JP2013114583A - Moving direction identification device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To improve monitoring efficiency by estimating a moving direction of a detected person and dismissing detection of a person outside a monitoring target from the moving direction.SOLUTION: A moving direction identification device comprises: a time change extraction unit 2 for dividing each of consecutive images on a time axis into multiple areas and calculating an intensity change within the areas as a state change quantity for each divided area; a movement vector calculation unit 3 for calculating a difference in the calculated state change quantity for each image frame group that consists of multiple consecutive images, calculating a centroid of the calculated difference, comparing a position of the calculated centroid among the consecutive image frame groups and calculating a movement quantity of the centroid as a feature quantity; a person detection unit 4 for determining the presence of a person in a picture; and a moving direction determination unit 5 for determining a moving direction of the person. The moving direction determination unit 5 consists of a cascade strong identifier by AdaBoost, compares the feature quantity stored in a sample storage unit 5a with the calculated feature quantity of each area and determines the moving direction of the person.

Description

  The present invention relates to a movement direction identification device that identifies a movement direction of a person from an image captured by a camera.

  In recent years, a technique for detecting a person from an image has been developed, and as a technique for performing identification based on a feature amount obtained by histogramating luminance changes of pixels in a local region, a feature vector HOG that is a histogram of luminance gradient direction is used. There is a method disclosed in Non-Patent Document 1 using the image processing method, or a method disclosed in Patent Document 1 used as a feature amount by performing pixel state analysis (PSA) based on a change in luminance on the time axis in a plurality of image frames. Proposed.

JP 2009-301104 A

N.Dalal and B. Triggs, “Histograms of Oriented Gradients for Human Detection”, IEEE Computer Vision and Pattern Recognitio, pp. 886-893, 2005

  However, if you want to monitor only the person who is approaching the camera using the camera, or who is moving in a specific direction, but who actually crosses the camera regardless of the direction of movement, But there was a problem that would detect it.

  Therefore, in view of such problems, the present invention estimates the moving direction of the detected person and rejects the detection of the person who is not the monitoring target from the moving direction, so that the moving direction can be improved. The object is to provide a device.

In order to solve the above-mentioned problem, the moving direction identification device according to the first aspect of the present invention is an image frame group that generates an image frame group composed of a plurality of images that are continuous on the time axis based on the captured image of the camera. Calculated in the generation means, the state change amount calculation means for dividing the individual image into a plurality of regions, and calculating the brightness change in the region as the state change amount for each divided region, and the generated individual image frame group The moving area calculating means for calculating the difference between successive images of the state change amount as the moving area, the center of gravity calculating means for calculating the center of gravity of the calculated moving area, and the position of the center of gravity are compared between successive image frame groups. A feature amount calculation means for calculating the amount of movement of the center of gravity as a feature amount, a person detection unit for determining the presence of a person in the video, and a movement direction determination unit for determining the movement direction of the person. Department A sample storage unit that stores feature amounts calculated from learning samples classified according to direction, and is configured by a cascade-type strong classifier using Adaboost, and when the person detection unit determines the presence of a person, the sample storage The moving direction of the detected person is determined by comparing the feature amount stored in the section with the calculated feature amount of each region.
According to this configuration, the movement direction is identified by determining the presence of a person from a region where the luminance changes between a plurality of images that are continuous on the time axis. Therefore, even if the person being imaged is grasped, any person who moves outside the monitoring target can be ignored and the monitoring efficiency can be improved.

According to a second aspect of the present invention, in the configuration of the first aspect, the person detection unit includes a sample storage unit that stores a feature amount calculated from a learning sample classified according to the presence or absence of a person, and is based on Adaboost It is configured by a cascade type strong classifier and is characterized in that the presence of a person is determined by comparing the feature amount stored in the sample storage unit with the calculated feature amount of each region.
According to this configuration, it is possible to prevent a pillar, planting, or the like from being determined as a person, and it is possible to prevent a sudden change in brightness from being determined as a person, thereby improving the detection accuracy of the person.

  According to the present invention, even if the person being imaged is grasped from the captured image of the camera, the person who moves outside the monitoring target can be ignored, and the monitoring efficiency can be improved.

