KR20090032808A - Object classifier method and system for controlling the same - Google Patents

Abstract

An object classifying method and a system thereof are provided to prevent an inputted image from being influenced by the rotation of a subject regardless of the difference in the size of the image, corresponding to the distance between a camera and the subject, and the inclined degree of the subject corresponding to the photographing angle. A central axis extracting unit(110) extracts the central axis of an inclined image. A central axis conversion unit(120) converts the central axis which the central axis extracting unit extracts so as to coincide with the y-axis. An image vector extraction unit(130) extracts a feature vector by dividing the converted image in the lattice type based on the HOG(Histogram of Oriented Gradients) algorithm. An image determination unit(140) determines whether the image is an image of a person or vehicle by inputting the extracted feature vector into an LSVM(Linear Support Vector Machine).

Description

OBJECT CLASSIFIER METHOD AND SYSTEM FOR CONTROLLING THE SAME

The present invention relates to a method and a system for classifying an object, and more particularly, to a method and a system for classifying an object.

Recently, as terrorism and theft accidents occur frequently, interest in security is increasing day by day. As a result, CCD cameras (charge-coupled device cameras) are being installed not only in urban skyscrapers, but also in major living environments such as apartment complexes.

For security and crime prevention by such a CCD security camera, it is necessary to capture a moving object, and to classify what the object is to the captured object.

However, in the object classification system used in the conventional CCD security camera, the difference in the size of the captured image occurs depending on the distance between the camera and the subject, and the size of the subject depends on the shooting angle. Since the degree of inclination is different, there is a problem that the object classification rate of the subject or the photographed image by the object classification system is significantly reduced.

The present invention has been made to improve the prior art as described above, in the object classification system used in the CCD camera, the input image is the size difference of the image according to the distance between the camera and the subject of the subject according to the shooting angle It is an object of the present invention to provide an object classification system that is robust to rotation of a subject regardless of the degree of inclination.

In addition, the present invention can automatically classify people and vehicles in a security environment to provide a high object classification rate when photographing a subject such as a CCD camera, and is stable even when classifying objects in an image distorted by environmental noise. An object of the present invention is to provide a method and system for classifying objects.

In order to achieve the above object and solve the problems of the prior art, the object classification method according to one aspect of the present invention, extracting the tilted central axis of the image, converting the central axis of the tilted image to match the y-axis, Dividing the transformed image into a grid to extract a feature vector, and determining whether the image is a person or a vehicle by using the extracted feature vector.

According to another aspect of the invention, the step of extracting the tilted central axis of the image, extracts the tilted central axis of the image using a Radon transform (Radon Transform).

According to another aspect of the present invention, the step of extracting a feature vector by dividing the transformed image into a grid, the feature vector is extracted using a histogram of oriented gradients (HOG) algorithm.

According to another aspect of the present invention, the HOG algorithm extracts a feature vector of the converted image using a cell consisting of a plurality of pixels and a block consisting of the plurality of cells.

According to another aspect of the invention, the step of determining whether the person or the vehicle, the feature vector is input to the LSVM (Linear Support Vector Machine) to determine whether or not the person or vehicle.

According to yet another aspect of the present invention, in the determining of whether the person or the vehicle is input, when the output information is positive by inputting the feature vector to the person classification LSVM, the person is determined to be a person.

According to another aspect of the present invention, the step of determining whether the person or the vehicle, if the output information is positive by inputting the feature vector to the vehicle classification LSVM, it is determined as a vehicle.

According to another aspect of the invention, the step of extracting a feature vector by dividing the transformed image into a grid, obtaining a gradient of the x-axis and y-axis of the image, the x-axis of the image Obtaining a magnitude and orientation value of each pixel of the image from the longitudes of the pixels and the y-axis, and normalizing the overlapping cells of the plurality of pixels. Generating a block feature vector for a block composed of the plurality of cells based on the magnitude value and the direction value, and extracting the feature vector using the block feature vector.

According to another aspect of the invention, the step of obtaining the gradient of the x-axis and the y-axis of the image, using the [-1 0 1] mask to obtain the gradient of the x-axis and the y-axis of the image. .

