JP2016091326A - Camera image person counting method and camera image person counting apparatus - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a camera image person counting method and apparatus for counting the number of persons accurately.SOLUTION: A method of counting the number of objects moving in an image from a camera image includes the steps of; estimating motion of a moving object; setting a line for determining a moving state of the moving object in the image, setting a mesh parallel to a line for determining the size of the moving object, and correcting distortion of a scene of the camera image; correcting overlap of the moving objects overlapping each other, with respect to the estimated motion, to extract grains of the moving object; correcting the size of the moving object with respect to the extracted grains of the moving object; and estimating the number of moving objects in a speed count conversion section, on the basis of the grains of the moving object with the corrected size.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a camera image number counting method and a camera image number counting device, and more particularly to a method and apparatus for measuring the number of objects moving in an area by image processing.

  With the camera image monitoring method, features such as the number of moving objects (persons and objects) (number of people and number of objects) and movement trajectory are easily obtained from images acquired from the camera without attaching special sensor tags to the moving objects. As a means that can be specified, the importance is increasing year by year. Therefore, it has become one of the central issues in the field of image processing and computer vision.

Tomokazu Nomura, “Multiple Regression Model -Econometric Analysis Exercise”, Internet (http://www.econ.kobe-u.ac.jp/~nomura/lecture/11s/multiple_regression.pdf) November 5, 2011 J.L.Barron, D.J.Fleet, and S.S.Beauchemin, “Performance of optical flow techniques”, International Journal of Computer Vision, vol. 12, no. 1, pp. 43-77, 1994. Takagi Takataka, “Image Geometric Correction 1”, Kochi University of Technology, Internet (http://www.infra.kochi-tech.ac.jp/takagi/Geomatics/10GeoCorrection.pdf) Wikipedia Free Encyclopedia, “Distance Space”, Internet (http://en.wikipedia.org/wiki/%E8%B7%9D%E9%9B%A2%E7%A9%BA%E9%96% 93) November 12, 2013 Wikipedia Free Encyclopedia, “Edge Detection”, Internet (http://en.wikipedia.org/wiki/%E3%82%A8%E3%83%83%E3%82%B8%E6%A4% 9C% E5% 87% BA) March 19, 2013

  In image processing for an outdoor environment using a camera image monitoring method, changes in lighting, overlap between objects (occlusion), differences in height or size of objects, crowding due to crowds, or apparent height or size in the field of view Extreme changes occur. These events cause the following problems in counting the number of subjects and the number of subjects. First, due to the overlapping of the objects with the increase / decrease in the number of persons (or the number of persons), the objects themselves cannot be clearly grasped, and the error in counting the number of persons (number of persons) increases. Secondly, the difference in the height of the person (or the size of the object) causes an error in the number of people (number counting). Thirdly, since it is a person counting method (number counting method) using movement information of movement of a person (or an object) as it is, it does not individually correct the overlap or height difference.

  Such a problem is a major factor that lowers the detection accuracy in the automatic counting method of the number of persons and the number of subjects by the camera image monitoring method. The present invention has been made in view of such problems, and an object of the present invention is to provide a camera image number counting method and a camera image number counting device for accurately counting the number of people.

  In order to solve such a problem, a first aspect of the present invention is a camera image number counting method for measuring the number of objects moving within an image from a camera image, wherein the movement is estimated by a motion estimation unit. In the step of estimating the movement of the object, and in the mesh deformation setting unit, a line for determining the movement state of the moving object is set in the image, and parallel to the line for determining the size of the moving object When the moving object overlaps within the image, the congestion degree correcting unit corrects the overlap of the moving object with respect to the estimated movement. Extracting the particles to be moved, and a size correction unit to determine the size of the object to be moved with respect to the extracted particles to be moved. The method comprising positive for, on the basis of the target particle movement to the magnitude of the corrected the extraction, the speed number conversion unit, characterized in that it comprises a step of estimating the number of subjects in the mobile.

