JP2008009498A - Fall detection device, program, fall detection method, and fall detection system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a technology for precisely detecting the fall of a person. <P>SOLUTION: After various processing is performed by the respective processing parts of a control part 12 to the image data(photographic images) of respective frames configuring a moving image photographed by a camera, the central position of a convex semi-circle is detected on the approximate curve of an outline included in a photographic image by a shoulder position detecting/tracing part 160. Then, an evaluation value(vertical change evaluation value) for evaluating the secular change of the center position to the vertical direction is calculated by a fall detection part 180 so that the fall of a person can be detected based on at least the vertical change evaluation value. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a technique for detecting a fall.

  In recent years, a large number of surveillance cameras are provided in public places and the like, and they are mainly used for the purpose of grasping the occurrence of abnormalities and congestion.

  An example of this abnormality is a person's fall (fall).

  Various techniques for detecting abnormalities such as a person falling over in a room such as a bathroom or a toilet using image data acquired by a surveillance camera have been proposed (for example, Patent Documents 1 to 4). For example, in the technique proposed in Patent Document 1, an abnormality is detected by extracting a high-frequency component from image data and determining a moving state and a stationary state. Further, in the techniques proposed in Patent Documents 2 to 4, a difference area from the background image (reference image) is extracted for each image data, and the height information of the difference area (that is, the human body image) per time The human body's overturning motion is detected by the change of the area and the change of the area.

JP 2002-304680 A JP 2000-253382 A JP 2000-207665 A JP 2000-207664 A

  However, in the techniques proposed in Patent Documents 1 to 4, if a large number of people are crowded in a limited space such as an escalator or a moving sidewalk, whether or not one person falls down It is difficult to detect this.

  For example, in the method of detecting falls using human feature values, if there are multiple people, the feature values will change even if the overlap or alignment of people changes, resulting in false detection. To do. In particular, for example, in an escalator or the like, when photographing the vicinity of the boarding area with a camera from the front side in the direction of travel, a tall person stands behind and a relatively short person (for example, a child) in front of the person falls down. However, since it is impossible to detect a fall by distinguishing one person from another, it is not possible to detect the fall of the person in front.

  In addition, as a technique for detecting a person by distinguishing one person from another, a technique for detecting a head or a face based on a shape, a feature amount, or a color is known. For example, in order to detect a face, There is a restriction that a person must be photographed from the front side. In addition, when the head is detected by shape recognition such as a circle or an ellipse, the proportion of the head occupying the whole person is not so large, and other rounded parts other than the head are likely to be erroneously detected as the head. . Furthermore, in the method of detecting a face by color, erroneous detection is likely to occur due to the influence of illumination, the background or the color of clothes, or the blurring of the skin itself. That is, false detection of a person's fall is likely to occur depending on the environment and subject conditions.

  The present invention has been made in view of the above problems, and an object of the present invention is to provide a technique capable of accurately detecting a person's fall.

  In order to solve the above-described problem, the invention of claim 1 is an image receiving means for receiving a plurality of frames of captured images in time sequence, and an approximate curve of an outline included in the captured image, which is convex upward and has a predetermined For the curve having a radius of curvature in the range, the reference position that maintains a predetermined relationship with the curve regardless of the frame, the captured image of each frame constituting the captured image of a plurality of frames that can be received by the image receiving means A position detection means for detecting from the above, and an evaluation value for calculating a vertical change evaluation value related to a change in at least the reference position in the vertical direction based on a plurality of reference positions respectively detected from a plurality of frames of captured images by the position detection means And a fall means for performing a fall detection process for detecting a fall of a person based on at least the vertical change evaluation value. To.

  The invention of claim 2 is the fall detection device according to claim 1, wherein the fall detection process includes a process of detecting a state in which the vertical change evaluation value exceeds a predetermined value. To do.

  The invention according to claim 3 is the fall detection device according to claim 2, wherein the evaluation value calculating means is based on a plurality of reference positions respectively detected from a plurality of frames of captured images by the position detecting means. Means for calculating a movement evaluation value relating to the movement of the reference position, and the movement evaluation value is set in advance after the fall detection process detects that the vertical change evaluation value exceeds a predetermined value. It includes processing for detecting a state included in the range of values.

  According to a fourth aspect of the present invention, there is provided the fall detection device according to any one of the first to third aspects, wherein the captured images of the plurality of frames are in a photographing direction with respect to a person conveyance path in the person conveyance device. Are images taken at different timings depending on the imaging device set substantially parallel to each other.

  The invention according to claim 5 is the fall detection device according to claim 4, wherein the person conveying device includes an escalator.

  The invention according to claim 6 is a program for causing the fall detection device to function as the fall detection device according to any one of claims 1 to 5 by being executed by a computer included in the fall detection device. is there.

  Further, the invention of claim 7 is directed to (a) an approximate curve of a contour included in a photographed image, wherein the curve has a curvature radius in a predetermined range within a convex range and is not related to each frame. And a reference position that maintains a predetermined relationship with each other from a captured image of each frame constituting a captured image of a plurality of frames whose shooting timings are before and after each other, and (b) a captured image of a plurality of frames in step (a) Calculating a vertical change evaluation value related to a change in at least the reference position in the vertical direction based on a plurality of reference positions respectively detected from (c), and (c) detecting a human fall based on at least the vertical change evaluation value. And a step of performing a fall detection process.

  The invention according to claim 8 is the fall detection method according to claim 7, wherein the fall detection process includes a process of detecting a state in which the vertical change evaluation value exceeds a predetermined value. To do.

  The invention according to claim 9 is the fall detection method according to claim 8, wherein the step (b) is set at a plurality of reference positions respectively detected from a plurality of frames of captured images in the step (a). And calculating a movement evaluation value related to the movement of the reference position based on, and the movement evaluation value is set in advance after the fall detection process detects a state where the vertical change evaluation value exceeds a predetermined value. And a process for detecting a state included in the range.

  The invention according to claim 10 is the fall detection method according to any one of claims 7 to 9, wherein the captured images of the plurality of frames are captured in a capturing direction with respect to the transport path of the person in the person transport apparatus. Are images taken at different timings depending on the imaging device set substantially parallel to each other.

  An eleventh aspect of the present invention is the fall detection method according to the tenth aspect, wherein the person conveying device includes an escalator.

  According to a twelfth aspect of the present invention, there is provided an imaging device that acquires captured images of a plurality of frames at different timings, and the fall detection device according to any one of the first to third aspects.

  The invention according to claim 13 is the fall detection system according to claim 12, wherein the imaging direction of the imaging device is set substantially parallel to the conveyance path of the person in the person conveyance device. Features.

  The invention according to claim 14 is the fall detection system according to claim 13, wherein the person conveying device includes an escalator.

  According to the first aspect of the present invention, an approximate curve of a contour included in a photographed image, which is convex upward and has a radius of curvature in a predetermined range, the curve is not related to each frame. A reference position that maintains a predetermined relationship is detected from the captured image of each frame, and at least a person's fall is detected based on an evaluation value related to a change in the reference position in the vertical direction. Can be done well.

