an optical flow obtained at the foot of a person. As shown in FIG. 14, P (h) when the real space movement amount M is m1 is smaller than P (h) when the real space movement amount M is m2. In other words, the per-height voting coefficient P (h) for the real space movement estimated when it is assumed that the optical flow is obtained in the upper body is the actual value estimated when the optical flow is obtained in the lower body. It is larger than the voting coefficient P (h) for each height with respect to the amount of space movement. Therefore, as for the voting result in one optical flow, the smaller real space movement amount obtains a larger number of votes than the larger real space movement amount.

  The weighting voting unit 206 performs weighted voting of the moving amount on the image using the voting coefficient P (h) for each height calculated by the voting coefficient calculating unit 205 for each height. As described above, in the estimation result of the real space movement amount obtained when it is assumed that the optical flow is a high body part, P (h) has a high value (a value close to 1). On the other hand, in the estimation result of the real space movement amount obtained when it is assumed that the optical flow is a low body part, P (h) is a low value (a value close to 0). In this way, the voting results are weighted for each height of the person's body observed by the fixed camera 10.

The real space movement amount vote estimation unit 207 estimates the real space movement amount. The real space movement amount voting estimation unit 207 estimates the real space movement amount based on the result of the weighted voting unit 206 performing weighted voting. For example, the real space movement amount voting estimation unit 207 acquires the real space movement amount having the largest weighted cumulative value as the estimation result.
FIG. 15 is a diagram illustrating a specific example of processing of the real space movement amount vote estimation unit 207. FIG. 15A shows a weighted vote. Reference numerals 402-1 to 402-3 denote votes obtained from the optical flows 400-1 to 400-3, respectively. As shown in FIG. 15A, each vote is weighted. Therefore, even if each vote has the same number of votes, there is a difference in weight. FIG. 15B is a diagram illustrating a specific example of the voting result. The horizontal axis represents the real space movement amount M, and the vertical axis represents the weighted vote count TW. The weighted vote count TW represents a cumulative value of weights. The real space movement amount voting estimation unit 207 acquires the real space movement amount M having the largest weighted vote number TW as an estimation result.

The passing person estimation unit 208 estimates the passing person number between frame images. The passing number estimating unit 208 stores the number of people in the predetermined area measured by the in-area number measuring unit 202, the real space moving amount estimated by the real space moving amount vote estimating unit 207, and the camera parameter storage unit 204. The number of people passing between frame images is estimated by using camera parameters.
FIG. 16 is a flowchart showing the flow of processing of the passing person estimation method in the second embodiment of the present invention.

  First, the frame image input unit 201 inputs images in chronological order (step S201). Next, the in-area number counting unit 102 measures the number of persons in the predetermined area from the frame image input by the frame image input unit 201 (step S202).

  The on-image movement amount calculation unit 203 detects a feature point of a person from two temporally adjacent frame images input by the frame image input unit 201. The on-image movement amount calculation unit 203 detects the feature points of the person from the two frame images, associates the feature points of the two frame images, and acquires an optical flow. The on-image movement amount calculation unit 203 calculates the on-image movement amount of the person on the image based on the acquired optical flow (step S203).

  The voting coefficient calculation unit 205 for each height uses the camera parameters stored in the camera parameter storage unit 204 and the number of people in the predetermined area measured by the in-area number measuring unit 202 to determine the voting coefficient P ( h) is calculated (step S204).

  The weighting voting unit 206 uses the voting coefficient P (h) for each height calculated by the voting coefficient calculating unit 205 for each height as the amount of movement on the image calculated by the on-image movement amount calculating unit 203, and assigns weights. Voting is performed while performing (step S205).

The real space movement amount voting estimation unit 207 estimates the real space movement amount based on the result of the weighting voting unit 206 performing the weighting voting (step S206).
The passing number estimating unit 208 stores the number of people in the predetermined area measured by the in-area number measuring unit 202, the real space moving amount estimated by the real space moving amount vote estimating unit 207, and the camera parameter storage unit 204. The number of passing people between the frame images is estimated using the camera parameters (step S207).

According to the passing person estimation system configured as described above, it is possible to correct an error in the amount of movement of a person in real space that occurs when an occlusion occurs. In the first embodiment, the real space movement amount correction unit 107 estimates the error after estimating the real space movement amount of the person, and corrects using the correction coefficient C. Therefore, the error of the real space movement amount due to the occlusion can be reduced. As a result, the number of people passing through the area can be estimated more accurately.
In the second embodiment, errors in the real space movement amount can be reduced by performing weighted voting for each person's body height when estimating the real space movement amount. That is, it is possible to correct the height deviation of the person photographed by the fixed camera 10. Therefore, it is possible to more accurately measure the number of people passing through the area.