According to the present invention, in the object classification system used in the CCD camera, the input image is rotated of the subject regardless of the difference in the size of the image according to the distance between the camera and the subject and the degree of inclination of the subject according to the shooting angle ( It is possible to provide an object classification system that is robust against rotation.

In addition, according to the present invention, it is possible to automatically classify people and vehicles in the security environment to provide a high classification rate of the object when shooting the subject of the CCD camera, stable even when classifying objects of the image distorted by environmental noise An object classification method and system can be provided.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings and the contents described in the accompanying drawings, but the present invention is not limited or limited to the embodiments. Like reference numerals in the drawings denote like elements.

1 is a view showing an object classification system according to the present invention, Figure 2 is a view showing an image vector extraction unit according to the present invention.

As shown in FIG. 1, the object classification system according to the present invention includes a central axis extractor 110, a central axis converter 120, an image vector extractor 130, and an image determiner 140. It is composed.

The central axis extractor 110 extracts the tilted central axis of the image.

The central axis converter 120 converts the central axis of the inclined image extracted by the central axis extractor 110 to match the y axis. In this case, according to the present invention, the inclined central axis of the image may be extracted using a Radon transform.

The image vector extractor 130 extracts a feature vector by dividing the converted image into a grid, and may extract the feature vector using a histogram of oriented gradients (HOG) algorithm.

The HOG algorithm refers to an algorithm that divides an image into a checkered square shape and calculates a local histogram of orientation in each of the square grids, in particular, a cell consisting of a plurality of pixels and the plurality of cells. Using a block, a feature vector of the transformed image is extracted.

In more detail, as illustrated in FIG. 2, the image vector extracting unit 130 may include a hardness calculator 131, a pixel value extractor 132, a normalizer 133, and a feature vector generator ( 134, and a feature vector extractor 135.

The hardness calculator 131 obtains gradients of the x-axis and y-axis of the image, and the pixel value extractor 132 determines the hardness of the image from the x-axis and y-axis hardness of the image calculated by the hardness calculator. Obtain the Magnitude value and Orientation value of each pixel, and the normalizer 133 normalizes the cells composed of the plurality of pixels by overlapping and normalizing the feature vector generator. The block 134 generates a block feature vector for a block composed of the plurality of cells based on the magnitude value and the direction value, and the feature vector extractor 135 uses the block feature vector to generate a feature vector. Extract.

In particular, the hardness calculator 131 obtains gradients of the x-axis and the y-axis of the image using a mask [-1 0 1].

The image determination unit 140 determines whether the image is a person or a vehicle by using the extracted feature vector. In this case, the image determination unit 140 may determine whether the vehicle is a person or a vehicle by inputting the feature vector to a linear support vector machine (LSVM).

3 is an image diagram according to an example of the present invention.

In the present invention, a Radon transform is applied to an image as shown in FIG. 3.

Radon Transform is an algorithm for finding lines similar to the Hough Transform. In the present invention, the radon transform is used to transform the tilted image (subject) axis to match the Y-axis. Can be used to

4 is a diagram illustrating a process of converting a central axis of an image to match the y-axis by using an edge component of the image according to an example of the present invention.

Radon transform is used to detect how inclined by the edge information of the inclined image 410, and converts by that angle so that the central axis coincides with the Y-axis, thereby tilting the image. Do not be affected by (Rotation).

The radon transformation as described above is performed by Equation 1 below.

(x is the hardness value in the x-axis direction, and y is the hardness value in the y-axis direction.)

The expression in the delta function is a Hough transform expression and returns '1' for coordinates with edge components, otherwise it returns '0'.

5 and 6 are graphs of values of coordinates with edge components for any angle of an image according to an example of the present invention, and more specifically, FIGS. 5 and 6 are shown in FIG. The image is a graph of the sum of the coordinates with the edge components for any arbitrary angle, from zero degrees to 180 degrees.

As shown in FIG. 5, it can be seen that the maximum values are set at '45 degrees' and '135 degrees'.

(x is the hardness value in the x-axis direction, and y is the hardness value in the y-axis direction.)

In order to find the strongest edge component, the second derivative value shown in FIG. 6 is obtained by using Equation 2 as described above, and the angle corresponding to the minimum value is the strongest angle on the graph.

