JP2015158795A - Object detection device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an object detection device which can reduce a calculation amount by focusing on the ratio of dynamic state pixels to static state pixels in a cell and eliminating a vector operation.SOLUTION: An object detection device comprises: an image input part 1 which generates an image frame group; a temporal change extraction part 2 which calculates the dispersion value of luminance from luminance change information on the same portion for each pixel of a reference pixel and an image frame preceding by a predetermined number or after the predetermined number from the reference frame to determine that the pixel with the dispersion value of the luminance larger than a prescribed dispersion value is in the dynamic state and pixels in other states are in the static state; an inter-frame feature extraction part 3 which forms a cell consisting of a plurality of adjacent pixels in the reference frame, calculating the ratio of the dynamic state pixels to the static state pixels for each cell with addition and division, and normalizes it to calculate the motion feature quantity; and a detection part 4 which detects a moving object by generating a strong discriminator on the basis of a weak discriminator generated from the calculated motion feature quantity and a learning sample.

Description

  The present invention relates to an object detection device that detects an object from a captured image of a camera.

In recent years, with the increase and sophistication of crime, the demand for security cameras with high security functions is increasing even in ordinary households. Such a conventional security camera is configured so that the camera is activated and starts imaging when the human sensor senses, so that even if a suspicious person is in the imaging range of the camera, it may be located away from the camera. Since it was not captured, it was insufficient for crime prevention.
For this reason, the inventors of the present invention proposed a technique for detecting a moving object by video processing in Patent Document 1.

JP 2013-190943 A

The object detection apparatus disclosed in Patent Document 1 can detect an object such as a person or a vehicle with an inexpensive CPU by reducing the amount of calculation, and can be applied to a security camera installed in a general home.
However, since this calculation amount requires only two dimensions for one cell, the calculation amount can be reduced as compared with the calculation using the HOG feature, but the vector calculation (square sum) when extracting the feature In the cascade type discriminator based on AdaBoost and division with the square root of the denominator, time-consuming processing has occurred, and the amount of calculation has not been sufficiently reduced until it can be implemented by an inexpensive CPU.

  Therefore, in view of such problems, the present invention focuses on the ratio of moving state pixels and static state pixels in a cell instead of vector intensity, and also enables object detection that can reduce the amount of calculation by simple co-occurrence expression. The object is to provide a device.

  In order to solve the above-mentioned problems, the invention of claim 1 is directed to an image frame group generation means for generating an image frame group consisting of continuous images on a time axis based on a captured image of a camera, and a plurality of generated image frames Any image frame is used as a reference frame, and the maximum value of the change in luminance of the same part for each pixel of the image frame up to a predetermined number on the time axis from the reference frame or a predetermined number on the time axis is calculated. Brightness change calculation means, and a brightness variance value is calculated from the reference frame and the brightness change information of the same part for each pixel of the image frame up to a predetermined number on the time axis from the reference frame or a predetermined number on the time axis. A variance value calculating means for determining a pixel having a maximum luminance change amount greater than a predetermined value and / or a pixel having a luminance variance value greater than a predetermined value as a moving state A pixel state determination means for determining pixels in other states as a static state and a cell composed of a plurality of adjacent pixels in the reference frame are formed, and the ratio of the moving state pixel and the static state pixel is added and divided for each cell. From the weak classifier by learning, the feature quantity calculating means for calculating the feature quantity of motion, the weak classifier generating means for generating a weak classifier expressing the co-occurrence from the feature quantity and the learning sample, And an object detecting means for detecting an object moving by the generated strong classifier.

  According to the present invention, when extracting a feature value, the feature value is calculated based on the ratio of the moving state pixel and the static state pixel obtained by addition and division, thereby performing a troublesome calculation for calculating the vector intensity. This eliminates the need for this, and the amount of computation can be reduced by using a simple co-occurrence expression compared to the conventional cascade booster using Adaboost, thereby reducing the load on the calculation means and detecting objects such as people and vehicles. Can be realized at high speed and at low cost.

