KR101480072B1 - Vehicle pattern histogram generating method and venicle pattern detecting method using the same - Google Patents

Abstract

The present invention relates to a method of generating a vehicle pattern and a method of detecting a vehicle pattern using the method. In a method of generating a vehicle pattern in a cell of a predetermined size, A pair of horizontal patterns that are symmetrical patterns with respect to the horizontal center line, a pair of vertical patterns that are symmetrical patterns with respect to the vertical center line, a mutually symmetrical pattern with respect to the vertical center line, Wherein the upper inclined pattern and the lower inclined pattern are symmetrical with respect to the upper inclined pattern and the vertical center line of one pair or a plurality of pairs of the upper inclined pattern and the lower inclined pattern, .
According to the present invention, the performance of the vehicle pattern histogram is improved by applying the vehicle pattern histogram to the recognizer or classifier that has been developed, and the object similar to the vehicle object can be detected quickly.

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a vehicle pattern histogram generation method and a vehicle pattern detection method using the same, and more particularly to a VEHICLE PATTERN HISTOGRAM GENERATING METHOD AND VENICLE PATTERN DETECTING METHOD USING THE SAME,

The present invention relates to a method of generating a vehicle pattern histogram and a method of detecting a vehicle pattern using the same, and more particularly, to a method of detecting a vehicle pattern by using a binocular It is possible to derive data on the type and occurrence frequency of the pattern (model), which allows detection of the double vehicle pattern, and a pattern containing information on the outline pattern of the object of interest, that is, the vehicle body (or vehicle pattern) The present invention relates to a vehicle pattern histogram generation method and a vehicle pattern detection method using the same, and more particularly, to a vehicle pattern histogram generation method and a vehicle pattern histogram generation method in which the presence or absence of a vehicle object is roughly and quickly detected in an input image by comparing a histogram of a part or all of a histogram and a binary input image.

One of the criteria by which a person recognizes or distinguishes objects at night is the external model of the object. This means that the external information of the object can be used as a feature of automatic object recognition.

In order to recognize an object such as an object, a vehicle, or a person from an image input from a conventional closed circuit system (CCTV) or a digital camera, a feature of the object is first defined, extracted, and learned. Next, an object is automatically recognized by using a learning result in a detector or a classifier. In this automatic recognition system, how to extract a feature is very important factor, which has a great influence on the performance of the recognizer. However, unfortunately, there is no guideline for the feature that satisfies the commercialization level until now to be. Haar, HoG (Histogram of Oriented Gradient), and LBP (Local Binary Pattern) features, which are widely used as characteristics of the classifier, are not significantly different from those mentioned above. Therefore, in order to improve the performance and processing speed of the recognizer, it is necessary to study various methods of defining and extracting features. Especially, it is necessary to study to identify the object type even if it is a vehicle.

We also used bootstrap, boosting or adaboosting algorithms to detect or recognize objects of interest. In addition to these algorithms, we use Haar, HoG (Histogram of Oriented Gradient), or LBP (Local Binary Pattern) together. The boosting sequence algorithm combined with these features has a fatal disadvantage that the classification (detection) speed is very slow, not to mention the learning speed. Particularly, in the case of the global search, the classification speed is remarkably lowered and the recognition rate is very low. In order to compensate for these drawbacks, it is possible to compensate for the aforementioned disadvantages by detecting the existence and location of the object of interest, in particular, the vehicle object, in a rough and fast manner, and passing the result to the boosting or adaboosting recognizer.

As a conventional art of this field, there is a device and method for detecting a region of interest, and a recording medium (Korean Patent Registration No. 10-1272448, registered on May 31, 2013) in which a program implementing the above method is recorded. However, Detecting at least two vertices for determining a region of interest (ROI) using a predetermined reference point through a line drawing input, and re-detecting the remaining vertices for specifying a region of interest based on the detected vertices at a predicted and predicted point And an interest region detecting apparatus and method for detecting a region of interest including a target object.

However, this method also involves a complicated and time-consuming process of detecting a region of interest.

It is an object of the present invention to provide a mechanism for defining vehicle patterns appearing on an outline of an object to extract an outline feature of a vehicle object and calculating the occurrence frequency of predefined vehicle patterns in the binary edge input image, In which the source code complexity and the execution time are minimized in generating the vehicle pattern histogram.

