JPH08272926A - Detecting device for number of people - Google Patents

Abstract

PURPOSE: To detect the number of people at a high speed with high precision irrelevantly of the degree of congestion of people or changes of clothing in seasons by reducing the influence of a change in the installation place or installation environment of a television camera in a case wherein the contrast between object people and a floor as a background is low or in a case wherein luminance varies with a lighting state, etc., and when the number of the peoples is detected by processing the video signal of the television camera. CONSTITUTION: The device which picks up the image of a people to be detected by the television camera is provided with a feature extraction part 6 consisting of an arithmetic means which updates a background image in the absence of people, a shape correcting means which corrects the binarized image of a digital image signal, a pixel number compressing means which decreases the number of pixels of the image obtained by the shape correcting means, and a correlative value detecting means which finds a correlative value by superposing a specific pattern on the image obtained by the pixel number compressing means, and a people number judging part 7 which input correlative value parameters obtained by arithmetic processing based on the correlative value, density parameters showing the density level of the image signal from the television camera, and environmental parameters featuring the installation place of the television camera to a neural network.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an apparatus for detecting the number of people in a specific place such as a department store, an exhibition hall, a building, an underground mall, etc. by processing a video signal of a television camera.

[0002]

2. Description of the Related Art A method of converting a video signal of a television camera into a digital signal and processing an image to be detected from the obtained digital image signal is widely used. For example, FIG. 5 is a block diagram of a conventional people detection device. Figure 5
In FIG. 1, 1 is an image input unit configured by a television camera or the like for capturing an image of a person to be detected and outputting an image signal, and 2 is an image signal input from the image input unit 1 for removing noise components and performing binarization or the like. A preprocessing unit for processing and outputting a binarized image signal, 3
Is an object extraction unit for inputting the image signal of the preprocessing unit 2 to extract a detection target person image, and 4 is a target person image of the object extraction unit 3 and a reference area image which is input in advance and stored therein. An area comparison unit 5 for comparison is a counting unit for counting images larger than the reference area image.

Next, the operation will be described. Image input section 1
When an image of a specific place such as a department store or an exhibition place is output and an image signal is output, the preprocessing unit 2 takes in the image signal and removes noise components by masking, and then performs binarization processing. Output the binarized image signal. The target extraction unit 3 extracts a detection target person image from the binarized image signal. A reference area image input in advance is stored in the area comparison unit 4, and the reference area image is compared with the detection target person image of the target extraction unit 3 to obtain a person image of a portion larger than the reference area image. The counting unit 5 counts these images and outputs a person number signal.

By thus digitally processing the image signal of the image input section, the number of persons to be detected can be detected.

[0005]

In the conventional number-of-people detecting device as described above, it is necessary to confirm that there is no person in the background image of the reference area and manually input the background image within the field of view of the camera. Every time the installation environment changed, such as when you put something, you had to input a new standard background image each time. Since area determination is performed based on a simpler binarized image signal, when a large number of people are crowded or when the contrast between the people and the background is low, for example, clothes and hair color become the background. When the color is similar to the floor, the image area (number of pixels) of the detection target person becomes small, and the color of clothes and the state of clothes change depending on the season. In addition, it was difficult to set the threshold level when performing the binarization process due to the change in the brightness of the installation location of the image input unit. In addition, when a television camera provided on the ceiling is used to capture an image of a person to be detected, the detection accuracy is lowered because the image area of the person to be detected varies depending on the mounting height of the television camera. After all, the detection accuracy of the conventional people detection device is greatly affected by the degree of congestion of people, clothes, installation location of TV camera, and installation environment.

The present invention has been made in order to solve the above-mentioned problems, and a background image in a state without a person can be automatically updated, and the brightness depending on the degree of congestion of people, clothes, floor color, lighting condition, etc. It is possible to obtain a person number detection device that can detect the number of people with high accuracy and high speed without being affected by the installation location or installation environment of the TV camera by reducing the influence of changes in the height. To aim.

[0007]

A person number detecting apparatus for solving the above-mentioned problems includes a calculating means for updating a background image in a state where there is no person, and a shape correcting means for correcting a binarized image of a digital image signal. A pixel number compression means for reducing the number of pixels of the image obtained by the shape correction means, and a correlation value detection means for obtaining a correlation value by superimposing a predetermined pattern on the image obtained by the pixel number compression means. A characteristic extraction unit, and a correlation value parameter obtained by applying a calculation process based on the correlation value,
It is configured by a number-of-persons discriminating unit having a neural network as an input, a density parameter representing a density level of an image signal from the television camera, and an environment parameter characterizing an installation environment of the television camera.