It is a block diagram which shows an example of the moving direction identification apparatus which concerns on this invention. It is explanatory drawing which shows the time series relationship of an image and an image frame group. It is a conceptual explanatory drawing which shows the flow which extracts an object.

  Embodiments of the present invention will be described below in detail with reference to the drawings. FIG. 1 is a block diagram of a moving direction identification device according to the present invention. An image input unit 1, a time change extraction unit 2, a motion vector extraction unit 3, a person detection unit 4, a movement direction determination unit 5, and a result output unit 6 are shown. I have. The time change extraction unit 2, the motion vector extraction unit 3, the person detection unit 4, and the movement direction determination unit 5 are integrally configured by a CPU or DSP in which a predetermined program is installed.

  The image input unit 1 includes a camera for imaging a person to be detected, and outputs a digital video signal. This video signal is, for example, a 256-color color video.

The time change extraction unit 2 sequentially extracts images (image frames) from the video signal output from the image input unit 1 at regular time intervals, and groups them for each predetermined number of images to generate an image frame group. FIG. 2 shows a time-series relationship between the extracted image and the image frame group, and shows a case where one image frame group is composed of three frame images.
Further, the time change extraction unit 2 divides each image into a plurality of regions, and calculates the luminance change in the region as the state change amount from the luminance information of the pixel for each divided region.

  The motion vector extraction unit 3 compares the state change amounts between the same regions of adjacent images on the time axis, and calculates a difference. Then, the centroid position of the difference obtained for each image frame group is obtained, and the movement amount of the centroid between adjacent image frame groups is calculated as a feature amount.

The person detection unit 4 uses the learning samples classified by the person and the background and the feature quantity extracted by the motion vector extraction unit 3 to use a known cascade booster classifier made up of strong boosters. Determine the existence of Reference numeral 4a denotes a sample storage unit that stores feature quantities of learning samples.
Note that the presence of a person may be determined by detecting a person from a camera-captured video by the technique of Non-Patent Document 1.

  The moving direction determination unit 5 uses a well-known cascade-type discriminator based on Adaboost using a learning class classified by person and background and the feature quantity extracted by the motion vector extraction unit 3. The moving direction of the person is determined. Reference numeral 5a denotes a sample storage unit that stores the feature amount of the learning sample.

  The result output unit 6 includes a sound output unit such as a speaker or a buzzer for reporting sound, a display unit for displaying an image, and the like, and when the movement direction determination unit 5 detects a person moving in a specific direction set in advance. , Notify the sound and display the captured image.

  The operation of the moving direction identification device configured as described above will be specifically described below. FIG. 3 is a conceptual explanatory diagram showing a flow of determining a person detection and moving direction, and will be described with reference to FIG. Here, the person detection by the person detection unit 4 and the determination of the movement direction of the person by the movement direction determination unit 5 are performed with the same flow except that learning samples described later are different.

  First, the time change extraction unit 2 extracts, for example, continuous images (image frames) P1, P2,... Pn at intervals of 100 ms from the camera-captured video of the image input unit 1, and groups them by a predetermined number as shown in FIG. A group of image frames is generated. The preceding and following image frame groups may have a common image.

In the time change extraction unit 2, each individual image is further divided into a plurality of preset regions, and the state change amount in the region is calculated from the luminance information of the pixel for each divided region. FIG. 3A shows a state where each image is divided into a plurality of regions, and shows a state where one image is divided into 12 regions. FIG. 3B is an explanatory diagram of the state change amount. Here, as an example, the change in the divided area indicated by E in FIG. 3A is shown. Actually, the state change amount is calculated in all regions.
Here, for convenience of explanation, a case is shown in which individual image frame groups are composed of two frame images. Also, F in FIG. 3A indicates a person who has entered the imaging range of the camera of the image input unit 1.

Next, the motion vector extraction unit 3 calculates a difference in state change amount between the same regions of adjacent images on the time axis as a moving region, and obtains the center of gravity of the moving region of each divided region for each image frame group. It is done. FIG. 3C shows the center of gravity G of the region E thus obtained for each image frame group.
Further, the positions of the centers of gravity between adjacent image frame groups are compared, and the amount of movement is calculated as a feature amount. FIG. 3D shows the calculated feature amount, and the feature amount is calculated as a vector V.