  The second aspect of the present invention is the camera image number counting method according to the first aspect, wherein the step of correcting the size is based on the line and the mesh for determining the moving state of the moving object. A dynamic model based on the relationship with the set region is applied to reduce an estimation error of an apparent moving speed in the image due to the difference in size.

  According to a third aspect of the present invention, there is provided a camera image number counting device for measuring the number of objects moving within an image from a camera image, the movement estimating unit estimating the movement of the moving object, A mesh that sets a line for determining the movement state of the moving object in the image, sets a mesh parallel to the line for determining the size of the moving object, and corrects the distortion of the scene of the camera image When the moving target overlaps in the image, a deformation setting unit, a congestion degree correcting unit that corrects an overlap of the moving target with respect to the estimated movement and extracts the moving target particle, and the extracted A size correction unit that corrects the size of the moving target with respect to the moving target particle, and the moving pair based on the extracted target particle of the motion that has been corrected in size. Characterized in that it comprises a speed number conversion section for estimating the number of.

  According to a fourth aspect of the present invention, there is provided the camera image number counting device according to the third aspect, wherein the size correction unit is set by the line and the mesh for determining a moving state of the moving object. Applying a dynamic model based on the relationship with the region to reduce the estimation error of the apparent moving speed in the image due to the difference in size, and correcting the size.

  In the present invention, in camera image processing, in addition to conventional preprocessing, the number of people can be accurately counted by correcting distortion of the coordinate system, correcting the apparent size, and correcting the degree of congestion.

It is a block diagram which shows the structure of the camera image number counting device concerning one Embodiment of this invention. It is a figure explaining the example of the number counting method by the conventional camera image. (A) shows a conventional number counting method, and (b) and (c) show examples of problems of the conventional number counting method. The correction method in the human flow counting method in this invention is shown. It is a flowchart which shows the flow of a process of the people counting method in one Embodiment of this invention. It is a figure explaining the typical relationship between the person and particle | grains which cross | intersect on the straight line of the virtual gate in the mesh defined in FIG. It is a figure which shows the particle | grains which show the pedestrian extracted by the edge extraction process of a pedestrian in the correction method by estimation of the congestion degree which concerns on one Embodiment of this invention, and edge extraction. (A)-(c) is a figure which shows the conventional edge extraction method, (d)-(k) is a figure which shows the edge extraction method in one Embodiment of this invention. It is a figure which shows the example in which the difference in a person's height produces an error in the count of the number of persons, and the example of particle | grain speed correction in one Embodiment of this invention. (A), (c) shows the particles on the straight line of the virtual gate, (b), (d) show the live action, and (e)-(g) show the pedestrian's virtual gate. It is a figure which shows the example of correction | amendment of the speed | velocity | rate of the particle | grains when crossing a straight line. It is a figure explaining the pushing force model which acts on the particle | grains in one Embodiment of this invention. (A)-(c) has shown the moving object in the mesh area | region containing the straight line of the virtual gate defined in FIG. It is a figure explaining the resistance force model which acts on the particle | grains in one Embodiment of this invention. (A)-(c) has shown the object which moves within the mesh area | region containing the straight line of the virtual gate defined in FIG. It is a figure which shows the result of the experiment with the conventional people counting method and the people counting method by one Embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

  FIG. 1 is a block diagram showing a configuration of a camera image number counting device according to an embodiment of the present invention. The camera image number counting device 100 of this embodiment includes a data input unit 101 that inputs image data from a camera or a microscope, and a data storage unit 102 that stores the input data. Further, the camera image number counting device 100 includes a background difference unit 103 that separates background images, a motion estimation unit 104 that estimates a motion to be counted, a mesh deformation setting unit 105 that corrects image scene distortion, and a geometric correction. Unit 106, a congestion correction unit 107 that corrects the overlap of counting targets, and a size correction unit 108 that determines and corrects the difference in size of counting targets. Furthermore, the camera image number counting device 100 includes a speed number conversion unit 109 that estimates the number of objects and a display unit 110 that displays a counting result.