It is a functional lineblock diagram showing an example of an object detection device concerning the present invention. It is explanatory drawing of the image frame group which shows the relationship between an image frame and a luminance value. It is an image explanatory drawing which shows the flow of the calculation which extracts a moving body with an image.

DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments embodying the present invention will be described below in detail with reference to the drawings. FIG. 1 is a functional configuration diagram showing an example of an object detection apparatus according to the present invention, which receives a video signal captured by a camera (not shown) and generates an image frame composed of images that are continuous on a time axis at a predetermined interval. An image input unit 1 that outputs the image, a time change extraction unit 2 that extracts a time change amount of luminance for each pixel using an arbitrary image frame as a reference frame, and an inter-frame feature that calculates a feature amount serving as an inter-frame difference feature for each pixel An extraction unit 3, a detection unit 4 that detects an object in an image frame by a strong classifier using the learning sample and the calculated feature amount, and a result output unit 5 that outputs a detection result are provided.
Of these components, the image input unit 1, the time change extraction unit 2, the inter-frame feature extraction unit 3, and the detection unit 4 constitute an object detection device, and the CPU or DSP executes a predetermined program. Is done. The whole including the result output unit 5 can be realized by a personal computer, for example.

  Hereinafter, the operation of each unit will be described in order. The image input unit 1 receives, for example, an imaging video signal of a night vision camera that is equipped with infrared illumination and can be monitored at night, and generates and outputs an image frame (still image), for example, every 0.03 seconds. In this way, the image input unit 1 outputs an image frame composed of digital data continuous on the time axis.

The time change extraction unit 2 calculates a luminance dispersion value S from an arbitrary pixel of the image frame output by the image input unit 1. This operation is finally performed on all pixels in the frame.
FIG. 2 is an explanatory diagram of an image frame group showing the relationship between the image frame and the luminance value. A reference frame is K, a past frame that is an image frame output immediately before the reference frame K, and an image frame that is output immediately thereafter. The relationship of the future frame which is an image frame is shown. In addition, the luminance value of a pixel at an arbitrary coordinate in the reference frame K is denoted by Ik, and the luminance at the same coordinate in each image frame is denoted by Ik (± n). Hereinafter, a description will be given with reference to FIG.

The variance value S of the pixels includes a variance value (first variance value) S1 from N past frames K-N to a reference frame K, and a variance value (second variance) from the reference frame K to N future frames K + N. Value) S2 is calculated.
The variance values S1 and S2 are calculated by, for example, Equation 1. In Equation 1, j is an integer from 1 to n.

Next, threshold processing is performed on the dispersion value thus obtained. The static state / moving state is determined by threshold processing. Specifically, if the first variance value S1 and the second variance value S2 are equal to or greater than the threshold value, it is determined to be a moving state, and the others are determined to be a static state, and thus binarized for each pixel (two gradations and Output brightness information.
Note that the threshold value may be a fixed value set in advance, but in the case of a fixed value, the determination varies greatly depending on the installation environment and brightness of the camera. It is preferable to determine an appropriate value for.

The inter-frame feature extraction unit 3 performs a predetermined calculation from the obtained moving state pixels and static state pixels of the entire image to normalize and calculate a feature amount.
Normalization is calculated based on the determination result of the static state / moving state for each pixel calculated by the time change extraction unit 2, and the feature amount is formed by a plurality of cells formed by a plurality of adjacent pixels. Calculated in units of blocks.
For example, one cell is composed of 5 × 5 pixels, and one block is composed of 3 × 3 cells.

FIG. 3 is an explanatory diagram of the feature amount calculation. B1 and B2 indicate blocks in the reference frame K, and C indicates a cell. Block B1 represents moving state pixels as white and still state pixels as black, and block B2 represents still state pixels as white and moving state pixels as black.
When the dynamic state pixel (white) shown in B1 is “1” and the static state pixel (black) is “0”, the normalized feature value Va when the static state pixel (black) is “0”, the static state pixel (white) shown in B2 is “1”, The normalized feature value Vr when the moving state pixel (black) is “0” is calculated by the following equation (2).

