JPH0991432A - Method for extracting doubtful person - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide the method for extracting a doubtful person by which an image from a monitor camera or the like is automatically analyzed to accurately and automatically extract the doubtful person. SOLUTION: Data of a small area of a face such as eyes and a mouth in an image being an object of monitor are collated with dictionary use image data to calculate the similarity. The similarity is discriminated based on a threshold level to extract the doubtful person whose part of face is hidden by a sun glass or a mask or the like automatically. The difference from each part of the face is obtained accurately by not collating the entire face of the center of face but collating each small area of the face with the dictionary data to extract accurately the doubtful person. The doubtful person is more accurately extracted independently of the direction of the face by using dictionary image data for face images from plural different angles so as to make the collation.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for extracting a suspicious person involved in the security field such as a bank ATM management system for automatically determining a suspicious person by analyzing image information from a surveillance camera. .

[0002]

2. Description of the Related Art A conventional suspicious person extraction method of this type basically determines whether or not a person is a suspicious person by seeing an image from a surveillance camera. Further, a method of extracting a suspicious person by image analysis is also known, but in the conventional method, the determination is made based on the degree of matching between the entire face or the central portion of the face and the dictionary image database of the face of a normal person.

[0003]

However, in the above-described conventional method for extracting a suspicious person by image analysis, the entire face or the central portion of the face is used as the matching portion,
There is a problem in that the background portion is included depending on the face orientation, which causes an erroneous determination. From such a situation, most people directly look at the image captured from the surveillance camera to determine whether or not it is a suspicious person, and the method relying on this human observation is not automated, There was a problem that the demand for labor saving could not be met.

The present invention has been made to solve the above problems, and an object thereof is to automatically analyze an image from a surveillance camera or the like to automatically and accurately extract a suspicious person. It is to provide a suspicious person extraction method that can be performed.

[0005]

In order to achieve the above-mentioned object, the present invention is a suspicious person extraction method for determining whether or not a person hides a part of a face from an image, and A first step of extracting a small area of a face such as eyes or mouth from an image and preparing it as dictionary image data;
A second step of extracting a small area of a face in the suspicious person extraction target image and calculating a correlation value with the dictionary image data obtained in the first step for each small area of the face; And a third step of determining whether or not the person is a suspicious person by comparing the correlation value for each small area of the face obtained in the step of 1. with a threshold value for determining that the person is a normal person. A suspicious person extraction method characterized by

In the above-mentioned first step, for a face image of a normal person captured from a plurality of different directions, a small area of the face is extracted and prepared as dictionary image data. This is suitable for accurately extracting the hidden suspicious person.

In the second step, the small area of the face in the target image for suspicious person extraction is extracted from the difference image with the background image to reduce erroneous determination due to inclusion of the background portion. It is suitable for more accurately extracting a suspicious person.

According to the present invention, the threshold value is used to determine the correlation value with the dictionary image data calculated for the small area of the face such as eyes and mouth in the image of the suspicious person extracted from the surveillance camera. It detects the state where a person in the image hides a part of the face with sunglasses or a mask, and automatically extracts a suspicious person. Here, by comparing with the dictionary for each small area of the face, not for the whole face or the central part, it is possible to obtain the dissimilarity of individual parts that cannot be obtained accurately for the whole face, etc. Enables extraction of suspicious persons. Further, by using the dictionary image data created by extracting the small area of the face from the face image of the normal person at a plurality of different angles, by performing the above collation, regardless of the orientation of the face,
Prevent background image information from entering the dictionary image data,
Enables more accurate determination of whether a person is a suspicious person. Furthermore, by extracting the small area of the face in the image of the monitoring target from the difference image with the background image, erroneous determination due to the inclusion of the background portion can be reduced, and a more accurate suspicious person can be extracted.

[0009]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described in detail below with reference to the drawings.

Here, a method of detecting a suspicious person in a bank ATM will be described as an example. The suspicious person referred to here is a person who hides a part of the face by using a tool such as sunglasses, glasses, a mask, etc., or by being extremely depressed.

FIG. 1 is a flow chart of a suspicious person detection method showing an embodiment of the present invention. The suspicious individual detection method includes [1] a step of extracting a face area. [2] a step of determining whether a part of the extracted face area is hidden (suspicious person determination).

First, the step of extracting the face area shown by [1] in FIG. 1 will be described. FIG. 2 is an explanatory diagram of the face area extracting method.

First, the background image data is previously captured by the set surveillance camera. Further, when viewed from the front of the surveillance camera, for example, face images of the same person are taken every 15 degrees within a range of ± 30 degrees in the vertical and horizontal directions, and the face area is divided into m _face × n _face blocks for each image. , The average of the gray values of each block is represented by a representative value. Then, this is stored in the face area search dictionary database as face area dictionary image data.

Next, the image of the user of the ATM is captured from the surveillance camera, and the image difference between this image and the background image prepared in advance is obtained. Then, it is determined whether or not there is a difference for each pixel depending on whether or not the image difference exceeds a threshold value. A difference value 100 is put in a portion having a difference, and a difference value 0 is put in a portion having no difference. The converted background difference image information is stored on the computer.