  In the camera image number counting device 100 of the present embodiment, a time series image of a certain environment is taken by a camera or a microscope, and the image data of the taken time series image is input to the data input unit 101. The input time-series image data is stored in the data storage unit 102. In the background difference unit 103, a difference image between successive images in the image data of the time-series image stored in the data storage unit 102 is processed, and a threshold is set for the magnitude of the luminance in the image that is equal to or greater than a certain value. The front part and the background part are separated. The motion estimation unit 104 estimates the motion to be counted from the image data obtained by separating the front part and the background part in the background difference unit 103. The mesh deformation setting unit 105 sets a line for determining the movement state of the counting target and a mesh parallel to the line in the image. The geometric correction unit 106 performs geometric correction so that the movement target of the in-view image moves to the front or back of the camera to have the same size, and corrects the distortion of the image scene when the camera is installed. The congestion correction unit 107 corrects the overlap of people with respect to the estimated movement, and extracts particles that are part of the counting target. The size correction unit 108 performs height correction on the extracted particles and size correction by a plurality of people. For the size correction unit 108, in order to reduce the estimation error of the apparent speed due to the difference in height or the number of people, a dynamic model based on the relationship between the region for determining the movement state of the movement target and the line segment is used. Apply. The speed number conversion unit 109 estimates the number of persons from the corrected image data based on speed information for a predetermined time, and the display unit 110 displays the result of the number counting by the speed number conversion part 109.

  FIG. 2 is a diagram illustrating an example of a conventional counting method using camera images. FIG. 2A shows a conventional number counting method, and FIG. 2B and FIG. 2C show examples of problems of the conventional number counting method. When there are a plurality of pedestrians 202 to 204 in the acquired image, the conventional person counting method virtually draws a straight line 201 in the image as shown in FIG. The number of subjects who cross the straight line 201 of the gate (virtual gate (VG)) is counted. However, depending on the relative line-of-sight angle between the camera and the pedestrian, as shown in FIG. 2 (b), a plurality of persons overlap on the line of sight (205), resulting in a large error in the number of people on the line. There is a case. In addition, as shown in FIG. 2C, even for the same person, the height changes significantly from the front to the back in the image (206 to 208), which may be a cause of erroneous recognition in image processing. That is, in the conventional method of counting people, a method of correcting the apparent congestion degree and size has been a big problem.

  FIG. 3 is a diagram illustrating an example of the number counting method in the speed number conversion unit 109 in FIG. The number counting method shown in FIG. 3 corrects the apparent distortion in the acquired image based on the image as shown in FIG. 2A, and counts the number of subjects 302 that cross the virtual gate straight line 301. To do. The present embodiment is characterized in that the influence of temporal changes in time-series data not considered in the conventional multiple regression model as described in the explanation of FIG. 2 is taken into consideration.

Here, for an explanatory variable x _j having a two-dimensional coordinate value on a certain image, its time series data is expressed as {x ₁ , x ₂ ,. . . , X _T } _j . However, the discrete time is from 1 to T. Next, in order to generate an objective variable related to time change, a weighting factor w (x), a virtual gate line 301 and a vector v _x indicating the motion of the object 302 (hereinafter referred to as “optical flow”) (non-patent document) 2))) is defined as follows, where θ _x is θ _x and | v _x | is the magnitude of the velocity by the optical flow, the moving speed when the apparent distortion is corrected is defined as follows.

