KR101438451B1 - Method of providing fast detection of moving objects from non-stationary camera video by dual-mode SGM, and computer-readable recording medium for the same - Google Patents

Abstract

The present invention relates to a technique for fast detecting a moving object from a video obtained by a non-stationary camera and, more specifically, to a technique for fast detecting a moving object from a video obtained by a non-stationary camera by compensating for the movement of the camera through model combination with respect to a video obtained by the non-stationary camera, preventing damage to a background model induced by foreground pixels and noise using a dual-mode single Gaussian model (SGM), and reducing complexity with the application of a background model based on grid. According to the present invention, a moving object can be accurately detected by preventing damage to the background model induced by foreground pixels and noise using the dual-mode SGM and allowing the background model to properly learn a change in the background. Further, complexity can be significantly reduced and a level of detection performance similar to other conventional algorithms can be achieved by applying a model in a grid unit while reducing the error of detection results considering an error in compensating for the movement of the camera.

Description

TECHNICAL FIELD The present invention relates to a dual-mode SGM-based high-speed moving object detection method for non-stationary camera images and a computer-readable recording medium therefor. medium for the same}

The present invention relates to a technique for detecting a moving object at a high speed from an image of an unfixed camera. More specifically, the present invention compensates camera movement through model mixing for images of non-stationary cameras, prevents background models from being corrupted by foreground pixels and noise by a dual mode SGM (single Gaussian model) And to a technique for detecting a moving object at a high speed from an image of a non-fixed camera by reducing a calculation amount by applying a background model based on the background model.

The ability to detect moving objects from images acquired by cameras has been recognized in the fields of computer vision and video surveillance, and various algorithms have been developed accordingly.

However, the conventional moving object detection method mostly relates to a technique for detecting a moving object in an image obtained from a fixed camera (C. Stauffer and W. Grimson. "Adaptive background mixture models for real-time tracking," In Computer Vision and Pattern Recognition "Proceedings of the IEEE, 2002, T. Ko, S. Soatto, and D. Estrin." Warping, IEEE Trans. Background subtraction, "In the Twenty-Third IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2010, Y. Sheikh and M. Shah." Bayesian modeling of dynamic scenes for object detection, "PAMI 2005).

Meanwhile, recently, non-fixed cameras such as a PTZ (Pan Tilt Zoom) camera have been actively applied. Non-fixed cameras are able to effectively monitor a wide range and have the ability to shoot while tracking specific objects. However, due to the movement of the camera itself, the non-stationary camera additionally generates various image effects which have not occurred in the fixed camera, and thus the existing moving object detection technique does not operate properly.

Accordingly, many conventional methodologies have been proposed for detecting a moving object from an image obtained by a non-fixed camera. However, the proposed schemes require too much computational complexity and are not suitable for real-time use in real-world applications.

See, for example, G. Georgiadis et al. In the case of "Motion estimation with independent motion estimation," In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2012, it takes about 30 to 60 seconds per frame in MATLAB simulation. Y. Sheikh et al. "The method proposed by In ICCV 2009 was also significant enough to take 2.6 seconds or more per frame.

S. Kwak et al. The method proposed in the 2011 IEEE International Conference on Incom- vision Vision (ICCV), which has excellent detection performance but is complicated in the optimization process as well as optical flow ) Operation.

Accordingly, these conventional techniques generally exhibit excellent results in the off-line, but are not applicable to real-time applications.

S. Kim et al. "Detection of Moving Objects with a Moving Camera Using a Non-Panoramic Background Model," Machine Vision and Applications (hereafter, "the paper") and J. Kim et al. The methods presented in "Fast moving object detection with non-stationary background," Multimedia Tools and Applications, 2012. have shown the speed to be applied in some real time. However, real-time performance is ensured only in high-performance PCs, and it is virtually impossible to obtain results in real time on a smartphone or embedded device.

On the other hand, O. Barnich and M. Van Droogenbroeck, which guarantee very low computational complexity for fixed cameras. There is a methodology of " ViBe: A universal background subtraction algorithm for video sequences, "IEEE Transactions on Image Processing, 2011 (hereinafter" Barnich paper "). However, this is a theory that can be effectively applied only to a fixed type camera, and when applied to a non-fixed type camera, there is a problem in that the performance is remarkably decreased with an increase in the calculation amount.

[Related Technical Literature]

1. Apparatus and method for automatically setting a motion detection area in a digital video recorder for surveillance (Patent Application No. 10-2004-0068862)

2. Pan-tilt-zoom camera and object detection method of the camera (patent application 10-2010-0043805)

It is an object of the present invention to compensate camera movement through model mixing for an image of an unfixed camera, to prevent the background model from being damaged by foreground pixels and noise by the dual mode SGM, And to provide a technique capable of detecting a moving object at a high speed from an image of a non-fixed camera.