It is an object of the present invention to provide a method of detecting a vehicle pattern using a vehicle pattern histogram that quickly and roughly detects an area having an outline characteristic similar to a vehicle object by utilizing the fact that objects of different kinds are different in outline information

According to another aspect of the present invention, there is provided a method for generating a pattern histogram in a sliding window area by moving a sliding window of a predetermined size in an input image,
Extracting a frequency of occurrences in the pattern histogram for a horizontal line pattern, a horizontal pattern, a vertical pattern, and upper and lower inclination patterns, which are pre-generated vehicle patterns;
Counting the number of occurrences of the plurality of extracted feature patterns exceeding respective thresholds,
And judging that the vehicle exists if the number of the counted patterns exceeds a predetermined value.

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

delete

Preferably,

And determining the thresholds of occurrence frequencies of the patterns with reference to the pattern occurrence frequency of the reference images.

delete

delete

Also preferably,

The step of counting the number of occurrences of the extracted feature patterns exceeding respective thresholds

 An edge is extracted by using an edge extraction algorithm,

The horizontal pattern, the vertical pattern, and the occurrence frequencies of the upper and lower inclination patterns, which are patterns that always occur among the extracted edge patterns, when the edge of the vehicle is clearly extracted .

As described above, according to the vehicle pattern histogram of the present invention and the vehicle pattern detection method using the same, the vehicle pattern histogram analyzing the outline of the vehicle object and calculating how much the vehicle pattern is distributed can be provided, The outline pattern analysis is facilitated.

According to the present invention, there is an effect that the type of the vehicle object can be inversely estimated through the data of the vehicle pattern histogram.

According to the present invention, the performance of the recognizer or classifier developed by applying the vehicle pattern histogram is improved.

According to the present invention, it is possible to quickly detect objects similar to vehicle objects.

According to the present invention, it is possible to detect a vehicle pattern faster than when a recognizer is used by reducing the processing time until recognition of the vehicle pattern.

1 is a diagram schematically illustrating an NxN cell space in a method of generating a pattern histogram according to the present invention.
FIGS. 2 and 3 are diagrams illustrating examples of 3x3 cell space to which a cell code is allocated in the pattern histogram generation method according to the present invention.
4 is a flowchart illustrating a process of an object detection method according to the present invention.
FIG. 5 shows a pattern histogram for a specific region of a binary edge image.
6 shows a pattern histogram of a template for a template image as a binary edge image of a reference image containing object information.
7 is a view showing an edge image processed with a vehicle image.
8 is a diagram showing a pattern histogram of a vehicle image edge.
9 to 17 are views showing an example of a vehicle pattern histogram according to the present invention.

Hereinafter, a vehicle pattern histogram generation method and a vehicle pattern detection method using the same according to the present invention will be described in detail with reference to the accompanying drawings.

In the present invention, a cell code is assigned in an N x N cell space in order to provide information on which object of interest is located at a certain position on the photographed image based on binary outline information according to a previously defined object, We propose a method of generating a pattern code that encodes an outline pattern of a target object and a method of generating a pattern histogram that accumulates the occurrence frequency of a pattern type of interest objects.

 At this time, the pattern histogram is a pattern distribution diagram showing how much the contour line pattern exists with respect to the analysis object.

The general histogram is a view of what value is in which image and how much. In contrast, the pattern histogram used in the present invention differs in finding out how much the pattern is in a specific region.

For example, if you zoom out of the input image by the size of the region of interest (assuming you are stripping out a 3x3 region) and zoom in, this becomes a pattern. In this way, all regions of interest are removed from the input image and enlarged and coded (as described in Equations 1 to 3 hereinafter). In other words, the pattern is coded to count frequencies (as described below by equations (1) and (2)). For example, if the number is 50, the number 50 is defined in advance as a certain pattern (the control unit (not shown) calculates a predefined number through a memory (not shown)). The control unit and the memory are general contents, and therefore, the drawings are not shown.

A similar shape will have a similar pattern, so you will later decide on a pattern with one value.

In the present invention, in order to encode a pattern, the number of cases that can be extracted from 3x3 is divided into 2 ^ 8 (8th power of 2), and the middle cell is excluded. The number of cases that can come out from 5x5 can be 2 ^ 24 (2 24 times). For reference, cells must be odd to operate efficiently.

However, the cell size can not be increased continuously, and the number of cases that can be increased is increased, but the accuracy is not guaranteed, and the speed of the pattern histogram is decreased because the capacity of the pattern histogram is increased.

