KR101582983B1 - Apparatus and method for fast object detection using scale invariant feature vector - Google Patents

Abstract

The present invention relates to an apparatus and a method for detecting an object at a high speed using a feature vector which is invariant according to an image size.
A fast object detection apparatus using a size invariant feature vector according to an aspect of the present invention calculates a share of a pixel having a shared region and an image patch cell among pixels of the image acquired by the image input unit, A feature vector extracting unit for extracting an invariant slope histogram feature vector according to the size conversion of the image patch cell using the result of the calculation and a gradient histogram feature vector for calculating the histogram of the orientated gradient, And an object detecting unit for detecting the object.

Description

[0001] APPARATUS AND METHOD FOR FAST OBJECT DETECTION USING SCALE INVARIANT FEATURE VECTOR [0002]

The present invention relates to an apparatus and a method for detecting an object at a high speed using a feature vector which is invariant according to an image size.

Techniques for detecting various objects such as vehicles, pedestrians, and faces have been applied to fields of intelligent automobiles, robots, surveillance cameras, games, and the like in the field of computer vision.

Conventional techniques for detecting a specific object in an image determine the position and size of the object by searching the entire image using a previously learned model and selecting an area corresponding to the object in the entire image.

As shown in FIG. 1, in order to detect an object of interest, a method of combining a histogram of an Oriented Gradient (HOG) and a support vector machine (SVM) ) Feature vectors are extracted, and a classification model is learned using extracted feature vectors and linear support vector machine classifiers.

For example, according to the conventional art, a slope histogram feature vector for 100 human images and 100 background images (non-human objects or backgrounds) is extracted and expressed in a vector space, and a plane equation (Optimal Hyper Plane) The object can be detected by distinguishing the clusters of the objects based on the object.

The object detection technique using the slope histogram and the support vector machine according to the related art improves the object detection performance as compared with the method using the Haar-like feature and the Adaboost feature, There is a problem that the computation time is rapidly increased.

Since the slope histogram (HOG) feature vector does not have the property of scale invariant, the object detection technique using the slope histogram according to the related art adjusts the size of the experiment image as shown in FIG. 2, The pyramid image is created and the object search is performed using the image patches of a fixed size (the same size as the size of the learned image samples) for each image, thereby increasing the calculation time.

3, for example, a sample image (eg, a learning image having a resolution of 48 × 96) is divided into 6 × 12 cells, The angle and size of the tilt are calculated for each cell (8 x 8 pixels). Further, the angle of the tilt is divided by a constant interval (bin, the bar graph width or interval in the x-axis in the bar graph shown in Fig. 3), and the magnitude of the gradient obtained from each pixel is added to each bin.

Each cell is combined with neighboring 2x2 cells to form a block, and the gradient histogram (HOG) is normalized in the block. The size nF of the entire feature vector dimension for the sample image is determined by the product of the number of blocks nB, the number of bins nbin, and the number of cells nC per block, as shown in the following equation .

[Equation 1]

According to recent research results, even if the image size is changed, the magnitude of the slope is proportional to a certain constant.

Slope histogram (HOG) When extracting feature vectors, normalization is performed for all bins, so theoretically, slope information can be used regardless of the size of the experiment image.

However, when it is desired to detect objects of various sizes in an image, when the number of pixels in a cell does not become an integer number (for example, when a feature vector is extracted for a 44 × 88 experimental image, 7.5 pixels), there is a problem that the gradient histogram (HOG) feature vector can not be accurately extracted.

SUMMARY OF THE INVENTION The present invention has been made to solve the above problems of the prior art, and it is an object of the present invention to extract a slope histogram feature vector having an invariant characteristic even for image patches of various scales, And to provide a high-speed object detection apparatus and method capable of performing high-speed object detection.

A fast object detection apparatus using a size invariant feature vector according to an aspect of the present invention calculates a share of a pixel having a shared region and an image patch cell among pixels of the image acquired by the image input unit, A feature vector extracting unit for extracting an invariant slope histogram feature vector according to the size conversion of the image patch cell using the result of the calculation and a gradient histogram feature vector for calculating the histogram of the orientated gradient, And an object detecting unit for detecting the object.