According to the present invention, a straight line component in a vertical direction is strongly shown by a person, and a straight line component in a horizontal direction is strongly shown by a vehicle. By doing this, it is possible to solve the constraint on the image in which the central axis is tilted.

7 illustrates a cell and a block according to an example of the present invention.

Histogram of Oriented Gradients (HOG) is an algorithm that divides an image into a checkerboard square shape and calculates a local histogram of orientation in each grid.

According to the HOG algorithm, gradients in the x and y-axis directions are obtained from the input image by using a [-1 0 1] mask, respectively.

(x is the hardness value in the x-axis direction, and y is the hardness value in the y-axis direction.)

(x is the hardness value in the x-axis direction, and y is the hardness value in the y-axis direction.)

From the gradient values in the x- and y-axis directions, the Magnitude value of each pixel is obtained using Equation 3 as described above, and the Orientation value of each pixel is obtained using Equation 4 as described above. You can get it.

Thereafter, the cells are normalized in units of blocks while overlapping cells.

Assuming that the size of the image (input image) is 64 X 128, the image is divided into square tiles. In this case, the size of one board is called 8 X 8 pixels, and the total number of cells is You can see that 8 X 16 = 128.

As illustrated in FIG. 7, four cells 710 are configured in a square shape to form a block 720 in order to local normalize.

In this case, the size of the block is 16 X 16 pixels, and a normalization factor can be obtained in units of blocks by using Equation 5 below.

(nf is a normalization coefficient, and V is a vector representing the Magnitude value of each pixel.)

When the normalization coefficient is obtained and divided by the value of each pixel, one local normalization is completed, and the same process is repeated by moving each cell to the right by using the cell 710 of FIG. Local Normalize (No. 7 in Landscape, No. 15 in Portrait)

At this time, a situation in which a cell overlaps with a block between blocks occurs. Even though the cells overlap, each of the cells has a different normalization scheme, so that the local information can be stronger. There is.

8 illustrates a normalization scheme according to an example of the present invention.

In the present invention, the block normalization scheme (Normalize Scheme) is based on the fact that the linear component, which is a common feature of humans and vehicles, is strong, and not the square form of the blocks 16 X 16 and the cells 8 X 8, FIG. 8. A normalization scheme using a 16 × 24 rectangular block and an 8 × 16 cell 810 as shown in FIG.

The rectangular type method according to the present invention has an advantage of having more information on linear components than the square type. In other words, when compared to the square block and the cell shape, the rectangular block and the cell structure not only have overlap between the blocks and the cells in the block, but also have overlaps between the cells and the cells. Having more image information has the advantage of extracting a feature that is more robust to the scale of the object classification.

9 is a table for summing the magnitudes of pixels corresponding to a range to which an orientation of each pixel belongs according to an example of the present invention.

By using the summation table illustrated in FIG. 9, a range of directions (Orientation) 910 according to the index of each pixel is found, and the magnitude of the pixel is added. The process is performed cell-by-cell, and if the values for all the pixels in a cell are passed to the bin, then the cell is voted on another independent bin, for a total of four bins in the block. The memory is formed.

For example, if the orientation information of an arbitrary position is 46.5 degrees and the corresponding Magnitude is 'k', the orientation value is an empty index corresponding to 40 to 60 degrees ( Bin index) The weighted vote value corresponds to '3' and becomes a 'c + k' value obtained by adding a 'k' value to a 'c' value, which is a conventional voted value (920). This process has a feature that if the edge information of a particular angle is large, it is weighted by its magnitude.

FIG. 10 is a view of connecting four bin memories of each block according to an example of the present invention.

When a total of four bin memories are completed for each block, as shown in FIG. 10, one feature vector for a block is completed by concatenation.

Once the voting for 105 blocks is completed, one large feature vector (3780) is completed.

A concatenate feature vector has a total of 36 values and contains information about one block. Since there are 105 such feature vectors, the dimension of each input image 64 X 128 is 3780.

The completed feature vector is input to a human classification linear support vector machine (LSVM) to classify whether the image is a person or a person. If the output information is positive, it is determined to be a person and vice versa.