  According to the second aspect of the present invention, a person is detected by detecting a person's fall using a process for detecting a state in which an evaluation value related to a change in the reference position in the vertical direction exceeds a predetermined value. It is possible to detect the falling state with high accuracy and to accurately detect the falling of the person.

  According to the invention of claim 3, when the evaluation value related to the movement of the reference position falls within the predetermined range after the evaluation value related to the vertical change of the reference position exceeds the predetermined value. By detecting that the person has fallen, it is possible to accurately detect the state where the person falls and the subsequent fall state, and to accurately detect the fall of the person. In addition, for example, it is possible to eliminate a problem that a person is detected to have fallen accidentally when the person runs through.

  According to the invention of claim 4, a plurality of frames of captured images respectively captured at different timings by the imaging device in which the imaging direction is set substantially parallel to the direction in which the person is conveyed in the person conveyance device. Even if a configuration that detects the fall of a person is used, it is possible to accurately detect the reference position that maintains a predetermined relationship with the approximate curve of the shoulder contour, so the fall of the person can be detected with high accuracy. Can be done well.

  According to the fifth aspect of the present invention, since the person conveyance device includes the escalator, it is possible to accurately detect the fall of the person in the person conveyance device such as the escalator.

  Further, according to the invention described in claim 6, it is possible to obtain the same effect as the invention described in claims 1-5.

  Further, according to the invention described in claims 7 to 11, it is possible to obtain the same effect as the invention described in each of claims 1 to 5.

  According to the invention described in claim 12, the same effects as those of the invention described in claims 1 to 3 can be obtained.

  Moreover, according to the invention of Claim 13 and Claim 14, the effect similar to the invention described in Claim 4 and Claim 5 can be acquired, respectively.

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

<Outline of the fall detection system>
FIG. 1 is a diagram illustrating a schematic configuration of a fall detection system 1 according to an embodiment of the invention. The fall detection system 1 is used, for example, in a monitoring system that detects the fall of a person when getting on and off an escalator. In the present embodiment, the monitoring location is the vicinity of the escalator boarding / exiting position. However, the present invention is not limited to this, and a general boarding / alighting position for a device that transports other persons such as a moving sidewalk (hereinafter also referred to as “person transporting device”). It may be near.

  As shown in FIG. 1, the fall detection system 1 includes a fall detection device 10 and N cameras (N is a natural number of 4 or more) (first camera, second camera, third camera,... N cameras) C1 to CN are connected via a communication line such as a network line NT so that data can be transmitted and received. In the present embodiment, the number of cameras is four or more. However, the number of cameras is not limited to this.

  The fall detection device 10 is a device that is installed in, for example, a security center of various facilities and detects a person's fall by analyzing moving images transmitted from the first to Nth cameras C1 to CN.

  The first to Nth cameras C1 to CN are video cameras that can shoot moving images, and are installed so as to be able to shoot the vicinity of entrances and exits of escalators where people are likely to fall. Yes. In addition, the first to Nth cameras C1 to CN are capable of shooting a moving image composed of 30 frames per second, and fallen the image data for one frame (that is, the shot image) every 1/30 seconds. Is sent to.

<Camera layout>
FIG. 2 is a schematic diagram illustrating an arrangement example of the first and second cameras C1 and C2. In FIG. 2, the outer edges of the shooting ranges of the first and second cameras C1 and C2 are indicated by broken lines, respectively. .

  As shown in FIG. 2, the escalator 100 conveys the person H in the arrow direction in the figure, for example. The first camera C1 is installed so as to be able to photograph the vicinity of the exit 101 of the escalator 100, and the second camera C2 is installed so that the vicinity of the entrance 102 of the escalator 100 can be photographed. Yes. Specifically, the first and second cameras C1 and C2 are installed on a substantially extended line of a path (hereinafter also referred to as “person transport path”) through which the escalator 100 transports the person H. The arrangement of the first and second cameras C1 and C2 is the same as that of a general camera that monitors the vicinity of the escalator.

  More specifically, the first camera C1 conveys the traveling direction of the escalator 100 in the vicinity of the exit 101 (that is, the person H so that the person H conveyed by the escalator 100 can be photographed from the front side. The second camera C2 is disposed in front of the escalator 100 in the vicinity of the entrance 102 so that the person H conveyed by the escalator 100 can be photographed from the back side. Has been. Therefore, the first camera C1 captures a moving image from the front of the movement of the person H when getting off the escalator 100 while maintaining a constant angle, and the second camera C2 holds the constant angle at the escalator 100. The motion of the person H when riding is shot from behind. As a result, the first camera C1 obtains an image (front image) that captures the person H descending from the escalator 100 from the front, and the second camera C2 captures the person H riding on the escalator 100 from the back (image). Back image) is obtained.

  In other words, the shooting direction of the first camera C1 and the traveling direction of the escalator 100 in the vicinity of the exit 101 are set to be substantially the same, and the shooting direction of the second camera C2 and the vicinity of the entrance 102 are set. It is set so that the traveling direction of the escalator 100 is substantially opposite. In other words, the shooting direction of the first camera C1 and the path (person transport path) through which the escalator 100 transports the person H in the vicinity of the exit 101 are set to be substantially parallel, and the second camera C2 is set. The shooting direction is set so that the escalator 100 transports the person H in the vicinity of the entrance 102 (person transport path) is substantially parallel. It should be noted that the person conveyance path and the shooting direction of each camera are set to be substantially parallel here when the person H is detected from the front or back as much as possible when detecting the contour of the shoulder of the person described later as a substantially semicircular shape. This is because it is preferable to capture. Accordingly, the term “substantially parallel” as used herein allows and includes an angle deviation of about 30 degrees from completely parallel, but the angle deviation is preferably as small as possible, for example, about 15 degrees or less. More preferably, there is no state.

  In FIG. 2, the arrangement of the first and second cameras C1 and C2 among the arrangements of the first to Nth cameras C1 to CN has been described as a representative example, but the arrangement of the third to Nth cameras C3 to CN is described. Since the arrangement is also set in the same manner as the arrangement of the first and second cameras C1 and C2, description thereof is omitted here.

<Configuration of the fall detection device>
FIG. 3 is a block diagram showing mainly the physical configuration (hardware configuration) of the fall detection device 10. In FIG. 3, only the first camera C1 is described as a representative of the first to Nth cameras C1 to CN in order to prevent the diagram from becoming complicated. A case in which image data is analyzed to detect a fall will be described as a representative example.

  The fall detection device 10 is configured using, for example, a general personal computer (personal computer). The fall detection device 10 mainly includes an interface (I / F) 11, a control unit 12, a storage unit 13, an operation unit 14, and a display unit 15.

  The I / F 11 is an interface for controlling transmission / reception of data via the network line NT with the first camera C1 and the like. In the I / F 11, a control signal for controlling the operation of the first camera C1 is output from the fall detection device 10 to the first camera C1, and image data output from the first camera C1 is input. Specifically, the image data (image data for 30 frames per second) respectively acquired at the timing (in this case, every 1/30 seconds) that moves back and forth in time by the first camera C1 is time-sequentially I / F11. And is received by the I / F 11.