According to another aspect of the present invention, there is provided a method of detecting a fast object using a size invariant feature vector, comprising: calculating a slope histogram for an image patch cell by considering a share of a pixel having an image patch cell and a shared region among pixels of the input image, Extracting an invariant slope histogram feature vector according to the size change of the image patch cell using the calculation result, and detecting an object in the image using the slope histogram feature vector.

The apparatus and method for detecting a fast object using a size invariant feature vector according to the present invention solves the problem of increase in computation time of an object detection technique using a skew histogram feature vector according to the related art, have.

Extracting a scale invariant gradient histogram (HOG) characteristic vector for an image patch of arbitrary scale, there is an effect of detecting objects of various sizes in an image without generating a pyramid image.

The object detection time can be shortened and the specific object can be robustly detected in various external environments. Therefore, the present invention can be applied to various object detection fields such as a vehicle safety assisting system and an intelligent method system.

The effects of the present invention are not limited to those mentioned above, and other effects not mentioned can be clearly understood by those skilled in the art from the following description.

1 is a conceptual diagram illustrating an object detection process using a slope histogram (HOG) and a support vector machine (SVM) according to the prior art.
2 is a conceptual diagram illustrating an object detection process in a multi-scale image using a gradient histogram (HOG) and a support vector machine (SVM) according to the prior art.
FIG. 3 is a conceptual diagram illustrating a gradient histogram (HOG) calculation process according to the related art.
4 is a block diagram illustrating a fast object detection apparatus using a size invariant feature vector according to an embodiment of the present invention.
5 is a conceptual diagram illustrating a process of applying a learning model to image patches of various scales according to an embodiment of the present invention.
FIG. 6 is a conceptual diagram illustrating a pixel having an image patch cell and a shared area in a process of calculating a pixel occupancy rate according to an embodiment of the present invention.
7 is a flowchart illustrating a fast object detection method using a size invariant feature vector according to an embodiment of the present invention.

BRIEF DESCRIPTION OF THE DRAWINGS The advantages and features of the present invention and the manner of achieving them will become apparent with reference to the embodiments described in detail below with reference to the accompanying drawings. The present invention may, however, be embodied in many different forms and should not be construed as being limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the concept of the invention to those skilled in the art. And the present invention is defined by the description of the claims. It is to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. In the present specification, the singular form includes plural forms unless otherwise specified in the specification. It is noted that " comprises, " or "comprising," as used herein, means the presence or absence of one or more other components, steps, operations, and / Do not exclude the addition.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the drawings.

The present invention has been proposed in order to solve the problem of the prior art in which the slope histogram (HOG) feature vector is incorrectly extracted because the number of pixels per image patch cell does not become an integer according to the size change of the input image.

4 is a block diagram illustrating a fast object detection apparatus using a size invariant feature vector according to an embodiment of the present invention.

The fast object detecting apparatus using the size invariant feature vector according to the embodiment of the present invention calculates the occupancy rate of the shared patch region of the image patch cell and the pixel having the shared region among the pixels of the image acquired by the image input unit 100, A feature vector extractor 200 for extracting a histogram feature vector, and an object detector 300 for detecting an object in the image using a gradient histogram feature vector.

The feature vector extracting unit 200 according to the embodiment of the present invention calculates the occupancy rate of the pixel of the image obtained through the image input unit 100 with respect to the shared area of the pixel including at least a part of the image patch cell .

If the number of pixels per cell of the image patch cell is not an integer, the feature vector extracting unit 200 extracts a feature value of a pixel included in at least a part of the image patch cell in order to calculate a slope histogram value allocated to an area of the image patch cell. Calculate the share for the shared area. That is, when the number of pixels per cell of the image patch cell is not an integer, the feature vector extracting unit 200 calculates the occupancy rate of the pixel including a part of the image patch cell in the image patch cell and calculates a slope histogram And the slope histogram feature vector of the image patch is extracted by accumulating the calculated slope histogram.