In order to accomplish the above object, there is provided a dual-mode SGM-based moving object high-speed detection method for non-stationary camera images according to the present invention, which uses two separate single Gaussian models having mean, variance, and age as background models, A first step of setting a dual mode SGM including a candy dart background model; A second step of dividing the input image of the non-fixed camera into a plurality of grid areas; A third step of predicting a moving direction of the background motion for each of the divided grids and compensating the camera motion for the input image by mixing the background model set at the previous time with respect to the adjacent grid area; A fourth step of applying a dual mode SGM to each of the grid areas in which the camera movement is compensated; A fifth step of examining the suitability of the dual mode SGM for the input image and selectively learning the input image for either of the adapted background model and the candy dart background model according to the review result; A sixth step of swapping an ad-hoc background model and a candy dart background model of the dual mode SGM based on the age of the background model; And a seventh step of detecting a moving object from the input image of the non-fixed camera from the attached background model.

When the input image is judged to be inadequate for both the background model of the dual mode SGM and the background of the candy dots, the candy dots background model is deleted and a candy dots background model is newly generated according to the input image. And an eighth step of expanding the learning to the second step.

If the input image has an error within a critical range with respect to the background model, it is determined that the background model is appropriate and the learning of the input image is performed on the background model. If the input image has an error exceeding the critical range for the background model of the apparant background, but the candidate background model has an error within the critical range, And performing the steps on the model.

In addition, in the present invention, the seventh step determines whether or not the distance from the center of the appalant background model to the grid pixel is within the critical range by the following equation

는 미리 설정한 임계 파라미터)(At this time,

Is a predetermined threshold parameter)

And, as a result of the determination, it can be configured to detect that it belongs to the mobile body when it has the threshold range or more.

을 계산하여 배경 영상의 이동 방향을 예측하는 단계; 그 예측된 이동 방향에 따라 원근변환

를 중심으로하는 움직임 보상 배경모델 영역

를 획득하는 단계; 영역

와 겹치는 그리드 그룹

과의 겹침 면적을 반영하여 혼합 가중치

(

)를 계산하는 단계; 그리드에 대해 다음의 수학식In the third step of the present invention, a homography matrix for the grid

And estimating a moving direction of the background image; According to the predicted moving direction,

A motion compensation background model region

; domain

And overlapping grid groups

Lt; RTI ID = 0.0 > weighted < / RTI &

(

); For the grid, the following equation

는

의 조건을 만족함)(In this case,

The

Lt; / RTI >

And compensating the movement of the background image according to the movement of the non-fixed camera.

를 초과하는 경우에 다음의 수학식 At this time, in the third step, after the motion compensation,

, The following equation

는 미리 설정된 감쇄 파라미터)(At this time,

Is a predetermined attenuation parameter)

And decreasing the age of the background model through the background model.

In the present invention, the single Gaussian model constituting the background model is expressed by the following equation

는 그리드 i의 평균;

는 그리드 i의 분산;

는 그리드 i의 에이지;

는 그리드 i의 픽셀 그룹;

는 그리드 i의 픽셀 개수;

는 픽셀 j의 픽셀 인텐시티)(At this time,

The average of grid i;

Is the variance of the grid i;

The age of the grid i;

A group of pixels of grid i;

Is the number of pixels in grid i;

Is the pixel intensity of pixel j)

As shown in FIG.

On the other hand, the computer-readable recording medium according to the present invention records a moving object high speed detection program for executing the above-described dual mode SGM-based moving object high speed detection method.

According to the present invention, it is possible to prevent a background model from being damaged by foreground pixels and noise through the dual mode SGM, and the background model can appropriately learn a background change, thereby enabling accurate moving object detection.

According to the present invention, it is possible to apply the model in a grid unit while reducing the error of the detection result in consideration of the compensation error of the camera motion, thereby dramatically reducing the computation amount, and exhibiting a performance similar to that of other algorithms in detection performance There are advantages.

The present invention can be suitably applied to a variety of fields such as smartphone image capturing and video surveillance. Particularly, in the video surveillance technology, it is possible to assist the video surveillance by separately displaying the moving object from the camera image, It has the advantage of automatically focusing.

1 shows an overall structure of a dual mode SGM based moving object high speed detection algorithm for a non-fixed camera image according to the present invention.
FIG. 2 is a diagram for illustrating a concept of compensating camera movement by model mixing in the present invention. FIG.
3 illustrates the concept of a dual mode SGM according to the present invention compared to a general SGM.
FIG. 4 is a diagram showing an execution speed improvement according to the present invention; FIG.
5 is a diagram illustrating an effect according to a result of performing a dual mode SGM in the present invention.
FIG. 6 is a diagram showing the effect of camera motion compensation through model mixing in the present invention. FIG.
7 is a diagram showing a result of a fast detection algorithm according to the present invention performed in a personal computer.
FIG. 8 is a diagram showing a result of a fast detection algorithm according to the present invention performed on a smartphone; FIG.

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

1 is a diagram showing an overall structure of a dual mode SGM based moving object high speed detection algorithm for a non-fixed camera image according to the present invention.