According to the present invention, there is provided an algorithm for creating a code in a specific space. In addition, if there are only 2 actual patterns in a 5x5 cell area, if the number of required patterns is only 2, write 0, 1, and then just map to another code. This is called mapping-mapping. That is, the code that is not used increases the histogram, so the intention is to use only the necessary code.

On the other hand, cell code assignment can be changed and changed.

Each of the cells of Figs. 1 and 2 is a cell code (one square is called a cell). It is not known what pattern comes in the actual pattern (input image pattern), but when it is combined, some number comes out. The number was called a pattern code.

To determine the size of the pattern histogram, all of the cell codes assigned to the cells in the cell space are added as shown in Equation (2), and finally 1 is added. For example, when a cell code is defined in a 3x3 cell space, the total sum of cell codes is 255, so that the size of the pattern histogram is 256. And the pattern code for indexing the pattern histogram is expressed by Equation (3).

As described above, an important feature of the present invention is to provide object information that can be used to detect an object of interest in an input image (this part has already been described and may be duplicated if necessary).

Referring to FIGS. 1 and 2, a method of generating a pattern histogram previously filed by the applicant of the present invention will be described.

First, the cell space 100 will be described.

A cell space (or a cell table) is a set of cells that store a cell code 110 and are classified according to a location, and is shown in FIG. 1, where N is an integer and an odd number.

The cell code will be described.

A cell code is given to each cell of the cell space in order to express one pattern as a code, and has the following constraint.

- Assign cell codes to cells in any order regardless of cell location. Cells and codes are one-to-one correspondence.

- The cell code should not overlap with other cell codes.

- The cell code must be present in binary if it is not '1'.

- No code is assigned to the center cell 120 in the center position or zero is given.

- Code 0 can not be used other than the center cell.

For example, if no code is allocated to the center cell in the NxN cell space, the length of the binary cell code is (NxN) -1. An example of a cell code assigned to a 3x3 cell space is shown in FIG.

The mapping of the cell code will be described.

The length of the cell code is determined by the size of the cell space. As a result, even if the number of patterns of interest is small, a large cell space causes inefficiency which leads to a long cell code length. This is because the length of the cell code is a factor that determines the size of the bin of the pattern histogram to be described later.

Therefore, when the number of patterns of interest is small, the cell code is mapped to another short code to reduce the cell code length. One-to-one mapping of only necessary cell codes to other short codes using a method such as a hash function or a lookup table. However, the mapping of cell codes can be omitted as needed.

The pattern code using the cell code will be described.

A short channel binary (0 or 1) pattern code is formed by summing the product of the partial area of the input image and the position of the cell space. The equation for generating the pattern code is shown in Equation 1 below.

In the above Equation 1, pcode (x, y) is a pattern code at position (x, y), N is the size of cell space, N / 2 is an integer operation, and decimals are discarded unconditionally. cell_space (i, j) is a cell space holding a predefined cell code, and the center cell code is 0. And region (x, y) is a partial region of the binary edge image, and the pattern code is calculated only when the center is 1. And * is a multiplication operator.

The pattern histogram by the pattern code will be described.

The number of beans total_bins, which means the size of the pattern histogram, is determined by Equation (2).

In Equation (2), cell_space () is a cell space, and 1 is added to the pattern histogram, and 1 can be removed as needed. The method of accumulating the occurrence frequency of the patterns for the binary input image is determined by Equation (3).

In the equation (3), phist [] is a pattern histogram, pcode (x, y) is a pattern code determined by the equation (1), and the symbol <--- is an assignment operator. The empty number of the pattern histogram is total_bins shown in Equation (2).

In the case of the conventional global search, the classification speed is significantly lowered and the recognition rate is very low. In order to secure such a disadvantage, it is possible to secure the above-mentioned disadvantages by detecting the existence and position of the object of interest in a rough and fast speed and forwarding the result to the boosting or adaboosting recognizer.

From this, the present invention has been started. Therefore, the object detection method using the pattern histogram can be used as a preprocessing step in the existing boosting algorithm, but it can also be used independently.

Referring to FIGS. 4 to 6, the object of the present invention can be classified into an object having an edge characteristic similar to the object of interest in the input image, based on the outline information of the object of interest analyzed in advance using the point having different outline information The area is detected quickly and roughly.

To do this, we analyze in advance the edge features of the template given as objects of interest. Based on this, it is judged whether the object of interest exists or not by comparing the edge feature of the given edge image with the input edge.