The slope histogram feature vector extracted by the feature vector extractor 200 according to the embodiment of the present invention has a scale invariant characteristic.

That is, according to the embodiment of the present invention, the feature vector extracting unit 200 extracts a feature vector from a multi-scale image patch, and the object detector 300 uses the feature vector to determine An object may be detected, and a learning model may be applied to image patches of various scales as shown in FIG.

Therefore, according to the embodiment of the present invention, it is possible to identify a specific object without creating dozens of pyramid images, and it is possible to drastically reduce the calculation time.

FIG. 6 is a conceptual diagram illustrating a pixel having an image patch cell and a shared area in a process of calculating a pixel occupancy rate according to an embodiment of the present invention.

The image patch block shown in Fig. 6 is composed of 2 x 2 image patch cells. It is also possible that the number of cells per image patch block is different from that of the image patch block, for example, to help those skilled in the art understand the embodiments of the present invention.

6 is a diagram showing an example in which the number of pixels per image patch cell is not an integer. In the case where the pixels of the input image are all included, partially included, or not all included in the image patch cell , And can be divided into three categories.

The feature vector extracting unit 200 according to the embodiment of the present invention calculates a slope histogram value assigned to an area of an image patch cell and includes a slope histogram of the pixel 10 including all of the area of the image patch cell, The inclination histogram considering the occupation rates of the pixels 20a to 20d and 30a to 30d including a part of them is summed to calculate the total inclination histogram value.

In order to help those skilled in the art understand the embodiments of the present invention, the first classification may be applied to one side of the sides of the image patch cell And the second classification is defined as the edge pixel 30a including the vertex of the image patch cell and divided by two sides of the image patch cell 30a To 30d).

The feature vector extracting unit 200 according to the embodiment of the present invention extracts feature pixel values of the divided pixel regions 20a, 20b, 20c, and 20d, which are sets of pixels divided by one side of the image patch cells, Which is a ratio occupied by a portion occupied by the user.

In addition, the feature vector extractor 200 according to the embodiment of the present invention extracts the occupancy rate of the edge pixels 30a, 30b, 30c, and 30d including the vertices of the image patch cell and divided by two sides of the image patch cell, The occupancy rate of the edge pixel with respect to the shared area is calculated by multiplying the occupancy rate of the divided pixel areas 20a and 20c, 20b and 20c, 20b and 20d, 20a and 20d divided by the two sides.

6, (px1, py1) is the x and y coordinates of the upper left vertex of the arbitrary ith image patch cell, and (px2, py2) is the x and y coordinates of the upper right vertex of the ith image patch cell. x, and y coordinates.

That is, the edge pixels 30a, 30b, 30c and 30d include vertices of (px1, py1), (px2, py1), (px2, py2) Which corresponds to pixels divided by sides.

As shown in FIG. 6, a set of pixels divided by one side of the image patch cell and having an area shared with the image patch cell is four (a first divided pixel region 20a to a fourth divided Pixel region 20d) is formed.

  The feature vector extracting unit 200 according to the embodiment of the present invention includes a pixel 10 including the entire region of the image patch cell and pixels 20a through 20d and 30a through 30d including a part of the region of the image patch cell, The slope histogram corresponding to the k-th bin for an arbitrary i-th cell is calculated using the following equation (2). &Quot; (2) "

[Equation 2]

는 i번째 셀의 영역 j에서 k번째 bin에 대한 히스토그램을 나타내고,

는 i번째 셀의 모서리 화소 j에서 k번째 bin에 대한 기울기의 크기(magnitude)를 나타내며,

는 i번째 셀의 영역 j에서 현재 셀에 할당되는 점유율(share rate)을 나타낸다. At this time,

Represents a histogram for the kth bin in the region j of the ith cell,

Represents the magnitude of the slope with respect to the k < th > bin in the edge pixel j of the ith cell,

Represents the share rate allocated to the current cell in the region j of the ith cell.