[0008]

Since the number-of-people detection device according to the present invention automatically updates the background image that serves as a reference when detecting the number of people, it is artificial even if the background changes, such as when an object is placed on the floor. There is no need to change the settings manually. In addition, even if people are crowded, clothes change, there is a change in brightness due to lighting conditions, and even the mounting height of the TV camera changes, the learning function of the neural network is used. Since the number of persons is discriminated, the number of persons to be detected can be detected with high accuracy and high speed.

[0009]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below with reference to FIGS. FIG. 1 is a block diagram showing the configuration of the number-of-people detecting device of the present invention.

Reference numeral 1 is an image input unit composed of the same television camera as in FIG. 5, and 2 is also the same as in FIG. 5, and the image signal of the image input unit 1 is input to remove noise components and perform processing such as binarization. A pre-processing unit 6 for outputting a binarized image signal and a feature extraction unit 6. The feature extraction unit 6 corrects the shape of the binarized image of the pre-processing unit 2 and further speeds up the calculation process. Correlation value detection processing is performed to calculate the image area of the overlapped portion by overlapping the person's pattern, and use this area as the correlation value of the pixel at the center position of the superimposed pattern, and obtain this over the entire image. Reference numeral 7 is a number-of-persons discriminating unit, which uses a neural network for discriminating logic of the number of persons. An output unit 8 displays the number of detected persons and prints out.

FIG. 2 is a flow chart showing the procedure of the background image updating process. When there is no person, the background image is automatically updated.

First, two images are continuously captured (procedure a). The commercial frequency is 50 H
Considering that z and 60 Hz, n / 20 seconds (n
Must be taken in at intervals of 1 or more) (especially for fluorescent lighting). Therefore, for example, capture is performed at intervals of 0.1 second.

Next, the absolute value of the difference between the two captured images (this is called the moving difference) is obtained (procedure b), and the difference image is binarized with a threshold value capable of removing noise (procedure). c), the time-varying part of the image, in this case, only the part where the person has moved is cut out.

After that, the shape of the image is corrected (procedure d). That is, in the peripheral portion of the detected image, the human image appears longer than in the central portion. Therefore, the image of the peripheral portion is compressed and the shape of the image is corrected so that the size of the image for one person becomes substantially equal. The shape correction may be determined such that a pixel having a longer distance from the image center has a larger movement amount to the image center. In the shape-corrected image, in order to shorten the processing time, a plurality of pixels are combined into one by compressing the number of pixels (procedure e), and the correlation value is calculated based on this (procedure f). First, an appropriate pattern close to the size of a person (for example, a square pattern of 3 × 3 pixels) is prepared, and the pattern is scanned pixel by pixel to obtain a correlation value between the pattern and the image. That is, each value (0 or 1) in the pattern and the pixel value (0
Alternatively, the value obtained by multiplying by 1) and adding the value over the entire pattern is set as the correlation value of the pixel at the center of the pattern. Then, the correlation value becomes high at the position where the pattern and the image shape of the person match well.

Since the peak of the correlation value corresponds to a person, it is determined whether or not there is a person by counting the number of peaks in step g. If the number of people is 0 (the number of peaks is 0), this is repeatedly executed, and when the number of people is 0 in m consecutive times (procedure h), it is considered that there is no person, and Save one image as a background image (procedure i). Here, the procedure h is performed because it is difficult to detect the image of the person by the movement difference when the lighting condition is bad or when there is almost no movement of the person. This is to prevent it from being saved as an image.

Next, the processing procedure for detecting the number of people will be described with reference to the flowchart of FIG.

In FIG. 3, first, an image for which the number of people is desired to be detected is captured in one screen (procedure j), and the absolute value of the difference between this image and the background image in the state where there is no person is stored in procedure i of FIG. (This is called a fixed difference) is calculated (procedure k). The characteristic of this fixed difference is that, unlike the case of moving difference,
Even if there is no movement of a person, an image of the person can be obtained if there is contrast with the background. Next, this fixed difference image is given a fixed density level (this is called a binarization level).
Is used as a threshold value (procedure 1). In step m,
The shape of the image is corrected by the same method as the procedure d in FIG. Furthermore, in order to increase the calculation speed of the number of people detection,
The number of pixels of the image is reduced (procedure n), and a pattern of a person having a predetermined size is superimposed on the reduced image to obtain a correlation value (procedure p). At this time, if the person's pattern is about the same size as the person's image and is as small as possible, the calculation speed can be increased and the accuracy can be improved.

Next, the parameters to be input to the number-of-persons discriminating section 7 shown in FIG. 1 are calculated. These parameters are
It is a characteristic parameter given to the input layer of the neural network described later, and is composed of the following three types of parameter groups.

(1) Correlation value parameter The correlation value parameter calculated in the procedure q of FIG. 3 is (1-1) the number of pixels whose correlation value is a predetermined value or more. (1-2) A value obtained by dividing the number of pixels having a correlation value equal to or larger than a predetermined value by the number of pixels of a predetermined pattern (this is called a correlation value coefficient). (1-3) Maximum peak value of correlation value. (1-4) The number of peaks of the correlation value. There are four, but at least one of these is normalized and given to the input layer of the neural network described later.

