KR101527962B1

KR101527962B1 - method of detecting foreground in video

Info

Publication number: KR101527962B1
Application number: KR1020140036204A
Authority: KR
Inventors: 김진영; 뷔넉남; 민소희; 김정기
Original assignee: 전남대학교산학협력단
Priority date: 2014-03-27
Filing date: 2014-03-27
Publication date: 2015-06-11

Abstract

The present invention relates to a method for extracting a motion object from a video image, which comprises the steps of: detecting, by an edge detector, an edge of an image frame for an inputted image frame; creating a correction image by adding a predetermined weighted value to the detected edge for the image frame; extracting a first motion object based on RPCA for corresponding correction image frames when the number of obtained correction image frames reaches the predetermined number of frames; performing Gaussian filtering to remove high frequency noise for the first motion object; and generating a second motion object by filling up an empty region in an outline region with a value corresponding to the outline among data obtained through the filtering. According to the present invention, the method can satisfactorily detect the motion object even for an image frame having a small amount of about ten frames, thus processing rate for extracting the motion object is enhanced.

Description

A method of extracting motion objects from a video image,

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a motion object extraction method for a video image, and more particularly, to a motion object extraction method for extracting a foreground from a continuously input video image.

In order to extract and recognize a moving object in a continuously input sequence, it is necessary to separate a foreground from a background.

A method of extracting motion objects by subtracting consecutive images from each other in order to detect motion objects and extracting motion images is disclosed in various Korean Patent No. 10-0377067.

However, if the background image is fixed, the background image may be easily stored and removed. However, in this case, noise objects are generated due to distortion of the camera, distortion caused by the change of illumination, and the like.

Therefore, even in a fixed background, robust methods for such distortion are needed.

In addition to the method of using the difference of before and after images, there is also a method of separating the background by using an optical flow method or a simple cumulative value of an image.

On the other hand, there is a method of separating a motion object (foreground) using RPCA (robust principle component analysis) in the recently proposed method.

However, in order to obtain a good motion object extracting performance when successively input video images are directly processed by RPCA, it is necessary to apply a method of performing arithmetic operations on a long time image frame, for example, 30 frames or more.

As a result, when RPCA is applied to real-time image processing, performance degradation is serious and even application is impossible. This is because a limited number of frame images must be used for real-time processing. Even if this is not real time, there is a problem that the increase in the number of frames increases the calculation amount and increases the calculation burden.

It is an object of the present invention to provide a method of extracting a motion image of a video image that can extract a good motion object by RPCA even with a small number of frames.

According to another aspect of the present invention, there is provided a method of extracting a motion object from a video image, the method comprising: Detecting an edge of an image frame by an edge detector with respect to an input image frame; I. Generating a corrected image by adding weight values set for edges detected for the image frame; All. Extracting a primary motion object based on RPCA for the corresponding corrected image frames when the number of corrected image frames obtained through the steps a and b reaches a set number of target frames; la. Performing Gaussian filtering to remove high frequency noise for the primary motion object extracted in the multi-step; hemp. And generating a secondary motion object by filling a blank area in the outline area of the data obtained through the step a) with a value corresponding to the outline.

Preferably, the step further comprises applying a canny edge detector to detect an edge.

In addition, the second step generates the corrected image by multiplying the edge detected for the gray image obtained by binarizing the image frame by a value of 255.

More preferably, And generating final motion object information by removing an area less than a predetermined basic size among the motion object areas generated through the step.

According to the method for extracting a motion object of a video image according to the present invention, a motion object can be detected well for a small number of image frames of about 10 frames or less, thereby providing an advantage that a processing speed for motion object extraction is improved.

FIG. 1 is a flowchart showing a motion object extraction process of a video image according to the present invention,
FIG. 2 is a flow chart showing the edge detection process of FIG. 1,
FIG. 3 is a view for explaining a blank area filling process of FIG. 1,
4 is an image showing an example of extracting a motion object according to the present invention.

Hereinafter, a method for extracting a motion object of a video image according to a preferred embodiment of the present invention will be described in detail with reference to the accompanying drawings.

FIG. 1 is a flowchart illustrating a motion object extraction process of a video image according to the present invention.

First, an edge of an image frame is detected by an edge detector with respect to an image frame sequentially input from a video image (step 10).

Here, the edge detector detects an edge by applying a canny edge detector, and a detailed description will be given later.

Next, a weighted value set for edges detected for the image frame is added to generate a corrected image, and the generated corrected image is stored in a queue (Que) for each frame (step 20).

In step 20, a gray image obtained by binarizing the image frame from 0 to 255 is multiplied by a value of 255 for edges detected to generate a corrected image.

Next, it is determined whether the number of corrected image frames obtained has reached the set target frame number T (step 30). If it is determined that the number of corrected image frames T has reached the target frame number T, (Step 40) based on the RPCA.

Preferably, the number of target frames is 8, 12, more preferably 10.

Thereafter, Gaussian filtering is performed on the extracted primary motion object to remove high-frequency noise (step 50), and a process of filling a hole in the outline area of the data obtained by the filtering with a value corresponding to the outline is performed (Step 60).

Here, the contour line refers to a line that follows the boundary line of an object formed between the edges in the form of a lung orbit.