The feature vector extracting unit 200 according to the embodiment of the present invention calculates the occupation rates of the first divided pixel region 20a to the fourth divided region 20d by using the following [Expression 3] - [Expression 6] Calculate each.

[Equation 3]

[Equation 4]

[Equation 5]

[Equation 6]

At this time, α ⁱ _r1, ⁱ _{r2 α,} α ⁱ _r3 And α ⁱ _r4 is a respective first split pixel region (20a) to the fourth divided-pixel image means a share for the patch cells, floor (x) is small up to an integer greater than x in the region (20d) shown in Figure 6, ceil (x) means the minimum integer greater than x.

That is, the (px1, py1) is (2.7, 7.3) shown in Figure 6. For example, in the case of (px2, py2) is (7.7, 2.3), α ⁱ _r1 is 0.3, α ⁱ _r2 is 0.7, α ⁱ _r3 Is 0.7 and? ^I _r4 is 0.3.

In addition, the feature vector extracting unit 200 according to the embodiment of the present invention extracts the vertexes (px1, py1), (px2, py1), (px2, py2) ), And calculates the occupancy of each edge pixel (the first edge pixel 30a, the second edge pixel 30b, the third edge pixel 30c, and the fourth edge pixel 30d) divided by two sides of the image patch cell , The occupancy rate of each edge pixel is calculated using the following [Expression 7] - [Expression 10].

[Equation 7]

[Equation 8]

[Equation 9]

[Equation 10]

At this time,? ^I _r1 ,? ^I _r2 ,? ^I _r3 , and? ^I _r4 are the first corner pixel 30a, the second corner pixel 30b, the third corner pixel 30c, when the fourth corner means a share for the image patch cells in the pixel (30d), and, (px1, py1) is (2.7, 7.3) and, (px2, py2) is (7.7, 2.3), for example, β ⁱ _r1 Is 0.21,? ^I _r2 is 0.49,? ^I _r3 is 0.21, and? ^I _r4 is 0.09.

The feature vector extracting unit 200 according to the embodiment of the present invention calculates the slope histogram allocated to the image patch cell using Equation 2 to Equation 10 as described above, It is possible to extract a slope histogram feature vector that can be applied irrespective of whether or not the histogram feature vector is used.

The object detection unit 300 according to the embodiment of the present invention converts the size of an image patch cell to search for the presence or absence of an object in the image using image patch cells of various scales, And detects the object by selecting the region of interest having the highest confidence using the Non-Maximum Suppression method.

7 is a flowchart illustrating a fast object detection method using a size invariant feature vector according to an embodiment of the present invention.

In the high speed object detection method using the size invariant feature vector according to the embodiment of the present invention, among the pixels of the image input in the test image input step (S100), the occupancy rate of the pixel having the image patch cell and the shared area is considered A step S200 of generating an integrated gradient histogram image and a step S300 of searching for an object in the image with respect to a multi-scale image patch.

In step S200 according to an exemplary embodiment of the present invention, a slope histogram allocated to an image patch cell is calculated in consideration of a share of the input image pixel with respect to the image patch cell, and the slope histogram feature vector is accumulated do.

In step S200, the feature vector extracting unit 200 according to the embodiment of the present invention uses the [Formula 2] - [Formula 10] , The occupancy rate of the pixel region 10, the pixels included in the divided pixel regions 20a to 20d, and the edge pixels 30a to 30d, all of which are included in the image patch cell, are calculated to obtain the slope histogram assigned to the image patch cell Do.

That is, in step S200, when the number of pixels per cell of the image patch cell is not an integer, the occupancy rate of the pixel divided by the side of the image patch cell is calculated, the slope histogram allocated to the image patch cell is calculated, The slope histogram feature vector of the image patch is extracted by accumulating the calculated slope histogram.

At this time, the step S200 calculates the occupancy rate of the divided pixels (20a to 20d in Fig. 6) and the shared pixels of the corner pixels (30a to 30d in Fig. 6) To 30d) is calculated by multiplying the occupancy rate of two adjacent pixel regions adjacent to the corresponding corner pixel.