When the feature amount of each region is calculated, first, the feature amount obtained by the person detection unit 4 is compared with the feature amount of the learning sample stored in the sample storage unit 4a. At this time, a person is detected using a well-known AdaBoost cascade classifier constituted by a strong classifier.
As the learning sample used at this time, data classified according to the presence or absence of a person is used. Specifically, one is, for example, shooting several hundred people of various postures or movements, and as a result, a feature quantity indicating a person's tendency is calculated from the image, and a probability density function is generated and stored. . As another class, feature quantities indicating a tendency other than a person are calculated as a result from, for example, several hundred captured images obtained by shooting a person other than a person, and a probability density function is generated and stored.
Using the generated probability density function as a learning sample, the feature amount calculated from the image input from the image input unit 1 is compared with the learning sample to detect a person.

If the person detection unit 4 detects a person, the movement direction determination unit 5 next determines the movement direction of the detected person. Also in this determination, a well-known cascade booster discriminator using Adaboost composed of a strong discriminator is used, and the determination is made by comparing with the feature amount of the learning sample.
As the learning sample used at this time, data classified according to the movement direction of the person is used. Specifically, when it is desired to detect a person moving in the direction approaching the camera, the feature amount is calculated from, for example, several hundred images obtained by photographing the person approaching as one class, and the probability density A function is generated and stored. In addition, for example, hundreds of images taken of a person who moves in a horizontal direction or a direction away from the person other than the approaching action, as a result, a feature amount indicating a direction tendency other than the approaching direction is calculated, and its probability density function Is generated and stored as a separate class.
Using the generated probability density function as a learning sample, the feature amount calculated from the image input from the image input unit 1 is compared with the learning sample to determine the moving direction.

  As a result of the determination, if it is determined that there is a person approaching, the result output unit 6 emits an alarm sound and displays a captured image of the camera.

In this way, the feature quantity calculated from a plurality of images that are continuous on the time axis is compared with the learning sample classified according to direction to determine the moving direction. A person who moves outside can be set to be ignored, and the monitoring efficiency can be improved.
Also, when detecting a person, it is possible to determine a pillar, planting, or the like as a person by comparing a feature amount calculated from a plurality of continuous images on the time axis with a learning sample and determining a moving direction. In addition, it is possible to prevent a sudden change in brightness from being determined as a person, and the detection accuracy of the person is improved.

  In the embodiment, the calculated luminance change amount (state change amount) includes a color change, a texture change, and an edge change.

  1. Image input unit 2. Time change extracting unit (image frame group generating unit, state change amount calculating unit) 3. Motion vector extracting unit (moving region calculating unit, center of gravity calculating unit, feature amount calculating unit) 4 .. Person detection unit, 4 a .. Sample storage unit, 5... Moving direction determination unit, 5 a .. Sample storage unit, 6.

Claims (2)

Image frame group generation means for generating an image frame group composed of a plurality of images continuous on the time axis based on a captured image of the camera;
State change amount calculating means for dividing each image into a plurality of regions and calculating a luminance change in the region as a state change amount for each divided region;
In each of the generated image frame groups, a moving area calculating unit that calculates a difference between consecutive images of the calculated state change amount as a moving area;
Centroid calculating means for calculating the centroid of the calculated moving area;
A feature amount calculating means for comparing the position of the center of gravity between successive image frame groups and calculating the amount of movement of the center of gravity as a feature amount;
A person detection unit that determines the presence of a person in the video;
A moving direction determination unit that determines a moving direction of the person,
The moving direction determination unit includes a sample storage unit that stores the feature amounts calculated from learning samples classified according to directions, and is configured by a cascade type strong classifier using Adaboost, and the person detection unit includes: When the presence of a person is determined, a moving direction identification device that determines the moving direction of the detected person by comparing the feature amount stored in the sample storage unit with the calculated feature amount of each region. . The person detection unit includes a sample storage unit that stores the feature amount calculated from a learning sample classified according to the presence or absence of a person, and is configured by a cascade type strong classifier using Adaboost, the sample storage unit The moving direction identification device according to claim 1, wherein the presence of a person is determined by comparing the feature quantity stored in the field and the calculated feature quantity of each region.

Images