However, FG (x) when the two-dimensional coordinate value x is used is binarized data (0, 1), and foreground (1) or background (0) for pixels near the straight line 301 of the virtual gate. ) Is assigned. sin (θ _x ) is a correction value that suppresses an unnatural increase in the number of apparent objects when moving slowly in parallel with the virtual gate straight line 301 over time. For FG (x), the background difference unit 103 in FIG. 1 takes the difference value of the luminance of the pixels in the current and previous past images for a continuous time-series image, and its absolute value is a constant value. Pixels having (for example, 15) or more are given as 1 and others are given as 0. That is, regarding an image region (pixel) in which a change in luminance occurred over a certain value, it was considered that there was a change in the movement of a person or the like.

  FIG. 4 is a flowchart showing a process flow of the people counting method in the present embodiment. In the present embodiment, a certain pedestrian road is photographed with a camera, and the number of people walking on the pedestrian road (counting target) is counted. In the number counting method of the present embodiment, first, in step 401, a pedestrian road is photographed obliquely from a certain fixed point. One spatio-temporal image of the pedestrian road obtained through a certain time obtained by the camera is stored in the data storage unit 102 of FIG. In step 402, the background difference unit 103 separates the front part and the background part by setting a threshold value for the magnitude of the luminance equal to or higher than a certain value.

  In step 403, the distortion of the scene due to the shooting direction of the camera is corrected. In the present embodiment, since the pedestrian road is photographed obliquely from a fixed point, it is necessary to correct the distortion of the scene due to the photographing direction of the camera. In step 403, the mesh deformation setting unit 105 selects an area for counting the number of persons visually from the image, and sets a straight line 411 that becomes a virtual gate so as to cross the pedestrian road. Note that the present embodiment is characterized in that the width of the straight line 411 is expanded to several pixels as compared with the conventional number counting method using the width of one pixel as the width of the straight line 411. Next, a mesh (lattice) 412 of a size that surrounds one pedestrian is set in parallel with the straight line 411 according to the scene, such as 5 × 4 to 50 × 10. By setting the mesh, it becomes possible to grasp the difference in the height of the person. At this time, both ends of the entire mesh parallel to the straight line 411 are set as end straight lines 413. Thereafter, the geometric correction unit 106 corrects the distortion of the scene. For correction of scene distortion, for example, a method (geometric correction) described in Non-Patent Document 3 is used. Even when the object to be moved within the field of view moves to the front or back of the camera by correcting the distortion of the scene, it can be corrected so as to have the same size.

  Next, the congestion correction unit 107 performs correction by estimating the degree of congestion (step 404) on the image on which the distortion of the scene has been corrected. Conventionally, when multiple people overlap when counting the number of people from a camera image, since one environment is photographed from one diagonal direction, the overlapping target is actually recognized as one target. In some cases, it was counted as a smaller number than the number of people present. In this embodiment, this problem is alleviated by incorporating correction based on congestion estimation.

  When correcting by estimating the degree of congestion, if there are multiple pedestrians (the number of people is unknown), obtain a histogram of edge feature values using multiple consecutive images and know the number of people created in advance The increase / decrease in the number of persons is estimated by referring to the histogram of the edge feature amount for each case. The purpose of this process is to eliminate the influence of edge feature amounts due to dynamic changes in the background, such as fluctuations in trees other than pedestrians.

  Next, the size correction unit 108 performs correction based on state estimation (step 405). The difference in human height is also a factor that causes an error in counting the number of people in image processing. In this embodiment, this problem is alleviated by adopting a correction method based on state estimation. In addition, the difference in the relative positional relationship between the apparent blob (blob) and the end straight line 413, which is created by one or more people in the mesh, is defined as a difference in state.

  On the other hand, in step 406, human motion is estimated. The human motion is estimated by the motion estimation unit 104 by the optical flow estimation method described in Non-Patent Document 2. The necessary optical flow obtained by the method described with reference to FIG. 3 is a component perpendicular to the virtual gate line.

  In step 407, particles having effective velocities are extracted from the optical flow estimated in step 406, and the flow of particles on the virtual gate line 411 of the person crossing the virtual gate line 411 in the image. Are extracted respectively.