  The control unit 12 includes a CPU 12a and a memory 12b, and realizes various functions and operations based on various signals from the operation unit 14, a program PG stored in the storage unit 13, and the like. For example, the control unit 12 reads and executes the program PG stored in the storage unit 13 to analyze the image data input from the first camera C1 or the like and detect a person's fall (fall down). A function for performing detection processing (falling detection function) is realized. More specifically, various functions related to the fall detection process are realized by executing a predetermined application program (here, program PG) in a computer running a predetermined OS (operation system). Note that various data generated in the course of various information processing in the control unit 12 is temporarily stored in the memory 12b.

  The storage unit 13 includes a hard disk or the like, for example, and stores a program PG for realizing a fall detection function. The program PG may be provided in a state where it is stored on a portable recording medium (CD-ROM or DVD), or may be downloaded from a predetermined download site on the network.

  The operation unit 14 includes, for example, a keyboard, a mouse, and the like. When the operation unit 14 is appropriately operated by a user of the fall detection device 10 (for example, a guard), the control unit 12 is operated according to the operation. Send various signals.

  The display unit 15 is configured by, for example, a cathode ray tube or a liquid crystal panel, and visually outputs various images in response to signals from the control unit 12.

<Functional configuration related to the fall detection function>
FIG. 4 is a functional block diagram conceptually showing the functional configuration related to the fall detection function realized by the control unit 12.

  As shown in FIG. 4, the control unit 12 mainly has a luminance image generation unit 120, a pixel thinning unit 130, a frame thinning unit 140, a person edge image generation unit 150, a shoulder position detection / detection function as a functional configuration related to the fall detection function. A tracking unit 160, a process switching unit 170, a fall-down detection unit 180, a fall-down state detection unit 190, and a warning screen generation unit 200 are provided.

  The luminance image generation unit 120 generates image data expressed by luminance from the image data output from the first camera C1 and input via the I / F 11 (hereinafter also referred to as “luminance image”). Specifically, the image data output from the first camera C1 is, for example, 24-bit RGB image data of 240 × 320 pixels, and the luminance image generation unit 120 determines each pixel from the RGB image data. By calculating the luminance value Y using the following equation (1), a luminance image of 240 × 320 pixels is generated.

    Y = 0.29891 × R + 0.58661 × G + 1.448 × B (1)

  The pixel thinning unit 130 thins the luminance image generated by the luminance image generation unit 120 every other line for both the vertical and horizontal lines, so that an image of 120 × 160 pixels (hereinafter referred to as a “reduced image”). To generate).

  The frame thinning unit 140 performs frame thinning every other frame (that is, adopts image data every other frame) for the reduced image generated by the pixel thinning unit 130 for 30 frames per second, so that 15 frames per second are obtained. The reduced image is output to the person edge image generation unit 150.

  The person edge image generation unit 150 regards a subject different from the background as a person, generates an image showing the outline of the person (hereinafter also referred to as “person edge image”), and outputs the image to the shoulder position detection / tracking unit 160. . Details of the person edge image generation unit 150 will be described later.

  The shoulder position detection / tracking unit 160 regards the contour of the person's shoulder as a shape having a predetermined range size similar to a semicircle (semicircular arc), and adds the person edge image generated by the person edge image generation unit 150 to the person edge image. The center position of an upwardly convex semicircle, which is an approximate curve of the included contour, is detected as a position (hereinafter also referred to as “reference position”) indicating an approximate place where a person's shoulder exists. Furthermore, the shoulder position detection / tracking unit 160 detects the center position from each person edge image input continuously in time, and stores information on the center position in the memory 12b by distinguishing each person. Thus, information (hereinafter also referred to as “shoulder tracking information”) TI that tracks the movement of the center position is generated and updated. The shoulder position detection / tracking unit 160 outputs a signal to the processing switching unit 170 when the shoulder tracking information TI is generated or updated for the image data of each frame. Details of the shoulder position detection / tracking unit 160 will be described later.

  The process switching unit 170 detects a state in which the person is in the middle of falling (hereinafter also referred to as a “falling state”) according to the situation (phase), and a person A process of detecting a state (hereinafter also referred to as a “falling state”) that has fallen down (hereinafter also referred to as a “falling state detection process”) is selectively switched and executed. The process switching unit 170 determines and determines the process switching process each time a signal indicating that the analysis of the image data of each frame has been completed is input from the shoulder position detection / tracking unit 160. Here, in the initial state, the process switching unit 170 causes the in-falling state detection unit 180 to execute the in-falling state detection process. Then, once the fall-in-progress state detection unit 180 detects a fall-in-progress state for a certain person, the fall-state detection unit 190 executes the fall-down state detection process for the certain person. Note that the fall state detection process is terminated when the center position of a certain person reaches outside the evaluation range. The reason for switching the processing is to cope with the detailed fall operation of the person who reaches the fall state through the fall state.

  The mid-fall detection unit 180 detects a mid-fall state by evaluating a change in the center position in the vertical direction based on the shoulder tracking information TI generated by the shoulder position detection / tracking unit 160. In addition, when the fall-in-progress state is detected, the fall-in-progress detection unit 180 outputs the detection result to the process switching unit 170, thereby realizing the process switching by the process switching unit 170. Further, the shoulder tracking information TI is updated based on the calculation result in the fall-down detection unit 180. Details of the fall-down detection unit 180 will be described later.

  The fall state detection unit 190 detects a fall state by evaluating a change in the center position based on the shoulder tracking information TI generated by the shoulder position detection / tracking unit 160. In addition, when the fall state is detected, the fall state detection unit 190 outputs the detection result to the warning screen generation unit 200. Further, the shoulder tracking information TI is updated based on the calculation result in the falling state detection unit 190. Details of the fall state detection unit 190 will be described later.

  The warning screen generation unit 200 generates image data of a warning screen that notifies that the person has fallen based on the detection result of the fall state detected from the fall state detection unit 190, and the display unit 15 Output. At this time, a warning screen is visibly output on the display unit 15.

  FIGS. 5 to 8 are functional block diagrams conceptually showing functional configurations included in the person edge image generation unit 150, the shoulder position detection / tracking unit 160, the fall-in-progress detection unit 180, and the fall-down state detection unit 190, respectively. . Hereinafter, the detailed functions of the person edge image generation unit 150, the shoulder position detection / tracking unit 160, the fall-in-progress detection unit 180, and the fall-down state detection unit 190 will be described with reference to FIGS.

○ Personal edge image generation unit 150:
As shown in FIG. 5, the person edge image generation unit 150 mainly includes a difference image generation unit 151, a smoothing processing unit 152, a binarization processing unit 153, an expansion processing unit 154, an edge enhancement unit 155, as functional configurations. A binarization processing unit 156, a mask processing unit 157, and a noise removal unit 158 are provided.