<Modification>
In the above-described embodiment, the fixed camera 10 and the passing number estimating device 100 are configured as different devices, but the fixed camera 10 and the passing number estimating device 100 may be configured integrally.
The frame image captured by the fixed camera 10 may be once recorded on a recording medium such as a hard disk, and the passing number estimating device 100 may perform the processing while reading the frame image.
The processing in step S102 and the processing in step S103 in the operation example in the first embodiment, and the processing in step S202 and the processing in step S203 in the operation example in the second embodiment may be performed in parallel.

  The embodiment of the present invention has been described in detail with reference to the drawings. However, the specific configuration is not limited to this embodiment, and includes designs and the like that do not depart from the gist of the present invention.

DESCRIPTION OF SYMBOLS 10 ... Fixed camera, 20 ... Shooting area | region, 100 ... Passage number estimation apparatus, 101 ... Frame image input part, 102 ... Number-of-areas measurement part, 103 ... Moving amount calculation part on an image, 104 ... Camera parameter memory | storage part, 105 ... Real space movement amount estimation unit, 106 ... correction coefficient calculation unit, 107 ... real space movement amount correction unit, 108 ... passing number estimation unit, 109 ... correction estimated value acquisition unit, 201 ... frame image input unit, 202 ... number of people in the area Measurement unit, 203 ... On-image movement amount calculation unit, 204 ... Camera parameter storage unit, 205 ... Height-by-height voting coefficient calculation unit, 206 ... Weighting voting unit, 207 ... Real space movement amount voting estimation unit, 208 ... Number of people passing through Estimating unit, 209... Corrected estimated value acquiring unit, 300-1 ... frame image, 300-2 ... frame image, 300-3 ... image 301-1 to 3. ... frame image, 313 ... arrow, 314 ... arrow, 401-1 to 3 ... real space movement amount, 305-1 to 4 ... rectangle (person), 404-1 ... real space movement amount, 404-2 ... real space movement Quantity, 402-1-3 ... weighted voting

Claims (6)

Using the captured image, the area people counting unit that measures the number of people in the area where the subject was observed,
An on-image movement amount calculation unit for calculating an on-image movement amount of the person from the captured image;
A corrected estimated value acquisition unit that acquires a value smaller than the real space movement amount estimated from the movement amount on the image as an estimation result of the real space movement amount;
An estimation apparatus comprising: The corrected estimated value acquisition unit
A real space movement amount estimation unit for estimating a real space movement amount from the movement amount on the image;
A correction coefficient calculation unit for calculating a correction coefficient for correcting an error of the real space movement amount estimated by the real space movement amount estimation unit;
A real space movement amount correction unit that corrects the real space movement amount estimated by the real space movement amount estimation unit using the correction coefficient;
The estimation apparatus according to claim 1, comprising: The corrected estimated value acquisition unit
A voting coefficient calculation unit for each height that calculates a voting coefficient for each height using the number of people in the region and the captured image;
A weighted voting unit for voting by weighting the amount of movement on the image using the voting coefficient for each height and weighting the amount of movement in real space;
The estimation device according to claim 1, further comprising: a real space movement amount voting estimation unit that estimates a real space movement amount by the weighted voting.   The passage number estimation unit for estimating the number of people who have passed through the imaging region using the real space movement amount acquired by the corrected estimated value acquisition unit and the number of people in the region of the captured image. The estimation apparatus of any one of 1-3. A region counting step for measuring the number of people in the region where the subject is observed using the captured image;
An on-image movement amount calculating step for calculating an on-image movement amount of the person from the captured image;
A corrected estimated value acquisition step of acquiring a value smaller than the real space movement amount estimated from the movement amount on the image as an estimation result of the real space movement amount;
An estimation method comprising: A region counting step for measuring the number of people in the region where the subject is observed using the captured image;
An on-image movement amount calculating step for calculating an on-image movement amount of the person from the captured image;
A corrected estimated value acquisition step of acquiring a value smaller than the real space movement amount estimated from the movement amount on the image as an estimation result of the real space movement amount;
A computer program for causing a computer to execute.

Images