  In step 408, the virtual force calculated in step 404 is applied to the particles extracted in step 407 to adjust the speed of movement of each particle. Here, the speed is adjusted by converting a virtual force into a speed through a speed conversion formula and determining a speed correction amount.

  In step 409, the particles whose speeds are adjusted are integrated to extract one target (person) to be counted. The result of the extracted number of people is displayed on the display unit 110.

  Next, the correction (step 404) based on the estimation of the congestion level in the congestion correction unit 107 will be described. FIG. 5 is a diagram for explaining a typical relationship between a person and a particle crossing a virtual gate line in the mesh defined in FIG. FIG. 5A shows the pedestrians 503 approaching the virtual gate straight line 501 by superimposing the pedestrians at two successive times. The pedestrian 503 can be replaced with particles 504 having movement. Before applying the correction method in the present embodiment, the particles 504 often face in various directions, and the magnitude of the velocity per unit time often varies.

  FIG. 5B shows particles when one pedestrian is present on the virtual gate straight line 501, and FIG. 5C shows two pedestrians approaching the virtual gate straight line 501. Shows the particles when they overlap.

  In the case of FIGS. 5B and 5C, there is a problem that the number of persons is often erroneously detected as one in image processing. In particular, it is known that the error increases as the number of pedestrians increases and the pedestrians overlap each other.

  FIG. 6 is a diagram illustrating a pedestrian edge extraction process in the correction method (step 404) based on estimation of the degree of congestion according to the present embodiment and particles indicating the pedestrian extracted by edge extraction. 6 (a) to 6 (c) in the upper stage are diagrams showing a conventional edge extraction method, FIG. 6 (a) is an image of a certain pedestrian road, and FIG. 6 (b) is FIG. The figure which extracted the edge from the image of (a), FIG.6 (c) shows the figure which extracted only the pedestrian from the edge of FIG.6 (b). However, the edge extraction method as shown in FIGS. 6A to 6C has a problem that the background feature amount that is not a processing target obstructs the counting of the stable number of people. Therefore, in the present invention, the edges of only the target pedestrian are narrowed down by the method described below.

  FIGS. 6D to 6K in the lower stage are diagrams showing the edge extraction method in the present embodiment. In this embodiment, edge extraction is performed on the image in the mesh created in step 403 in FIG. 4, and FIGS. 6 (d) and 6 (e) show the mesh part images set in FIG. 6 (f) is a diagram obtained by extracting edges from the image (d), FIG. 6 (g) is a diagram obtained by extracting edges from the image (e), and FIG. 6 (h) is a diagram of FIG. ) Is a histogram of the spatio-temporal edge of the diagram in (g), FIG. 6 (j) is the image of (d), and FIG. 6 (k) is the walk of the image of (e) on the virtual gate line. The example of the particle | grains corresponding to a person is shown. The first row of the lower row ((d), (f), (h) and (j)) shows the second row ((e), (g), (i) and (K)) is a case where there are a plurality of pedestrians.

In this embodiment, a reference edge histogram H _Normal and an edge histogram H _t obtained at an arbitrary time are represented as follows:

And the relative rate R _{t of the} degree of congestion at time t is obtained from H _Normal and H _t . Here, D (•) is a function for calculating the degree of similarity between two histograms, and performs distance calculation regarding edge strength and occurrence frequency. There are various methods for calculating the distance, and the Euclidean distance (see Non-Patent Document 4) or the like can be selected. However, since H _Normal is edge information obtained from the spatiotemporal image, the average height of the person is 1.6 m and the average walking speed is 0.5 m / s.

Edge histograms H _Normal and H _t are

Defined in Where x = {t, Normal}, H ⁿ is the length of the edge, and H ^l is the number of edges. The edge can be easily obtained by a spatial first derivative with respect to image luminance, a CANNY edge detector (see Non-Patent Document 5), or the like. When an edge is obtained from only one image, in general, there arises a problem that various buildings and roads in the background that are not targeted are detected together with a person. Therefore, in this embodiment, the edge feature amount is calculated from the time-series image.