  The difference image generation unit 151 includes an image (hereinafter also referred to as “background image”) G1 (for example, FIG. 9) obtained by capturing a background in advance by the first camera C1 and storing it in the storage unit 13 or the like, and a frame thinning unit. A difference image G3 (for example, FIG. 11) from the reduced image G2 (for example, FIG. 10) input from 140 is generated. The difference image G3 is an image that captures a subject that mainly includes a person, which is different from the background.

  The difference image G3 generated by the difference image generation unit 151 is subjected to smoothing processing in the smoothing processing unit 152, binarization processing in the binarization processing unit 153, and expansion processing in the expansion processing unit 154, respectively. As a result, an image (hereinafter also referred to as “person region specifying image”) that specifies the region in which the person is captured in the captured image is generated. In the smoothing processing unit 152, for example, a general smoothing process is performed in which a simple average is performed between the target pixel and its neighboring pixels. In the binarization processing unit 153, for example, a binarization process using a preset predetermined value as a threshold value or a binarization process using a value obtained from all pixel values according to a predetermined rule as a threshold value is performed. Further, in the expansion processing unit 154, for example, if there is at least one white pixel in the vicinity of the target pixel, a general expansion process is performed so that the target pixel is a white pixel.

  In this way, the person image specifying unit 151, the smoothing processing unit 152, the binarization processing unit 153, and the dilation processing unit 154 generate a person area specifying image that functions as mask information for extracting only moving objects such as people. Generated.

  On the other hand, the edge enhancement unit 155 performs edge enhancement processing for enhancing the contour extending in the horizontal direction on the reduced image G2 (for example, FIG. 10) input from the frame thinning unit 140, and then the binarization processing unit 156. In FIG. 12, an edge image (for example, FIG. 12) extracted by mainly emphasizing the lateral contour of the reduced image is generated by performing the binarization process. In the edge enhancement unit 155, for example, edge enhancement processing using a Sobel filter as shown in FIG. 13 is performed. In the binarization processing unit 156, for example, a binarization process using a preset predetermined value as a threshold value or a binarization process using a value obtained from all pixel values according to a predetermined rule as a threshold value is performed. In the following description, the edge image will be described assuming that the background is white and the edge is black, for example.

  The mask processing unit 157 obtains a logical product of the person region specifying image input from the dilation processing unit 154 and the edge image input from the binarization processing unit 156, so that an edge image focused on a moving object such as a person is obtained. (Personal edge image) G5 (for example, FIG. 14) is generated. The human edge image G5 generated by the mask processing unit 157 is subjected to a known noise removal process such as isolated point removal in the noise removal unit 158, and is output to the shoulder position detection / tracking unit 160. The person edge image generated here has, for example, a white background and an edge of a person or the like expressed in black.

  In this case, the selective use of edges related to moving objects such as people reduces the amount of calculation in the detection of the center position, which will be described later, and suppresses the problem that the center position is erroneously detected from the background. It is to do.

○ Shoulder position detection / tracking unit 160:
As shown in FIG. 6, the shoulder position detection / tracking unit 160 mainly includes a semicircular center detection unit 161 and a center position grouping unit 162 as functional configurations.

  The semicircular center detection unit 161 approximates the contour of the shoulder from the human edge image input from the human edge image generation unit 150 using the fact that the contour of the human shoulder is similar to the arc of the lower chord semicircle. The center position of the semicircle (that is, the coordinates of the center position) is detected. The center position is a reference position indicating a general place where a person's shoulder exists, and the center position can be detected by employing, for example, a Hough transform used for detecting a circle. In the detection and management of the center position, the coordinates of each pixel of the reduced image (vertical 120 pixels × horizontal 160 pixels) as shown in FIG. 10, for example, the lower left of the reduced image is the origin, and the right direction is the X-axis direction. The upper direction is treated as coordinates (X, Y) on the XY plane with the Y-axis direction as the upper direction.

  Here, a center position detection method using the Hough transform will be described.

  In the detection of the center position in the semicircular center detection unit 161, first, a score is calculated as an index when determining whether or not to detect a certain coordinate as the center position.

  FIG. 15 is a flowchart illustrating an operation flow of a score calculation process performed in the semicircle center detection unit 161.

  In step S1, the score is reset and set to 0, which is an initial value.

  In step S2, a count i (i is a natural number) for specifying a pixel to be determined (determination target pixel) among all the pixels constituting the person edge image is set to 1 which is an initial value.

  In step S3, the i-th pixel (Xi, Yi) of all the pixels constituting the person edge image is designated as the determination target pixel.

  In step S4, it is determined whether or not the pixel (designated pixel) designated in step S3 is a black pixel. If the designated pixel is not a black pixel (that is, if it is not a pixel corresponding to an edge such as a person), the process proceeds to step S5. If the designated pixel is a black pixel, the process proceeds to step S7.

  In step S5, it is determined whether all the pixels constituting the person edge image have been designated as the determination target pixels. Here, if all the pixels (120 × 160 = 19200 pixels in this case) are not designated as the determination target pixels, the count i is incremented by 1 in step S6, and the process proceeds to step S3. On the other hand, if all the pixels are designated as the determination target pixels, the flow of the score calculation processing operation is ended. That is, the processes after step S3 are repeated as appropriate until all the pixels are designated as the determination target pixels.

  In step S7, the center coordinates (Xc, Yc) of the circle passing through the black pixel (Xi, Yi) that is the designated pixel are obtained. Here, assuming that the maximum radius of the circle is Rmax, the center coordinates (Xc, Yc) are obtained within a range satisfying the following expressions (2) and (3).

(Xi−Rmax) ≦ Xc ≦ (Xi + Rmax) (2)
(Yi−Rmax) ≦ Yc ≦ Yi (3)

  That is, the coordinates of all the pixels included in the semicircle with the radius Rmax (first chord semicircle) centered on the black pixel (Xi, Yi) are obtained as the center coordinates (Xc, Yc). Therefore, in step S7, a large number of coordinates are obtained as center coordinates (Xc, Yc). The maximum radius Rmax is a value set in advance as an upper limit value of the size of the person's shoulder captured in the image, and may be specified in advance by the program PG, or may be arbitrarily determined by the operation of the operation unit 14 by the user. It may be possible to set to.

  In step S8, a count j for specifying a center coordinate (hereinafter also referred to as “score calculation target central coordinate”) for which a score is to be calculated among the many central coordinates (Xc, Yc) obtained in step S7. j is a natural number) which is set to 1 which is an initial value.

  In step S9, the j-th center coordinate (Xcj, Ycj) among the many center coordinates (Xc, Yc) obtained in step S7 is designated as the score calculation target center coordinate.

  In step S10, the distance between the center coordinate (Xcj, Ycj) specified in step S9 and the black pixel (Xi, Yi) that is the specified pixel is centered on the coordinate (Xcj, Ycj) by the following equation (4). It is obtained as a radius Rj of a semicircle passing through the black pixel (Xi, Yi).

Rj = √ {(Xi−Xcj) ² + (Yi−Ycj) ² } (4)

  In step S11, it is determined whether or not the radius Rj obtained in step S10 is not less than a preset minimum radius Rmin and not more than a preset maximum radius Rmax. That is, it is determined whether or not the following expression (5) is satisfied.