  Here, a, b,..., R are the count numbers of the static state / moving state of each cell, and as shown in Equation 2, here, the feature amount of the moving state and static state of the cell of interest is simply added. It is a numerical value calculated by the ratio of the block to the number of dynamic states obtained by division and the ratio of the block to the number of static states.

The detection unit 4 extracts a weak classifier that simply expresses a co-occurrence from the features calculated by the inter-frame feature quantity extraction unit 3, and generates a strong classifier through selection of the weak classifier by learning. , An object (here, a case where a person is detected will be described) is detected. Simple co-occurrence is calculated by, for example, Equation 3.
In Equation 3, h _{a · r} and h _{a + r} are the weak discriminator outputs, W ^j _{a +} is the sum of weights according to the value of the feature value Va of all positive samples, and W ^j _{r +} is the feature value Vr of all positive samples. The sum of the weights according to the values, W ^j _a− is the sum of the weights according to the value of the feature value Va of all negative samples, and W ^j _r− is the sum of the weights according to the value of Vr of all the negative samples.

The strong classifier is constructed by Real Adaboost, learns from a learning sample consisting of a human image to be detected and other images, and selects a more optimal cell combination to obtain a final strong classifier. When a person is detected by the strong classifier thus obtained, an object detection signal including the person's coordinate information, area information, etc. is output.
In this case, the learning sample used in this real Adaboost is a learning sample that is classified according to the person and the background to detect the person. Learning samples classified by background are used.

  The result output unit 5 includes a sound report unit that sounds an alarm when the detection unit 4 detects a person, a video display unit such as an LCD monitor that displays the detected video, a notification unit that reports to the outside, and the like. In response to the coordinate information output by 4, the video display unit displays a human area in addition to the display of the input video.

  As described above, when calculating the feature amount, it is necessary to perform a cumbersome calculation to calculate the vector intensity by calculating the feature amount based on the ratio of the moving state pixel and the static state pixel obtained by addition and division. By using simple co-occurrence expressions, it is possible to reduce the amount of calculation compared to the conventional cascade booster with Adaboost, so the load on the CPU for calculating the characteristic amount can be reduced, and people, vehicles, etc. It is possible to realize the detection of the object at high speed and at low cost.

  In the above embodiment, the moving state is set when the first variance value S1 and the second variance value S2 are equal to or larger than the threshold value. However, the moving state may be set when the maximum value of the luminance change amount is larger than a predetermined luminance value. Alternatively, the moving state may be set as long as both the dispersion value and the maximum value of the amount of change in luminance are greater than or equal to a predetermined value.

  1. Image input unit (image frame group generation unit) 2. Time change extraction unit (luminance change calculation unit, variance value calculation unit, pixel state determination unit) 3. Inter frame feature extraction unit (feature amount calculation) Means), 4 .... detection unit (weak classifier generation means, object detection means), 5 .... result output unit.

Claims (1)

Image frame group generation means for generating an image frame group consisting of continuous images on the time axis based on captured images of the camera;
Changes in luminance of the same part for each pixel of an image frame up to a predetermined number on the time axis from the reference frame or a predetermined number on the time axis from an arbitrary image frame among a plurality of generated image frames as a reference frame A luminance change calculating means for calculating the maximum amount,
A variance value calculating means for calculating a variance value of luminance from the reference frame and luminance change information of the same part for each pixel of the image frame up to a predetermined number on the time axis from the reference frame or a predetermined number of images after the reference frame; ,
A pixel state in which the maximum value of the luminance change amount is larger than a predetermined value and / or a pixel in which the variance value of the luminance is larger than a predetermined value is determined as a moving state, and pixels in other states are determined as a static state Judgment means,
A feature amount that forms a cell composed of a plurality of adjacent pixels in the reference frame, calculates a motion feature amount and normalizes the ratio of the moving state pixel and the static state pixel by addition and division for each cell. A calculation means;
Weak classifier generating means for generating a weak classifier expressing co-occurrence from the feature quantity and the learning sample;
An object detection apparatus comprising: an object detection unit configured to detect an object moving by a strong classifier generated from the weak classifier by learning.

Images