Ask for. However, C (i) is an edge in the image i. C ′ (i) is defined by a logical operation AND between the current and previous edge image and an exclusive logical XOR of the current edge image. According to the equation (4), it is possible to eliminate the edge in the extra background, and at the same time, obtain the edges of a plurality of pedestrians from one person who is the object to move.

Next, in this embodiment, the relative rate R _{t of the} degree of congestion is converted into a virtual temperature T _t by the following equation (5).

However, α and β are constants related to the slope and bias of the linear regression function, and are estimated from learning video data in advance.

  Here, in order to further alleviate the influence of background noise in the real environment, particles P (pixels) having effective velocities among particles on a virtual gate line, that is, of the person to be counted The particles that represent a part

It is set as P when satisfying. However, γ is a constant, and is estimated in advance from learning video data. U (P) is a uniform random number in the range of 0 to 1 in P.

The number of particles P is controlled by the virtual temperature T _t . That is, when the temperature T is high (low), the number of particles is increased (decreased). By determining the number of particles P using Equation 6, the degree of congestion can be adjusted and the counting of the number of people can be stabilized and robust.

  Next, correction (step 405) by state estimation in the size estimation unit 108 will be described. Mitigating the effects of variations in human height through state estimation.

  FIG. 7 is a diagram illustrating an example in which an error occurs in counting the number of persons due to a difference in the height of a person, and an example of particle velocity correction in the present embodiment. FIG. 7B shows a live image of a person 701 with a short height, FIG. 7D shows a live image of a person 702 with a high height, and FIG. 7A shows a virtual gate corresponding to FIG. 7B. Particles 704 on a straight line 703, FIG. 7 (c), show particles 704 on a virtual gate straight line 703 corresponding to FIG. 7 (d). Comparing FIG. 7A and FIG. 7C, it can be seen that the number of particles 704 corresponding to a person 702 with a high height is larger than the number of particles 704 corresponding to a person 701 with a short height. When the height is different, the apparent number of particles is also different, and the difference in the apparent number of particles may cause an error in counting the number of people.

  FIGS. 7E to 7G are diagrams showing examples of particle velocity correction when a pedestrian crosses a virtual gate straight line 712. FIG. 7A shows an example in which the velocity of the particles when the pedestrian 711 passes from the left side to the right side of the virtual gate straight line 712 is greatly corrected. FIG. 7B shows the moving direction of the particles. FIG. 7C shows an example in which a virtual force is applied in the reverse direction to correct the speed so as to decrease the speed, and FIG. 7C shows a case where the speed is corrected to increase again.

In the present embodiment, the speed correction based on the size estimation is performed by correcting the speed of the particles P extracted by the congestion degree correction unit 107, and a virtual gate straight line defined in FIG. This is based on the relative distance from the end straight lines at both ends of the mesh region and the difference in state such as the relative position between the virtual gate straight line and the human position. Here, the force _{F P} acting on the particles P,

Give by. In addition,

Is a pressing force acting on the particle P,

Is a resistance force acting on the particles P. This pressing force and resistance force contribute to the increase and decrease of the speed of P. These forces are converted into the speed of the particles P from the mathematical formula as will be described later. As already mentioned, since the number of people is estimated only by the component perpendicular to the virtual gate line, all speed corrections by pressing force and resistance force are performed along the virtual gate line. .

  Next, a method for obtaining the pressing force will be described. FIG. 8 is a diagram for explaining a pressing force model acting on particles in the present embodiment. FIGS. 8A to 8C show moving objects in the mesh area including the virtual gate straight line defined in FIG. 4. The moving object is a blob 800 and is shaded in the mesh. Is shown.