    Rmin ≦ Rj ≦ Rmax (5)

  The minimum radius Rmin is a value set in advance as a lower limit value of the size of the person's shoulder captured in the image, and may be defined in advance by the program PG, or may be arbitrarily determined by the user operating the operation unit 14. It may be possible to set to. In this way, by limiting the radius range to some extent according to the size of the person's shoulder, the amount of calculation in the Hough transform can be reduced, and the processing speed can be increased.

  In this step S11, if it is determined that the above equation (5) is not satisfied for the combination of the center coordinates (Xcj, Ycj) and the radius Rj, the process proceeds to step S13, and it is determined that the above equation (5) is satisfied. In step S12, the score for the combination of the center coordinates (Xcj, Ycj) and the radius Rj is increased by 1, and the process proceeds to step S13.

  In step S13, it is determined whether or not all center coordinates (Xc, Yc) obtained in step S7 are designated as score calculation target center coordinates. If all the center coordinates (Xc, Yc) are not designated as the score calculation target center coordinates, the count j is incremented by 1 in step S14, and the process proceeds to step S9. On the other hand, if all the center coordinates (Xc, Yc) are designated as the score calculation target center coordinates, the process proceeds to step S5. That is, until the score relating to all the central coordinates (Xc, Yc) is calculated for one designated pixel (Xi, Yi), the processes in steps S9 to S14 are repeated, and all the designated pixels (Xi, Yi) are obtained. When the score relating to the central coordinates (Xc, Yc) is calculated, the next pixel is designated as the designated pixel, or the flow of the score calculation processing operation is ended.

  Note that the score calculation processing operation (FIG. 15) in the semicircle center detection unit 161 is executed each time one frame of a person edge image is input from the person edge image generation unit 150.

  After calculating the score by the above-described score calculation processing operation (FIG. 15), the semicircle center detecting unit 161 calculates the following formula (6) for a predetermined number (here, the top five) scores from the highest score: If there is a combination of a center coordinate (Xcj, Ycj) to which a satisfying score is given and a radius Rj, the center coordinate (Xcj, Ycj) is detected as a coordinate of the center position (hereinafter simply referred to as “center position”).

    {(Score × 100) / (3.14 × Rj)}> 70 (6)

  FIG. 16 is a diagram exemplifying the position of a substantially semicircle that roughly indicates the contour of the shoulder detected for the person edge image G5 (FIG. 14) and the center position thereof. In FIG. 16, a substantially semicircle indicating the contour of the shoulder is indicated by a thick line, and the center positions Pc1 and Pc2 are indicated.

  In addition, about the score evaluation method in the semicircle center detection part 161, it is not restricted to what uses the said Formula (6), For example, you may use another relational expression. Further, in the semicircular center detection unit 161, when there are a plurality of center positions, a plurality of center positions where the distance between the center positions is less than or equal to a predetermined distance (for example, a quarter of the minimum radius Rmin). May exclusively adopt one central position according to a predetermined rule (for example, a rule that adopts the one with the largest assigned score or a rule that adopts the one with the largest radius Rj). . This process is a process for suppressing the occurrence of a problem in which a plurality of center positions are detected for the shoulder of one person due to the fact that the contour of the shoulder is not completely semicircular.

  Then, the detection of the center position in the semicircle center detection unit 161 is executed every time a person edge image is input from the person edge image generation unit 150, and each time the center position information is output to the center position grouping unit 162. Is done.

  The center position grouping unit 162 groups the information on the center position sequentially input from the semicircular center detection unit 161 into information on the center position for each person and tracks the center position (hereinafter referred to as “shoulder”). TI) (also called “tracking information”) is generated and stored in the memory 12b.

  As a method of grouping, for example, a center position (hereinafter also referred to as “n center position”) detected in an nth frame (n is a natural number) person edge image and an n + 1th frame person edge image are detected. A method of grouping combinations of center positions that are closest to the center position (hereinafter also referred to as “n + 1 center position”) as the center positions of the same person can be employed. The distance D between the n center position and the n + 1 center position is calculated by the following expression (7), where the coordinates of the n center position are (X1, Y1) and the coordinates of the n + 1 center position are (X2, Y2). .

D = √ {(X1−X2) ² + (Y1−Y2) ² } (7)

  By such grouping of the central positions, even when a plurality of persons are captured in one frame of the person edge image G5, it is possible to distinguish and track and manage the central positions for each person.

  FIG. 17 is a diagram illustrating first and second center position groups Gc1 and Gc2 obtained by grouping the center positions detected by analyzing the human edge image for five frames. FIG. 17 shows first and second center position groups Gc1 and Gc2 each having five center positions for the image area AR corresponding to the person edge image G5.

  However, in a person edge image generated in time sequence by the person edge image generation unit 150, that is, an image that moves around in time capturing the same angle, the shoulder of the person appears / disappears in time sequence as the person moves. . That is, a person's shoulder enters and exits the shooting range in time sequence. Then, when the shoulder of the second person exists in the shooting range when the shoulder of the first person goes out of the shooting range, the center position of the first person and the center position of the second person May be grouped by mistake. In order to avoid such a problem, in the center position grouping unit 162, for example, when the distance D between the n center position and the n + 1 center position is a predetermined distance (for example, 10 pixels) or more. Do not form the same group. That is, the group at the center position related to the first person disappeared from the image is regarded as the past, and all information related to the group is deleted from the shoulder tracking information TI. On the other hand, the center position relating to the second person existing on the image forms another group uniquely.

  Thus, by tracking the change in the center position that identifies the location of the shoulder that occupies a certain size (size) on the image, it is possible to accurately detect the change in the posture of the person.

  The center position grouping unit 162 outputs a signal indicating that the analysis process has been completed to the process switching unit 170 every time the analysis process of the person edge image for one frame is completed.

  FIG. 18 is a diagram illustrating the change over time of the center position when the person falls. In FIG. 18, the vertical axis indicates the height (Y coordinate) of the center position in the image, the horizontal axis indicates time, and the broken line indicates the change over time of the height (Y coordinate) of the center position. As shown in FIG. 18, when a person falls, first, the height of the center position changes greatly in the process of falling (time t1 to t2), and then the height of the center position in the fallen state. Hardly changes (time t2 to t3). Therefore, in the fall detection device 10 according to the present embodiment, a fall detection process is performed in which a fall-down state is detected by the fall-down detection unit 180, and then a fall-down state (substantially stationary state) is detected by the fall-down detection unit 190. Thus, for example, even when a person runs in front of the first camera C1, it can be prevented that the person has accidentally fallen, and as a result, It can be detected with high accuracy. For reference, FIG. 18 also shows the change over time in the height of the center position in the process of getting up from the collapsed state (time t3 to t4).

○ Falling detection unit 180:
As shown in FIG. 7, the in-fall detection unit 180 mainly includes a vertical movement amount calculation unit 181, a vertical movement amount evaluation unit 182, and a phase switching determination unit 183 as functional configurations.