In addition, the same feature vector is input to the vehicle classification Linear Support Vector Machine, and if it is positive, the vehicle is discriminated as a vehicle.

11 is a diagram illustrating an object classification method according to an example of the present invention.

The tilted central axis of the image is extracted (S110), and the central axis of the tilted image is converted to match the y-axis (S120).

In this case, the inclined central axis of the image is extracted using the Radon transform.

Thereafter, the transformed image is divided into a grid to extract a feature vector (S130). At this time, the feature vector is extracted using a histogram of oriented gradients (HOG) algorithm.

In particular, the HOG algorithm extracts a feature vector of the converted image using a cell composed of a plurality of pixels and a block composed of the plurality of cells.

Thereafter, the extracted feature vector is used to determine whether the image is a person or a vehicle (S140). At this time, the feature vector is input to a linear support vector machine (LSVM) to determine whether the image is a person or a vehicle. Done.

In order to determine whether the vehicle is a person or a vehicle, the feature vector is input to a person classification LSVM to determine whether the image is a person image (S150), and the image is a person image (S160). When the output information is positive, it is determined as a person.

If it is determined that the image is not a human image, the feature vector is input to the vehicle classification LSVM to determine whether the image is a vehicle image (S170). At this time, the vehicle classification LSVM determines the vehicle when the output information is positive (S180), and otherwise determines that the image is not an image of a person and a vehicle (S190).

FIG. 12 is a diagram illustrating a method of extracting a feature vector according to an example of the present invention. In detail, the center axis of the tilted image is changed to match the y-axis, and the transformed image is HOG ( This section describes how to extract feature vectors by dividing them into grids using the Histogram of Oriented Gradients algorithm.

First, gradients of the x-axis and y-axis of the image are obtained (S210), and the hardness may be obtained using a [−1 0 1] mask.

Magnitude values and orientation values of each pixel of the image are obtained from the longitudes of the x-axis and y-axis of the image (S220), and the cells composed of the plurality of pixels are overlapped. To normalize (S230).

Based on the magnitude value and the direction value, a block feature vector for a block including the plurality of cells is generated (S240), and a feature vector is extracted using the block feature vector (S250).

Embodiments of the invention also include computer-readable media containing program instructions for performing various computer-implemented operations. The computer readable medium may include program instructions, data files, data structures, etc. alone or in combination. The program command may be those specially designed and configured for the present invention, or may be known and available to those skilled in the art. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tape, optical media such as CD-ROMs, DVDs, and magnetic disks, such as floppy disks. Magneto-optical media, and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, flash memory, and the like. Examples of program instructions include not only machine code generated by a compiler, but also high-level language code that can be executed by a computer using an interpreter or the like. When all or part of the operation of the mobile terminal or base station described in the present invention is implemented as a computer program, a computer readable recording medium storing the computer program is also included in the present invention.

As described above, the present invention has been described by way of limited embodiments and drawings, but the present invention is not limited to the above embodiments, and those skilled in the art to which the present invention pertains various modifications and variations from such descriptions. This is possible.

Therefore, the scope of the present invention should not be limited to the described embodiments, but should be determined not only by the claims below but also by the equivalents of the claims.

1 is a view showing an object classification system according to the present invention.

2 is a diagram illustrating an image vector extracting unit according to the present invention.

3 is an image diagram according to an example of the present invention.

4 is a diagram illustrating a process of converting a central axis of an image to match the y-axis by using an edge component of the image according to an example of the present invention.

5 and 6 are graphs of values of coordinates with edge components for any angle of an image according to an example of the present invention.

7 illustrates a cell and a block according to an example of the present invention.

8 illustrates a normalization scheme according to an example of the present invention.

FIG. 9 is a table summating magnitudes of pixels corresponding to a range to which an orientation of each pixel belongs according to an example of the present disclosure.

FIG. 10 is a view of connecting four bin memories of each block according to an example of the present invention.

11 is a diagram illustrating an object classification method according to an example of the present invention.

12 is a diagram illustrating a method of extracting a feature vector according to an example of the present invention.