  Here, the right side of Expression (7) is specifically defined. First, the pressing force on the particles P is

It is defined as Here, F ₀ is 1. Further, σ indicates the degree to which the force acting according to the distance is attenuated. d _P is the distance between particle P and the nearest blob boundary patch 801. When the particle P is close to the boundary patch 801 of the blob 800, for example, when the particle P is on the boundary 801 of the blob 800 (FIG. 8A), d _P = 0, the pressing force becomes large, and the particle P Increase the speed of the. On the other hand, when the particle P is inside the blob, that is, when dP is the maximum distance (FIG. 8B), the pressing force is reduced, and the speed of the particle P is decreased. Therefore, when moving from FIG. 8B to FIG. 8A, the moving speed of the particles P increases when the blob leaves the virtual gate line. When a plurality of pedestrians cross the virtual gate straight line 802, the pressing force is reduced. This is because _dP is large.

  Next, a method for obtaining the resistance will be described. FIG. 9 is a diagram for explaining a resistance force model acting on the particles P in the present embodiment. 9A to 9C also show a moving object in the mesh region including the virtual gate straight line defined in FIG. 4. The moving object is a blob 900 and the mesh is included in the mesh. It is shown as a hung.

  Resistance is

Is defined. Here, d _{P, bb} represents the distance between the boundary patches 903 of the two blobs 900 along the moving direction of the particle P. d _{P, b-P} is the distance between the boundary patch 903 of the blob 900 on the moving direction side of the particle P and the particle P. F1 is a resistance force. When a patch having an average speed of 0.3 pixel / frame or more is connected as a moving blob, the blob is connected to both ends of the straight line 901 and the virtual gate 902 from both ends of the straight line. Receive resistance. The definition formula is shown below.

However, W _F is the width of the mesh area. _Fc was set to 1.0. Also,

Is half the average height of a person. From Equation 10, when a pedestrian is in contact with or passes through an end straight line (FIG. 9B), a resistance force is generated, and thus this resistance force is a particle on a virtual straight line 902 of the gate. Decrease the original speed of P. That is, the resistance force is translated to the particle P on the red line segment on the end straight line 901. In other words, it is assumed that such a potential field is formed. On the other hand, when the moving blob is not in contact with both ends of the end straight line 901 (FIG. 9A) or passes rightward (FIG. 9C), the end straight line The resistance force from both ends of 901 is zero. Resistance

When the moving blob 900 has the thickness d _{P, bb} , the resistance force is increased and the velocity of the particles P is adjusted to be smaller than the original velocity. This alleviates the problem that is a problem with the conventional method and is estimated to be faster than the actual speed due to high height.

By combining the pressing force and the resistance force described above, the force FP on the particle _P is determined.

Here, as described above, 1) When the pedestrian enters or leaves the virtual gate line, the resistance force is zero, and the pressing force increases, so that the particle P (virtual gate line) The speed of all the particles above) is adjusted to increase. On the other hand, 2) When the pedestrian crosses the virtual gate straight line, the pressing force acts on the particle P small, and when the blob boundary is close to the end straight line, the resistance force increases. . In this case, the particles P are decelerated. 3) When a plurality of pedestrians cross the virtual gate line (see FIG. 8C), the acting force becomes small. As described above, since not only d _P is large but also d _{P and b-P} are large, the pressing force and the resistance force are small, and the particle P maintains the original moving speed. It is a feature of the present invention that the estimation error of speed due to the difference in the number of pedestrians and the difference in height is corrected by the relative positional relationship (state) between the virtual gate straight line and the end straight line. one of.

  Up to this point, the change in force acting on the particles P has been described, but in order to correct the speed, it is necessary to convert this force into speed. About this, it calculates as follows. The speed of the modified particle P at time t is

It is. Here _{, VP, τ} is the velocity of the particle P at time τ, _{and FP, τ} is the force acting on the particle P at time τ. Note that ρ = 1 and k = 3 from the relationship between stability of speed estimation and calculation cost. In particular, it is effective to use a spatio-temporal image created by k> 1 and a plurality of continuous images.