  The vertical movement amount calculation unit 181 calculates the movement amount in the vertical direction of the center position for each group by referring to the shoulder tracking information TI. The vertical movement amount calculation unit 181 calculates a difference in the Y coordinate of the center position between the frames (hereinafter also referred to as “center position vertical difference”), and associates it with the information on the center position used for the calculation of the center position vertical difference. And stored in the shoulder tracking information TI. At this time, the shoulder tracking information TI is updated.

  The vertical movement amount evaluation unit 182 evaluates the movement amount of the person's shoulder during a predetermined period (for example, 1/3 second) by referring to the shoulder tracking information TI. Specifically, for example, by referring to the shoulder tracking information TI, the center position vertical difference between each frame is squared for a predetermined number of frames (for example, a total of 5 frames including the latest frame and the latest 4 frames). Then, the value divided by the comparison number (for example, 4) is calculated as an evaluation value (hereinafter also referred to as “vertical change evaluation value”) related to the vertical change in the center position over time. And when the said vertical change evaluation value exceeds predetermined value (for example, 20), since the change of a center position is comparatively large, it is evaluated that it is the state in the middle of a person falling (the state in the middle of a fall). That is, a state in the middle of falling is detected. Here, as the vertical change evaluation value, a value obtained by squaring and accumulating the center position vertical difference for a predetermined number of frames and dividing by the comparison number is adopted, but the present invention is not limited to this, and for example, a predetermined number of frames Any value may be used as long as the vertical change of the center position can be evaluated, such as a value obtained by simply summing the vertical differences of the center positions between the frames.

  The phase switching determination unit 183 detects the half-falling state detection in response to the fact that the vertical movement amount evaluation unit 182 evaluates the half-falling state for each person (that is, the group at the center position). It is determined to switch from the phase in which processing is performed (phase 1) to the phase in which the falling state detection processing for detecting a falling state (phase 2) is performed, and a signal indicating that is output to the processing switching unit 170.

○ Falling state detection unit 190:
As shown in FIG. 8, the falling state detection unit 190 mainly includes a movement amount calculation unit 191 and a dispersion amount evaluation unit 192 as functional configurations.

  The movement amount calculation unit 191 calculates the movement amount of the center position for each group by referring to the shoulder tracking information TI. The movement amount calculation unit 191 calculates the difference between the XY coordinates of the center position between the frames, that is, the difference in the horizontal and vertical directions (hereinafter also referred to as “center position horizontal / vertical difference”), and the center position horizontal / vertical. The shoulder tracking information TI is stored in association with the information on the center position used for the difference calculation. At this time, the shoulder tracking information TI is updated.

  The dispersion amount evaluation unit 192 evaluates the dispersion amount of the center position (that is, the amount of movement of the person's shoulder) in a predetermined period (for example, 1/3 second) by referring to the shoulder tracking information TI. Specifically, for example, referring to the shoulder tracking information TI, the center position horizontal / vertical difference between each frame is squared for a predetermined number of frames (for example, a total of 5 frames including the latest frame and the latest 4 frames). Then, a value obtained by accumulating and dividing by the number of comparisons (for example, 4) is calculated as an evaluation value (hereinafter also referred to as “movement evaluation value”) related to the amount of movement of the center position over time. When the movement evaluation value is less than a predetermined value (for example, 5), that is, within a predetermined value range (for example, 0 or more and 4 or less), the temporal change in the center position is extremely small. It is evaluated (detected) that the person has fallen to a state (falling state) and detects that the person has fallen. Here, as the movement evaluation value, a value obtained by squaring and accumulating the center position horizontal / vertical difference for a predetermined number of frames and dividing by the comparison number is adopted, but is not limited to this, for example, a predetermined number of frames Any value that can evaluate the movement of the center position, such as a value obtained by simply summing the center position horizontal and vertical differences between the frames, may be used.

  For example, when the person H falls down at the entrance 102 of the escalator 100, a scene in which the person H is transported in a fallen state by the escalator 100 is also assumed. Since the amount of change in position over time is a relatively small value, the fall state can be accurately detected by the fall state detection process.

  Further, when the variance amount evaluation unit 192 detects that the person has fallen, the variance amount evaluation unit 192 outputs a signal indicating that to the warning screen generation unit 200.

  Note that the timing of performing the above-described falling state detection process by the falling state detection unit 180 and the falling state detection process by the falling state detection unit 190 may be limited. For example, when the center position exists in a predetermined area on the image (for example, as shown in FIG. 19, the area from the 10th line to the 80th line from the bottom of the shaded area in the image area AR). Only in the middle, the fall state detection process or the fall state detection process may be performed. In general, it is predicted that the time from the fall-down state to the fall-down state is relatively short. Therefore, a period during which the fall-down state detection process is performed is set to a predetermined period (for example, 0.5 seconds or 1 second). ) And forcibly returning to phase 1 in which the state detection process during the fall is performed in response to the passage of the predetermined period.

<Operation related to fall detection processing>
FIG. 20 is a flowchart showing an operation flow related to the fall detection process in the fall detection system 1. This operation flow is started and realized when the program PG is read and executed by the control unit 12 when the user operates the operation unit 14 in the fall detection device 10 in various ways. Here, in order to simplify the description, the description will be focused on the case where image data of each frame constituting a moving image shot by the first camera C1 is sequentially input.

  In step ST1, image data for one frame is input from the first camera C1 to the luminance image generation unit 120 via the I / F 11.

  In step ST2, a luminance image is generated from the image data input in step ST1 by the luminance image generation unit 120.

  In step ST3, a reduced image is generated by performing line thinning (that is, pixel thinning) on every other vertical and horizontal lines for the luminance image generated in step ST2 by the pixel thinning unit 130.

  In step ST4, it is determined whether or not the reduced image generated in step ST3 by the frame thinning unit 140 is a reduced image related to an even-numbered frame. If it is determined that the image is not a reduced image related to the even-numbered frame, the reduced image is not adopted, and the process returns to step ST1. If it is determined that the image is a reduced image related to the even-numbered frame, the reduced image is adopted. Then, the process proceeds to step ST5. That is, frame thinning processing is performed. In addition, although it demonstrated as what employ | adopts the reduced image which concerns on an even-numbered frame here, you may employ | adopt the reduced image which concerns on an odd-numbered frame.

  In step ST5, the person edge image generation unit 150 generates a person edge image from the reduced image.

  In step ST6, the center position is detected by the shoulder position detection / tracking unit 160 from the person edge image generated in step ST5.

  In step ST7, the center positions of the center positions detected in step ST6 by the shoulder position detection / tracking unit 160 are grouped for each person. In the case where the process of step ST7 is the first time, only one point is detected for each person's center position, so that each center position forms each group, but the process of step ST7 is the second time. Thereafter, the center positions that change over time are grouped for each person, and shoulder tracking information TI that tracks the center position for each person is generated and stored in the memory 12b.

  In step ST8, it is determined whether or not the group at the center position is stored in the memory 12b, that is, whether or not the shoulder tracking information TI is stored in the memory 12b. If the shoulder tracking information TI is not stored in the memory 12b, the process returns to step ST1. On the other hand, if the shoulder tracking information TI is stored in the memory 12b, the process proceeds to step ST9. When no person is captured by the first camera C1, the shoulder tracking information TI is not stored in the memory 12b.