4 is a flowchart illustrating a process of an object detection method according to the present invention.

The processing method will be explained.

The process area setting is performed first (S2).

The entire single-channel binary (0 and 1) edge image is divided into several sub-areas. The processing method selects an area to be processed, and then processes a series of processes for the area. It then sets up another area and processes a series of processes. The processing area is set by a rectangle composed of a starting position called a window, a horizontal size, and a vertical size, and the method is as follows.

How to set the zone 1

Step 1. Set the window by increasing the initially fixed size window from the initial starting point by a certain amount in the positive horizontal direction.

Step 2. When the setting for one line is completed, the area is set in the horizontal line as in Step 1 at the changed vertical position by moving a certain amount in the positive vertical direction.

Step 3. Perform Step 2 for the entire image.

Step 4. Change the size of the window by a specific amount in the horizontal and / or vertical direction.

Step 5. Repeat steps 1-4 until the given window size is satisfied with the termination condition.

How to set the area 2

Step 1. Set the initial area by setting the size and position of the initial window.

Step 2. Set the area while keeping the position of the window fixed while adjusting the window size. Continue until the window size is equal to a predefined size.

Step 3. Move the position of the window by a certain amount in the positive horizontal direction and repeat step 2. However, if the position of the window is at a specific position on the line (end of line), proceed to step 4.

Step 4. Move the position of the window by a certain amount in the positive vertical direction and repeat steps 2-3.

Step 5. When the end condition reaches the vertical position of the given window, end the area setting.

Pattern histogram generation for the processing area (S4)

To find the pattern type and frequency of the selected region, we generate pattern histogram by calculating pattern code in NxN cell space. The description of the pattern histogram is a method of generating a pattern histogram, which is the contents filed by the applicant of the present invention.

FIG. 5 shows a pattern histogram for a specific region of a binary edge image, which is a pattern histogram for the red square region on the right side expanded in the left binary edge image.

Pattern histogram of templates

A template is a binary edge image of a reference image containing object information, and an example thereof is shown on the left side (a) of FIG. The pattern histogram of the template must be analyzed in advance, and the method of generating the pattern histogram has already been described above (see Equations 1 to 3), so that a detailed description thereof will be omitted.

We also need to analyze a number of templates for the objects of interest and analyze in advance which indices (pattern codes) in the pattern histogram and how to use them as features.

Feature comparison (S6)

The pattern histogram of the template is compared with the histogram of the processing region pattern to determine whether it is the object of interest. The number of pattern codes satisfying the expression (4) is counted.

In Equation 4, R [] is the pattern histogram of the processing region, T [] is the pattern histogram of the template, pi is the i-th index of the pattern histogram, and F is a set of patterns selected as the comparison pattern. That is, it is an index set of a pattern histogram to be compared. Si is a scale factor, X is a multiplication operator, > is a large / small operator, and a is a factor representing a window size ratio determined by equation (5).

After step S6, it is determined whether an object of interest is included (S8)

If the number of the feature pattern codes satisfying Equation (4) is greater than a given threshold value, it is determined that the region includes the object of interest or is not included in the region.

Then, it is determined whether the entire area has been processed (S10).

According to the object detection method using the pattern histogram according to the present invention, objects similar to the object of interest can be detected quickly.

According to the present invention, by reducing the processing time for recognizing an object of interest, it is possible to detect it faster than the conventional method using the recognizer.

5 and 6 illustrate vehicle objects as an example.

7 is an edge image obtained by processing a vehicle image in a Sobel fashion. In other words, the edge is extracted using the Sobel edge extraction algorithm.

In the present invention, in particular, a vehicle pattern for a vehicle object has been created in total of nine types as shown in Figs. 9 to 17 in advance.

That is, the vehicle pattern generated by the present invention as a pattern useful for detecting the vehicle will be described. First, edges of the input image are extracted using a Sobel edge extraction algorithm (see FIG. 7).

When the edges of the vehicle are clearly extracted, the patterns shown in Figs. 9 to 17 always occur. That is, the horizontal line pattern, the upper and lower horizontal patterns, the left and right vertical patterns, the left and right upper inclination patterns, and the left and right lower inclination patterns. The frequency of occurrence of such feature patterns is extracted using the above-described pattern histogram algorithm, and then it is determined whether or not a vehicle exists in the corresponding region. Below is a case of 3x3 of each pattern.