Finally, in the present embodiment, the speed number conversion unit 109 estimates the number of people walking on the pedestrian road for a certain time N (t ₁ , t ₂ ) as follows.

Here, μ is a coefficient for converting from the adjusted motion vector on the straight line (LOI) of the virtual gate to the number of people, and is estimated in advance from the video data for learning. The weighting factor ω (P) corrects the speed due to scene distortion in the camera scene. As for the number of people per unit time, it is only necessary to divide the number of people accumulated by the time width of time t _{2 to} t ₁ .

[Example]
FIG. 10 shows the results of an experiment between the conventional number counting method and the number counting method according to the present embodiment. One camera was installed on the road in the city area, and while recording the situation of pedestrians traveling for two days, the number of pedestrians coming and going was visually counted, and the visual count was taken as the number of correct answers. Subsequently, the representative conventional number counting method (a) and the number counting method according to the present embodiment are applied to the recorded video image. In the number counting method according to the present embodiment, only the congestion degree correction method (b) is applied. In the case of using the correction method (d) combining the congestion degree correction method and the size correction method only for the size correction method (c), the number of people detection rate was compared and evaluated. As a result, the detection rate improved in the order of a, b, c, and d.

  INDUSTRIAL APPLICABILITY The present invention is used in industrial fields related to monitoring, image sensing, and the like in a real environment in the electric power field, energy field, communication field, and sensing field.

DESCRIPTION OF SYMBOLS 101 Data input part 102 Data storage part 103 Background difference part 104 Motion estimation part 105 Mesh deformation | transformation setting part 106 Geometric correction part 107 Congestion correction part 108 Size correction part 109 Speed people correction part 110 Display parts 201, 301, 411, 501 703, 712, 802, 902 Virtual gate straight line 202-208 Moving object 302, 503, 701, 702, 704, 711 Pedestrian 412 Mesh 413, 502, 901 End straight line 504 Particle 800, 900 Blob 801, 903 Boundary patch

Claims (4)

A camera image counting method for measuring the number of objects moving within an image from a camera image,
Estimating a movement of the moving object in a motion estimation unit;
In the mesh deformation setting unit, a line for determining the movement state of the moving object is set in the image, a mesh parallel to the line for determining the size of the moving object is set, and the camera image Correcting the distortion of the scene;
When the moving object overlaps in the image, the congestion degree correction unit corrects the overlap of the moving object for the estimated movement and extracts the moving target particle;
In the size correction unit, correcting the size of the moving target with respect to the extracted moving target particles;
A method for counting the number of persons in a camera image, comprising: estimating a number of moving objects in a speed number conversion unit based on the extracted particles of the movement target whose size has been corrected.   The step of correcting the size applies the dynamic model based on the relationship between the line for determining the movement state of the moving object and the region set by the mesh, and the image based on the difference in size. 2. The camera image number counting method according to claim 1, wherein an estimation error of an apparent movement speed is reduced. A camera image counting device for measuring the number of objects moving in an image from a camera image,
A motion estimator for estimating the motion of the moving object;
A line for determining the movement state of the moving object is set in the image, a mesh parallel to the line for determining the size of the moving object is set, and the distortion of the scene of the camera image is corrected. A mesh deformation setting unit;
When the moving object overlaps in the image, a congestion degree correcting unit that corrects an overlap of the moving object for the estimated movement and extracts the moving target particle;
A size correction unit that corrects the size of the moving target with respect to the extracted moving target particles;
A camera image number counting device, comprising: a speed number conversion unit that estimates the number of moving objects based on the extracted movement target particles whose size has been corrected.   The size correction unit applies a dynamic model based on a relationship between the line for determining a moving state of the moving target and a region set by the mesh, and the size correction unit 4. The camera image number counting device according to claim 3, wherein the size is corrected by reducing an estimation error of an apparent moving speed.