  In step ST9, a count A (A is a natural number) for specifying a group related to the center position to be evaluated among groups related to the center position included in the shoulder tracking information TI is set to 1 which is an initial value. .

  In step ST10, a group related to the center position of the Ath person is designated as an evaluation target.

  In step ST11, it is determined by the process switching unit 170 whether or not the evaluation target specified in step ST10 is in phase 2. Here, if it is not in phase 2, the process proceeds to step ST12, and if it is in phase 2, the process proceeds to step ST15.

  In step ST12, a fall-in-progress state detection process by the fall-in-progress detection unit 180 is performed for the group at the center position of the Ath person designated as the evaluation target in step ST10.

  In step ST13, it is determined whether or not it is in the middle of a fall depending on the processing result in step ST12. Here, if it is in the state of falling down, the process proceeds to step ST14, and the A person who is the object of evaluation in step ST14 is switched to phase 2 by the phase switching determination unit 183. On the other hand, if it is not in the middle of falling, the process proceeds to step ST18.

  In step ST15, the fall state detection process by the fall state detection unit 190 is performed for the group at the center position of the Ath person designated as the evaluation target in step ST10.

  In step ST16, it is determined whether or not it is in a fallen state according to the processing result in step ST15. Here, if it is in the fall state, the process proceeds to step ST17, where the warning screen image data is generated by the warning screen generation unit 200 in step ST17, and the warning screen is visually output on the display unit 15. A warning sound may be output together with the warning screen. At this time, for example, a security guard can confirm an image currently captured by the first camera C1 on the display unit 15 of the security center, and determine whether or not an emergency measure is necessary. On the other hand, if not in the fall state in step ST16, the process proceeds to step ST18.

  In step ST18, it is determined whether or not all groups related to the center position included in the shoulder tracking information TI are designated as evaluation targets. If all groups are not designated as evaluation targets, the count A is incremented by 1 in step ST19, and the process proceeds to step ST10. On the other hand, if all groups are designated as evaluation targets, the process returns to step ST1. That is, the processes of steps ST10 to ST19 are repeated as appropriate until all groups are designated as evaluation targets.

  Then, each time image data for one frame is input from the first camera C1 to the luminance image generation unit 120 via the I / F 11, the processes of steps ST1 to ST19 are performed.

  As described above, in the fall detection system 1 according to the embodiment of the present invention, in the fall detection apparatus 10, the image data of each frame constituting the moving image shot by the first to Nth cameras C <b> 1 to CN is imaged. The center position of an upwardly convex semicircle that is an approximate curve of the contour included in is detected. Then, a vertical change evaluation value for evaluating the temporal change of the center position in the vertical direction is calculated, and a person's fall is detected based on at least the vertical change evaluation value. Adopting such a configuration, for example, a certain amount of people regardless of various environmental factors such as a situation in which a plurality of people are mixed in a limited space and a plurality of people are overlapped on an image, a difference in lighting, etc. Since the reference position (here, the center position) related to the shoulder occupying the size area can be detected with relatively high accuracy, it is possible to accurately detect the person's fall.

  Further, the fall of a person is detected using a fall-in-progress state detection process that detects a state in which the vertical change evaluation value related to the change in the center position in the vertical direction exceeds a predetermined value. According to the fall-in-progress state detection process using the center position that roughly specifies the location of the shoulder, it is possible to accurately detect the state in which the person is falling down, so that the fall of the person can be accurately detected. it can.

  Further, when the movement evaluation value related to the movement of the center position falls within the predetermined range after the vertical change evaluation value related to the vertical change of the center position exceeds the predetermined value (here, by the fall state detection process) When a fall state is detected, it is detected that the person has fallen. By adopting such a configuration, it is possible to accurately detect a state in which a person falls (a state in the middle of a fall) and a subsequent fall state (a fall state). For this reason, it is possible to accurately detect the fall of the person.

  In addition, the images were taken at different timings by the first to Nth cameras C1 to CN in which the photographing direction was set substantially parallel to the direction in which the person H was conveyed by the person conveying device (here, the escalator 100). A person's fall is detected from a plurality of photographed images. Even when a configuration conforming to the general arrangement of a surveillance camera related to an escalator or the like in which a person H conveyed by such a person conveying apparatus is photographed from the front direction or the back direction is adopted, the contour of the shoulder The center position (that is, the reference position) that maintains a predetermined relationship with the approximate curve can be accurately detected. Accordingly, it is possible to accurately detect the fall of the person.

  In particular, by applying the fall detection system 1 to the escalator 100 included in the person conveyance device, the fall of the person can be accurately detected in the escalator 100 or the like where the person is likely to fall.

<Modification>
As mentioned above, although embodiment of this invention was described, this invention is not limited to the thing of the content demonstrated above.

  For example, in the above embodiment, the fall detection device 10 is configured to include the display unit 15. However, the present invention is not limited to this. For example, a configuration in which the display unit 15 is excluded from the fall detection device 10 is referred to as a fall detection device. A fall detection system may be configured by connecting a display device that displays various images to the fall detection device.

  Also, in the above embodiment, in order to track and evaluate the movement of the person's shoulder, the semi-circle that is an approximate curve of the outline of the person's shoulder is used as a reference position indicating the approximate location where the person's shoulder exists. Although the center position is detected, the present invention is not limited to this. For example, a position that is uniquely determined by the relationship with the semicircle, such as one end of the arc of the semicircle or the center point, that is, a position that maintains a predetermined relationship with the semicircle may be detected as the reference position. Furthermore, in the above embodiment, the approximate curve of the contour of the person's shoulder is a semicircle, but the present invention is not limited to this, and may be, for example, an arc of 1/3 of a circle or a partial arc of an ellipse. Therefore, in order to track and evaluate the movement of a person's shoulder, an approximate curve of the contour included in the photographed image, which is convex upward and has a curvature radius in a predetermined range, is determined by each frame. Instead, a reference position that maintains a predetermined relationship with the curve may be detected from the captured image of each frame.

  In the above embodiment, the analysis of one frame image has been described as being completed before the next frame image is input. However, the present invention is not limited to this. For example, the analysis of one frame image is performed as follows. If it is not completed before the image of the frame is input, the analysis of the image of the next frame may be started during the analysis of the image of one frame.

  In addition, in the above embodiment, after detecting a fall-in-progress state in which the vertical change evaluation value of the center position exceeds a predetermined value, by detecting a fall-down state in which the movement evaluation value of the center position is included in a predetermined range, Although it is detected that the person has fallen, the present invention is not limited to this. For example, it may be detected that the person has fallen only by detecting a fall-in-progress state in which the vertical change evaluation value of the center position exceeds a predetermined value. . However, when a configuration in which a person's fall is detected simply by detecting a fall-in-progress state in which the vertical change evaluation value of the center position exceeds a predetermined value, for example, when the person H runs through the escalator 100, etc. There is a possibility that a problem of detecting that a person has fallen by mistake occurs. Therefore, from the viewpoint of improving the accuracy of human fall detection by suppressing erroneous detection, the movement evaluation value of the center position is detected after detecting a fall-in-progress state in which the vertical change evaluation value of the center position exceeds a predetermined value. It is preferable to detect that the person has fallen by detecting a fall state included in the predetermined range.