<Explanation of symbols for the main parts of the drawings>

100: object classification system 110: central axis extraction unit

120: central axis conversion unit 130: image vector extraction unit

140: image determination unit

Claims (21)

Extracting the tilted central axis of the image; Converting a central axis of the tilted image to coincide with a y axis; Dividing the converted image into a grid to extract a feature vector; And Determining whether the image is a person or a vehicle by using the extracted feature vector Object classification method comprising a. The method of claim 1, Extracting the tilted central axis of the image, An object classification method comprising extracting a tilted central axis of the image using a Radon transform. The method of claim 1, Extracting the feature vector by dividing the transformed image into a grid, And extracting the feature vector using a histogram of oriented gradients (HOG) algorithm. The method of claim 3, HOG algorithm, And a feature vector of the transformed image is extracted using a cell composed of a plurality of pixels and a block composed of the plurality of cells. The method of claim 1, Determining whether the person or the vehicle is, And classifying the feature vector into a linear support vector machine (LSVM) to determine whether the feature vector is a human or a vehicle. The method of claim 5, Determining whether the person or the vehicle is, And inputting the feature vector into a human classification LSVM to determine that the output information is positive. The method of claim 5, Determining whether the person or the vehicle is, And classifying the feature vector into a vehicle classification LSVM and determining that the output information is positive. The method of claim 1, Extracting the feature vector by dividing the transformed image into a grid, Obtaining gradients of the x and y axes of the image; Obtaining Magnitude values and Orientation values of each pixel of the image from the longitudes of the x and y axes of the image; Normalizing the cells including the plurality of pixels by overlapping the cells; Generating a block feature vector for a block consisting of the plurality of cells based on the magnitude value and the direction value; And Extracting a feature vector using the block feature vector Object classification method comprising a. The method of claim 8, Obtaining the gradient of the x-axis and y-axis of the image,  [-1 0 1] A method of classifying an object according to claim 1, wherein the gradient of the x and y axes of the image is obtained. The method of claim 8, The magnitude is calculated using the following equation.

(x is the hardness value in the x-axis direction, and y is the hardness value in the y-axis direction.) The method of claim 8, And the direction value is calculated using the following equation.

(x is the hardness value in the x-axis direction, and y is the hardness value in the y-axis direction.) A computer-readable recording medium having recorded thereon a program for executing the method of claim 1. A central axis extracting unit extracting a tilted central axis of the image; A center axis converting unit converting the center axis of the tilted image to match the y axis; An image vector extracting unit dividing the converted image into a grid to extract a feature vector; And An image determination unit that determines whether the image is a person or a vehicle by using the extracted feature vector Object classification system comprising a. The method of claim 13, The central axis extraction unit, An object classification system, characterized by extracting the tilted central axis of the image using a Radon transform. The method of claim 13, The image vector extraction unit, And extracting the feature vector using a histogram of oriented gradients (HOG) algorithm. The method of claim 15, The HOG algorithm, And a feature vector of the transformed image is extracted using a cell composed of a plurality of pixels and a block composed of the plurality of cells. The method of claim 15, The image determining unit, And inputting the feature vector into a linear support vector machine (LSVM) to determine whether a feature is a person or a vehicle. The method of claim 15, The image vector extraction unit, A hardness calculator for calculating gradients of the x-axis and y-axis of the image; A pixel value extracting unit that obtains Magnitude values and Orientation values of each pixel of the image from the hardness of the x and y axes of the image calculated by the hardness calculator; A normalizer for overlapping and normalizing a cell composed of the plurality of pixels; A feature vector generator configured to generate a block feature vector for a block including the plurality of cells based on the magnitude value and the direction value; And Feature vector extractor for extracting feature vectors using the block feature vectors Object classification system comprising a. The method of claim 18, The hardness calculation unit,  And a gradient of the x-axis and y-axis of the image using a [-1 0 1] mask. The method of claim 18, The pixel value extraction unit, The object classification system, characterized in that to calculate the magnitude using the following equation.

(x is the hardness value in the x-axis direction, and y is the hardness value in the y-axis direction.) The method of claim 18, The pixel value extraction unit, The object classification system, characterized in that for calculating the direction value using the following equation.

(x is the hardness value in the x-axis direction, and y is the hardness value in the y-axis direction.)

Images