Next, the background difference image information is displayed on the x-axis (horizontal axis),
A projected image projected on each of the y-axis (vertical axis) is obtained. Here, from the histogram curve of the projected image on the x-axis, two inflection points and the distance XL between the two inflection points are raised, and from the histogram curve of the projected image on the y-axis, the histogram curve rises. A point and an inflection point that first appears, and a distance YL from a rising point of the histogram curve to the first inflection point are obtained. Then, these two values XL and YL are set as the approximate size of the face area. Based on the approximate size of the face area, the block size of the matching area is XL / m _face for the size in the x direction,
The size in the y direction is changed before and after YL / n _face . Further, the search area in the image is determined based on the approximate size of the face area and the position of each point. For example,
On the x-axis, a range between two inflection points in the histogram curve and on the y-axis, a range from the rising point of the histogram curve to the first inflection point is defined as a search area. This search area is an area that includes many pixels with a difference value of 100 in the background difference image information.

Next, the area composed of m _face × n _face blocks in the search area is used as a matching area and is matched with the dictionary image data prepared in advance in the face area searching dictionary database. Here, by determining whether or not half or more pixels of the difference value 100 of the background difference image information are included in the matching area, it is determined whether or not to match with the dictionary image data.
It is preferable in that the calculation cost can be reduced by avoiding invalid collation that is unlikely to be extracted as a face area. In the case of matching, the average of the gray values of each of m _face × n _face blocks is represented by a representative value to generate block image data.
Then, this m _face × n _face block image data is considered as a vector expression, and the inner product with the dictionary image data is taken,
Use it as a similarity measure. The matching area with the highest value is extracted as the face area.

Subsequently, it is determined whether or not a part of the extracted face area shown in [2] in FIG. 1 is hidden (determination of suspicious person).
The steps to be performed will be described. FIG. 3 is an explanatory diagram of the suspicious person determination method.

First, in the same manner as in the previous step, the eye area (divided into m _eyes × n _eyes blocks) and the mouth area (m _mouth × n _mouth blocks) of the same person as the dictionary image data of the face area are divided. (Division) dictionary image data is prepared for each. That is, for each of the face images of the same person taken every 15 degrees within the range of ± 30 degrees in the vertical and horizontal directions when viewed from the front of the surveillance camera, the eye area is m _eyes × n _eyes.
The mouth area is divided into m _mouth × n _mouth blocks and the average of the gray values of each block is represented by a representative value. The eye area is stored in the eye area searching dictionary database as eye area dictionary image data, and the mouth area is stored in the eye area searching dictionary database as eye area dictionary image data.

Next, the face area obtained in the previous step is searched, block image data of the collation area is generated in the same manner as in the previous step, and the highest degree of similarity with the above-mentioned eye area dictionary image data is obtained. The indicated area is obtained as the eye area, and the area having the highest degree of similarity with the mouth area dictionary image data is obtained as the mouth area. As the similarity used here, for example, the inner product of the block image data used in the previous stage and the dictionary image data can be used.

Finally, if at least one of the eye area and the mouth area has the above-mentioned degree of similarity lower than the set threshold value, it is judged as a suspicious person. In addition, there is an example in which the similarity is higher as the value indicating the correlation is smaller, and in general, when the similarity exceeds the set threshold and the similarity is low, the suspicious person is determined.

[0021]

As is apparent from the above description, according to the present invention, a target image from a surveillance camera or the like is automatically analyzed by a computer or the like, so that a suspicious person is automatically extracted. be able to. At that time, the matching is performed not for the whole face or the central part but for each small area of the face, so that it becomes possible to obtain the dissimilarity of each part that cannot be obtained accurately for the whole face, etc. A suspicious person can be extracted.

Further, when the above collation is performed using the dictionary image data created by extracting small areas of the face from a plurality of face images at different angles, the orientation of the face is not particularly limited. , It is possible to more accurately extract a suspicious person.

Further, when the small area of the face in the image to be monitored is taken out from the difference image with the background area, erroneous determination due to the inclusion of the background image can be reduced, and more accurate. Enables extraction of suspicious persons.

[Brief description of drawings]

FIG. 1 is a flowchart of a suspicious individual detection method according to an exemplary embodiment of the present invention.

FIG. 2 is a diagram illustrating an example of a face area extraction method in the suspicious individual detection method of the above-described embodiment.

FIG. 3 is a diagram illustrating an example of a suspicious person determination method in the suspicious person detection method of the above-described embodiment.

Claims (3)

[Claims] 1. A suspicious person extraction method for determining whether or not a human hides a part of a face from an image, and extracts a small area of a face such as an eye or a mouth from a face image of a normal person for a dictionary. First prepared as image data
And a second step of extracting a small area of the face in the suspicious person extraction target image and calculating a correlation value with the dictionary image data obtained in the first step for each small area of the face. A third step of determining whether a person is a suspicious person by comparing the correlation value for each small area of the face obtained in the second step with a threshold value for determining that the person is a normal person; A method for extracting a suspicious person, comprising: 2. The suspicious individual according to claim 1, wherein in the first stage, a small area of the face is extracted and prepared as dictionary image data for a face image of a normal person captured from a plurality of different directions. Extraction method. 3. The second step is to extract a small area of the face in the suspicious person extraction target image from the difference image with the background image.
Alternatively, the suspicious person extraction method according to claim 2.