The horizontal line pattern (Figure 9)

refers to a pattern that horizontally passes the center of a cell to a thickness of 1 pixel. 9 shows an example in which the size of the cell is 3x3. The pattern code is 34 and the shape of the pattern is as shown in FIG. In Fig. 9, the bright part represents the edge and the black part represents the background.

The upper and lower horizontal patterns (Figs. 10 and 11)

Refers to a pattern when the edge is oriented in the horizontal direction. Figs. 10 and 11 show an example in which the cell size is 3x3. The pattern codes are 62 (FIG. 10) and 227 (FIG. 11), and the patterns are shown in FIGS. 10 and 11. In the drawing, a bright part represents an edge and a black part represents a background.

Left vertical and right vertical patterns

Refers to a pattern when the edge is directed in the vertical direction. Figs. 12 and 13 show an example in which the cell size is 3x3. The pattern codes are 143 (FIG. 12) and 248 (FIG. 13), and the patterns are shown in FIGS. 12 and 13. In the drawing, a bright part represents an edge and a black part represents a background.

Left upper and right upper inclined patterns

Quot; refers to a pattern when the top of the edge is inclined. Figs. 14 and 15 show an example in which the cell size is 3x3. Pattern codes are 30 (FIG. 14) and 60 (FIG. 15), and the patterns are shown in FIGS. 14 and 15 below. In the drawing, a bright part represents an edge and a black part represents a background.

Left lower and right lower inclined patterns

Refers to a pattern when the lower edge of the edge is inclined. FIGS. 16 and 17 show an example in which the cell size is 3x3. Pattern codes are 195 (Fig. 16) and 225 (Fig. 17), and the patterns are shown in Figs. 16 and 17 below. In the drawing, a white part represents an edge (a bright part) and a black part (a brightness part represents a background).

A method of detecting a vehicle pattern will be described.

The vehicle is detected using the correlation between the occurrence frequency of the feature patterns and the occurrence frequency of the feature patterns.

The method of detecting the vehicle pattern can be proceeded as follows.

ST-1. The pattern histogram is generated in the sliding window area while moving the sliding window of a certain size in the input image.

ST-2. The occurrence frequency in the pattern histogram of the horizontal line pattern, the horizontal pattern, the vertical pattern, and the upper and lower inclination patterns, which are the previously-generated vehicle patterns, is extracted.

ST-3. Thresholds of occurrence frequency of each pattern (arbitrarily determined) are determined with reference to pattern occurrence frequency of reference images.

ST-4. And counts the number of occurrences of the extracted feature patterns exceeding respective thresholds.

ST-5. If the number of counted patterns exceeds a predetermined value, it is determined that the vehicle exists.

According to the vehicle pattern histogram and the vehicle pattern detection method using the same according to the present invention, it is possible to easily analyze the outline pattern of the vehicle object by providing a vehicle pattern histogram that can calculate how much the vehicle pattern is distributed by analyzing the outline of the vehicle object in advance .

According to the present invention, the type of the vehicle object can be inversely estimated through the data of the vehicle pattern histogram.

According to the present invention, its performance can be improved by applying a vehicle pattern histogram to a recognizer or classifier that has already been developed.

According to the present invention, objects similar to vehicle objects can be detected quickly.

According to the present invention, by reducing the processing time until recognition of the vehicle pattern, it is possible to detect it faster than in the case of using the recognizer.

100: 3x3 cell space
110: decimal cell code
120: center cell

Claims (6)

delete delete Generating a pattern histogram in a sliding window area while moving a sliding window of a predetermined size in an input image,
Extracting a frequency of occurrences in the pattern histogram for a horizontal line pattern, a horizontal pattern, a vertical pattern, and upper and lower inclination patterns, which are pre-generated vehicle patterns;
Counting the number of occurrences of the plurality of extracted feature patterns exceeding respective thresholds,
And judging that the vehicle exists if the number of counted patterns exceeds a predetermined value.
The method of claim 3,
Further comprising the step of determining thresholds of occurrence frequencies of the patterns with reference to the pattern occurrence frequency of the reference images.
delete The method of claim 3,
The step of counting the number of occurrences of the extracted feature patterns exceeding respective thresholds
An edge is extracted by using an edge extraction algorithm,
A horizontal pattern, a vertical pattern, and an occurrence frequency of the upper and lower inclination patterns, which are patterns that always occur among the extracted edge patterns when the edge of the vehicle is clearly extracted A method for detecting a vehicle pattern.

Images