After step 60, an area of the generated motion object area that is less than the set basic size is removed (step 70), and final motion object information is generated (step 80).

Here, the removal target motion object may be appropriately applied in consideration of the number of pixels of the applied image, and preferably, the object corresponding to the region of 1000 to 2000th of the number of pixels of the frame image is removed.

As an example, in the case of an HD (1280x720) class image, an area smaller than a size corresponding to 300 pixels may be set to be removed.

Hereinafter, the motion object extraction process will be described in more detail.

First, let's look at the definition of RPCA.

RPCA was proposed by E. J. Candes et al. (E.J. Candes, et al., "Robust Principal Component Analysis," ACM, Vol 58, 2009). First, assume that there is a large-scale data matrix M. The PCA then expresses the matrix M as: " (1) "

In the above matrix representation, Lo has a low rank and No is a small perturbation matrix. Here, RPCA is extended to be expressed by the following equation (2).

In Equation (2)

Is the nuclear norm of the matrix

silver

norm. When the above equation is solved, L of low rank represents the background, and sparse S represents the motion object (foreground).

The above problem can be solved by convex optimization.

In the motion object separation of the image processing, the matrix M becomes a matrix of vectorized image frames.

However, when the RPCA is applied in real time, a problem arises. As described above, there is a large amount of calculation and a considerable time frame image must be scaled. In order to solve this problem in the present invention, (RPCA) using edge information.

That is, a corrected image which is an edge-enhanced image is generated, and an RPCA is performed from the generated corrected image.

When edge information is used, irregular noise may be generated inside the motion object. To solve this problem, Gaussian filtering and hole filling are used.

First, edge information is extracted using a canny edge detector as an edge detector.

The edge image detected by the canny detector

Let's say. Where k is the frame index. The Canny edge detector performs processing as shown in Fig.

First, the input image is smoothed through a Gaussian filter (step 11).

These Gaussian filters are used for noise reduction.

The Gaussian filter uses a 5x5 next binaural filter.

Where A is the input image.

Next, the gradient or gradient is calculated for the Gaussian filtered image (step 12).

Gradient calculations are done using the Sobel edge detector below.

Afterwards, the intensity of the gradient

(Step 13).

Next, in the obtained G image, a local maxima which is a value larger than the surrounding value is found and determined as a candidate of an edge (step 14), and hysteresis is performed on the determined edge candidate (step 15).

That is, a strong edge is determined, and a weak edge survives when it is connected to a strong edge.

The edge image obtained by the canyon edge detector is binarized, that is, represented by 0 and 1

(K) is obtained through Equation (3) to obtain an edge enhancement image as described in step 20 above.

Where I (k) is a P × Q two-dimensional image. That is, the gray (black and white) image of the original image is multiplied by the value of the edge binary image by 255, and added to the original image. This image is the edge-enhanced corrected image.

These corrected images are stored in a buffer, which can be expressed by Equation (4) below.

In Equation (4)

Dimensional image signal

Into a vector.

Thus, Buff (k) is a PQ 占 N matrix, and past N-1 images are added to the current frame number k and put in Buffers Buff (k).

RPCA is applied to the above Buff (k) matrix to obtain a matrix S (k). S (k) is a sparse matrix of the matrix Buff (k) and corresponds to a motion object (forground). In other words

, And L (k) is a low rank and corresponds to a background.

At this time, RPCA

subject to

And is calculated by the Principal Component Pursuit by Alternating Directions algorithm by the calculation method shown in Table 1 below. In Table 1 below, M corresponds to Buff (k).

Through this process, the image is reconstructed from S (k) corresponding to the motion object (forground), and Gaussian filtering is performed on the image as described above.

The Gaussian filter

Dimensional Gaussian filter is shown below.

Then, hole filling is performed on the noise-removed motion object estimation image using Gaussian filtering, and a well-known morphological filter is used for filling the empty space.

In this empty space filling process, as shown in FIG. 3, a point of a hole in a boundary of a moving object A having a noise removal process and a bounded boundary is subjected to the following Equation 4 for Xo Repeat the process of filling empty space.

Here, B is a symmetric strucuture as shown in Fig. 3, A ^c is a complementary image, and is a complementary image. Equation (4)

.

Finally, the moving object is separated, and the object occupying the small area is removed to remove the noise object.

As an example of such a processing procedure, as shown in FIG. 4, which is performed on 10 frames, the motion object can be extracted with a small number of frames of about 10 frames or more.

Claims

delete

A method for extracting a motion object from a video image,
end. Detecting an edge of an image frame by an edge detector with respect to an input image frame;
I. Generating a corrected image by adding weight values set for edges detected for the image frame;
All. Extracting a primary motion object based on RPCA for the corresponding corrected image frames when the number of corrected image frames obtained through the steps a and b reaches a set number of target frames;
la. Performing Gaussian filtering to remove high frequency noise for the primary motion object extracted in the multi-step;
hemp. And generating a secondary motion object by filling a blank area in the outline area of the data obtained through the step a) with a value corresponding to the outline,
Wherein said step further comprises applying a canny edge detector to detect an edge,
The second step generates the corrected image by multiplying the edge detected for the gray image obtained by binarizing the image frame by a value of 255,
Wherein the number of the target frames is 8 to 12.