  In the above embodiment, the center position is tracked for each person, and the fall state detection process and the fall state detection process are selectively performed according to the phase for each person. When a state in the middle of a fall for one person is detected, it may be switched to Phase 2 where the fall state detection process is simply performed for all the persons. When such a configuration is employed, the period during which the fall state detection process is performed may be appropriately limited to return to Phase 1 in which the fall state detection process is performed. As a method of appropriately limiting the period during which the fall state detection process is performed, for example, a method of returning to phase 1 when the center position deviates from a predetermined region on the image (for example, a hatched region in FIG. 19), A method of limiting the period during which the fall state detection process is performed to a predetermined period set in advance (for example, 0.5 seconds or 1 second) and returning to Phase 1 after the predetermined period elapses may be considered.

  In the above-described embodiment, in the fall-in-progress state detection process, a fall-in-progress state is detected based on the change of the center position in the vertical direction (that is, the one-dimensional direction). Although a fall-in-progress state is detected based on movement in two dimensions (XY directions), the present invention is not limited to this. For example, even in the fall-down state detection process, the center position changes (moves) in the vertical direction (that is, one-dimensional direction). The fall state may be detected based on the above. However, from the viewpoint of suppressing false detection in a scene where a person runs from the escalator and then goes down in the horizontal direction, the fall state detection process is based on the movement of the center position in the two-dimensional plane. It is preferable to detect the state of falling down.

  In the above embodiment, each functional configuration realized by the control unit 12 has been described as being realized by cooperation of a hardware configuration and software. However, the present invention is not limited to this. A part or all of the included functional configuration may be realized by a hardware configuration.

It is a figure which illustrates schematic structure of the fall detection system which concerns on embodiment of this invention. It is a figure which shows an example of arrangement | positioning of the camera which concerns on embodiment of this invention. It is a block diagram which mainly shows the physical structure of a fall detection apparatus. It is a functional block diagram which shows notionally the function structure which concerns on a fall detection function. It is a block diagram which illustrates the functional composition of a person edge image generation part. It is a block diagram which illustrates the functional composition of a shoulder position detection / tracking part. It is a block diagram which illustrates the functional composition of a fall way detecting part. It is a block diagram which illustrates the functional composition of a fall state detection part. It is an image figure of a background image. It is an image figure of a reduction image. It is an image figure of the difference image which corresponds to the difference of a reduction image and a background image. It is an image figure of an edge image. It is a figure which illustrates the filter characteristic of a Sobel filter. It is an image figure of a person edge image. It is a flowchart which illustrates the operation | movement flow of score calculation. It is a figure which illustrates the approximate semicircle which shows the outline of the shoulder on an image, and the position of the center. It is a figure which illustrates a mode that a center position changes on an image. It is a figure which illustrates the change of a center position. It is a figure which illustrates the field which tracks a center position on an image. It is a flowchart which illustrates the operation | movement flow of a fall detection process.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Fall detection system 10 Fall detection apparatus 12 Control part 100 Escalator 150 Human edge image generation part 160 Shoulder position detection / tracking part 170 Process switching part 180 Falling-down detection part 190 Falling state detection part C1-CN 1st-Nth camera PG program TI shoulder tracking information

Claims (14)

Image accepting means for accepting captured images of a plurality of frames in time sequence;
A reference position that is an approximate curve of an outline included in a captured image and has a predetermined relationship with the curve regardless of each frame, with respect to a curve having an upwardly convex curvature radius within a predetermined range. Position detecting means for detecting from the captured image of each frame constituting the captured image of a plurality of frames received by the image receiving means;
Evaluation value calculation means for calculating a vertical change evaluation value related to a change in the vertical direction of at least the reference position based on a plurality of reference positions respectively detected from a plurality of frames of captured images by the position detection means;
A fall detection means for performing a fall detection process for detecting a fall of a person based on at least the vertical change evaluation value;
A fall detection device comprising: The fall detection device according to claim 1,
The fall detection process
A fall detection device comprising a process of detecting a state in which the vertical change evaluation value exceeds a predetermined value. The fall detection device according to claim 2,
The evaluation value calculating means is
Means for calculating a movement evaluation value related to movement of a reference position based on a plurality of reference positions respectively detected from a plurality of frames of captured images by the position detection means;
The fall detection process
A fall detection device comprising a process of detecting a state in which the movement evaluation value is included in a preset value range after detecting a state in which the vertical change evaluation value exceeds a predetermined value. The fall detection device according to any one of claims 1 to 3,
The captured images of the plurality of frames are
A fall detection device, characterized in that it is an image taken at a different timing by an imaging device whose photographing direction is set substantially parallel to a person conveyance path in the person conveyance device. The fall detection device according to claim 4,
The person conveying device is
A fall detection device including an escalator.   The program which makes the said fall detection apparatus function as a fall detection apparatus in any one of Claims 1-5 by being performed by the computer contained in a fall detection apparatus. (a) A reference position that is an approximate curve of an outline included in a photographed image and has a predetermined relationship with the curve regardless of each frame with respect to a curve that is convex upward and has a curvature radius in a predetermined range Detecting from a captured image of each frame constituting a captured image of a plurality of frames whose shooting timings are before and after each other,
(b) calculating a vertical change evaluation value related to a change in the vertical direction of at least the reference position based on the plurality of reference positions respectively detected from the captured images of the plurality of frames in the step (a);
(c) performing a fall detection process for detecting a person's fall based at least on the vertical change evaluation value;
A fall detection method comprising: The fall detection method according to claim 7,
The fall detection process
A fall detection method comprising a process of detecting a state in which the vertical change evaluation value exceeds a predetermined value. The fall detection method according to claim 8,
Step (b)
Calculating a movement evaluation value related to movement of the reference position based on a plurality of reference positions respectively detected from the captured images of the plurality of frames in the step (a),
The fall detection process
A fall detection method, comprising: detecting a state in which the movement evaluation value is included in a preset value range after detecting a state in which the vertical change evaluation value exceeds a predetermined value. A fall detection method according to any one of claims 7 to 9,
The captured images of the plurality of frames are
A fall detection method, characterized in that the images are taken at different timings by an imaging device in which the photographing direction is set substantially parallel to a person conveyance route in the person conveyance device. The fall detection method according to claim 10,
The person conveying device is
A fall detection method comprising an escalator. An imaging device that acquires captured images of a plurality of frames at different timings;
The fall detection device according to any one of claims 1 to 3,
A fall detection system characterized by comprising: The fall detection system according to claim 12,
The imaging direction of the imaging device is
A fall detection system, characterized in that the fall detection system is set substantially parallel to a conveyance path of a person in a person conveyance device. The fall detection system according to claim 13,
The person conveying device is
A fall detection system characterized by including an escalator.

Images