(2) Density Parameter The density parameter calculated in step r of FIG. 3 is (2-1) the binarization level of the difference image obtained in step l. (2-2) The density gradation of the background image obtained in step i of FIG. 2 or FIG. , And at least one of them is normalized and given to the input layer of the neural network described later.

(3) Environmental parameter The environmental parameter calculated in step s of FIG. 3 is (3-1) an amount that depends on the mounting height of the television camera. That is, since the size of the image of a person changes depending on the mounting height of the television camera, it is necessary to change the size of the pattern of the person when obtaining the correlation value in step p. The amount depending on the mounting height of the television camera is defined as the normalized size of the human pattern. (3-2) Season-dependent quantity. In other words, people's clothing changes in color and condition depending on the season. Especially in summer and winter, the difference is usually large. For example, the main system is white in summer, and black or navy in winter. The quantity depending on the season is mainly defined as the normalized clothing information. At least one of the above parameters is given to the input layer of the neural network described later.

Step t is a neural network, and the number of persons is discriminated by giving the above-mentioned correlation value parameter, density parameter, and environment parameter to the input layer of the neural network.

FIG. 4 shows the structure of the neural network. The neural network used here is a hierarchical network based on the error back propagation method, and has a three-layer structure including an input layer 10, an intermediate layer 11, and an output layer 12. Input layer 10
Is input with the correlation value parameter, the concentration parameter, and the environment parameter obtained in the procedure q, the procedure r, and the procedure s of FIG. 3, respectively, and the number of persons (0
Set people to N people. The intermediate layer 11 may be determined experimentally from the viewpoint of the speed of error convergence and the amount of calculation, but for example, the number of units is half the total number of units in the input / output layers.
The threshold function of the unit is a sigmoid function (not shown), a well-known delta rule is used as a learning rule, and a network is constructed by recursive learning that modifies the weight of the connection between units.

As described above, according to the neural network of the present embodiment, since the number of parameters given to the input layer can be small, the network can be smaller in scale than the case where the human image is directly input to the neural network, and the rotation of the pattern and It has the feature that it does not depend on the size.

[0025]

As described above, according to the present invention, the background image, which is the reference when detecting the number of people, is automatically updated. Even if it changes, there is no need to artificially change the setting. In addition, people may be crowded, clothes may change depending on the season, brightness may change due to lighting conditions, etc.
Furthermore, even if there is a change in the installation location or installation environment such as a change in the mounting height of the TV camera, the small-scale neural network configuration has the effect of being able to detect the number of people with high accuracy and high speed.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of a person number detection apparatus according to an embodiment of the present invention.

FIG. 2 is a flowchart showing a background image update process.

FIG. 3 is a flowchart illustrating a processing operation for detecting the number of people.

FIG. 4 is a configuration diagram of a neural network.

FIG. 5 is a block diagram showing a configuration of a conventional people detection device.

[Explanation of symbols]

 1 Image Input Section 2 Pre-Processing Section 3 Target Extraction Section 4 Area Comparison Section 5 Counting Section 6 Feature Extraction Section 7 Number of People Discrimination Section 8 Output Section 10 Input Layer 11 Intermediate Layer 12 Output Layer

Claims (4)

[Claims] 1. A person number detecting device for converting a video signal of a television camera into a digital signal, calculating a person image using a binarization level of the obtained digital image signal as a threshold value, and detecting the number of persons, Arithmetic means for updating the background image in the absence of people, shape correction means for correcting the binarized image of the digital image signal, pixel number compression means for reducing the number of pixels of the image obtained by the shape correction means, and the pixels The image obtained by the number compression means, a correlation value detection means for obtaining a correlation value by superimposing a predetermined pattern on the image, a correlation value parameter obtained by performing a calculation process based on the correlation value, A person who uses the density parameter representing the density level of the image signal from the TV camera and the environment parameter characterizing the installation environment of the TV camera as an input layer of the neural network. An apparatus for detecting the number of people, comprising: a number determination unit. 2. The correlation value parameter is the number of pixels having a correlation value of a predetermined value or more, or a value obtained by dividing the number of pixels having a correlation value of a predetermined value or more by the number of pixels of a predetermined pattern, or the maximum peak value of the correlation value, Alternatively, the number-of-persons detecting device according to claim 1, wherein the number of peaks of the correlation value is at least one element. 3. The number of people according to claim 1, wherein the density parameter is at least one of a binary level of the digital image signal and a density gradation of a background in the absence of a person. Detection device. 4. The number-of-people detection device according to claim 1, wherein the environmental parameter is at least one element of an amount depending on a mounting height of the television camera or an amount depending on a season.