In operation S300, an invariant slope histogram feature vector is used to search for the presence or absence of an object in the image according to the size change of the image patch cell.

The fast object detection method using the size invariant feature vector according to the embodiment of the present invention includes the steps of grouping a region of interest corresponding to the same object (e.g., a person) (S400), a region- (S400) of selecting a region of highest reliability by using a region of interest (e.g., Suppression).

The embodiments of the present invention have been described above. It will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims. Therefore, the disclosed embodiments should be considered in an illustrative rather than a restrictive sense. The scope of the present invention is defined by the appended claims rather than by the foregoing description, and all differences within the scope of equivalents thereof should be construed as being included in the present invention.

100: image input unit 200: feature vector extracting unit
300:

Claims (11)

Calculating a share of the pixel having the image patch and the shared area among the pixels of the image acquired by the image input unit with respect to the shared area to calculate a histogram of the orientated gradient for the image patch cell, A feature vector extractor for extracting an invariant slope histogram feature vector according to a size change of the image patch cell using the result; And
An object detection unit for detecting an object in the image using the slope histogram feature vector and detecting an object of various sizes in the image using the slope histogram feature vector that is invariant in accordance with the size change of the image patch cell,
And a size invariant feature vector.
The apparatus of claim 1, wherein the feature vector extractor
Calculating the occupancy rate of the pixel including at least a part of the area of the image patch cell with respect to the shared area when the number of pixels per cell of the image patch cell is not an integer
Fast Object Detection System Using Pinyin Invariant Feature Vector.
The apparatus of claim 1, wherein the feature vector extractor
Calculating an occupancy rate of a portion included in an area of the image patch cell among the pixels divided by sides of the image patch cell, accumulating a slope histogram allocated to the area of the image patch cell, Extracting a histogram feature vector
Fast Object Detection System Using Pinyin Invariant Feature Vector.
4. The apparatus of claim 3, wherein the feature vector extractor
Calculating the occupancy rate of a portion included in the region of the image patch cell among the divided pixel regions that are a set of pixels divided by the same side of the side of the image patch cell
Fast Object Detection System Using Pinyin Invariant Feature Vector.
5. The apparatus of claim 4, wherein the feature vector extracting unit
Wherein the occupancy rate of the edge pixel to the shared area is calculated from the edge pixel including the vertex of the image patch cell and divided by two sides of the image patch cell among the divided pixels, By multiplying the occupancy rate of the divided pixel region divided by the two sides
Fast Object Detection System Using Pinyin Invariant Feature Vector.
delete (a) calculating a slope histogram for the image patch cell by taking into account the occupancy rate of a pixel having the image patch cache and the shared area among the pixels of the input image to the shared area, Extracting an invariant slope histogram feature vector according to a magnitude transformation of the histogram feature vector; And
(b) detecting an object in the image using the slope histogram feature vector, and searching for objects of various sizes in the image using the slope histogram feature vector that is invariant in accordance with the size change of the image patch cell
And a size invariant feature vector.
8. The method of claim 7, wherein step (a)
Calculating an occupancy rate of a pixel divided by sides of the image patch cell when the number of pixels per cell of the image patch cell is not an integer, accumulating a slope histogram allocated to the image patch cell, Extracting the slope histogram feature vector
A Fast Object Detection Method Using In - Size Invariant Feature Vector.
9. The method of claim 8, wherein step (a)
Calculating a share of the divided pixel region, which is a set of pixels divided by the same side of the side of the image patch cell, to the shared region
A Fast Object Detection Method Using In - Size Invariant Feature Vector.
10. The method of claim 9, wherein step (a)
Calculating a share of a corner pixel including a vertex of the image patch cell and divided by two sides of the image patch cell among the divided pixels, for a shared area of the edge pixel, The occupancy rate is calculated by multiplying the occupancy rate of the divided pixel region divided by the two sides
A Fast Object Detection Method Using In - Size Invariant Feature Vector.
delete

Images