으로 가정한다. 시간 t에서 그리드 i의 픽셀 그룹을

라 하고 해당 픽셀 그룹의 픽셀 개수를

이라 하며 시간 t에서 픽셀 j의 픽셀 인텐시티(pixel intensity)를

라 하면, 본 발명에서 해당 그리드(픽셀 그룹

)에 적용되는 SGM 모델의 평균

, 분산

, 에이지

는 다음의 [수학식 1]에 의해 업데이트된다.
First, the SGM applied to the background model will be described in the present invention. The SGM of the present invention reflects the mean, variance, and age. In the present invention, an input image is divided into grid areas in order to reduce the computation amount. In the mathematical modeling,

. At time t, the group of pixels of grid i

And the number of pixels of the pixel group

And the pixel intensity of the pixel j at time t

In the present invention,

The average of the SGM models applied to

, Dispersion

, Age

Is updated by the following equation (1).

,

,

는 시간 (t-1)에서 모델 혼합을 통해 움직임 보상이 이루어진 SGM 모델을 나타낸다. 또한, [수학식 1]에서

와

는 다음의 [수학식 2]와 같이 정의된다.
At this time,

,

,

Represents the SGM model in which motion compensation is performed through model mixing at time (t-1). Further, in Equation (1)

Wow

Is defined as the following equation (2).

와

를 모델링함으로써 그리드 내의 일부 특이한 픽셀에 의해 배경 모델이 잘못 학습되는 현상을 방지할 수 있어 바람직한 결과를 얻을 수 있었다.Equation (2) is a specialization of the SGM model in view of the fact that the present invention is applied in a grid unit. According to Equation (2)

Wow

It is possible to prevent the background model from being mis-learned by some unusual pixels in the grid, so that a preferable result can be obtained.

는 시간이 경과함에 따라 변화하게 되는데, 이를 통해 비고정 카메라의 입력 영상에 대해 카메라 움직임을 보상하는 과정에서 발생하는 보상 오류 및 일부 특이한 픽셀에 의해 배경 모델이 잘못된 영향을 미치는 것을 방지할 수 있다.
Further, according to Equation (1)

The compensation error caused in compensating the camera motion for the input image of the non-fixed camera and the background model may be prevented from being erroneously influenced by some unusual pixels.

Referring to FIG. 1, the overall flow of the moving object high-speed detection algorithm according to the present invention will be described. When an image of an unfixed camera is input (p1), the input image is preferably divided into uniform grid areas, and the camera motion is compensated by applying a model mixture as shown in Fig. 2 (p2 ).

That is, in order to compensate for the influence caused by the motion of the non-fixed camera, the present invention mixes the background models obtained at the previous time with respect to the adjacent grid area selected according to the prediction motion. In the non-fixed camera image, the pixels of the background part are also changed much in accordance with time. In particular, in the present invention, there is a risk that a large-scale error occurs due to the application of the background model in units of a grid, ], It is possible to compensate the camera movement relatively accurately through the model mixing as described below.

Then, a dual-mode Sigma Gaussian Model (SGM) (p3) is applied to the input image on which the camera motion compensation is performed through the mixture of the models. The present invention reduces the amount of computation by applying a background model in units of a grid instead of applying a background model per pixel. As described later with reference to FIG. 6, since the camera movement is compensated for the non-fixed camera image through model mixing, a good result can be obtained even when a background model is applied in a grid unit.

This dual-mode SGM is modeled specifically for the grid structure as described above, and the background is represented by a pair of an apparent backgroud model (p3a) and a candidate background model (p3b). Thereby preventing the background from being damaged by noise or foreground objects. 3 is a diagram showing characteristics of a dual mode SGM through comparison with a general SGM. The solid line in Fig. 3 (b) showing the dual-mode SGM indicates an adherent background model and the dotted line indicates a candy dart background model.

Meanwhile, the dual-mode SGM-based high-speed moving object detection method for non-stationary camera images according to the present invention is a relatively simple model, but has an age meaning information observation memory for information observation per pixel in each model image , Which keeps the correct learning rate variable and thus also learns errors that occur in areas that the model can not handle.

Also, comparison and swap (p4) are performed on the ad-hoc background model (p3a) and the candy dart background model (p3b). When learning the background model based on the input image, the two models are compared and learning proceeds according to which model is more suitable. In the dual mode SGM, depending on the age, the appalant background model and the candy dent background model are determined. Preferably, the aged background is the apparent background model and the aged one is the candy dart background model. On the other hand, when it is judged that both the adherent background model and the candy dart background model are inappropriate for the input image, the existing candy dart background model is deleted, and a candy dart background model is newly generated according to the input image.

Finally, the moving object detection result is obtained from the attached background model p3a in the dual mode SGM (p5).

Hereinafter, a dual mode SGM-based moving object high speed detection method for a non-fixed camera image according to the present invention will be described in detail with reference to FIG. 1 with reference to FIG. 2 to FIG.

FIG. 2 is a diagram for illustrating the concept of compensating camera movement (p2) by model mixing in the present invention. When the general coordinate transformation of the motion of the non-fixed camera is performed, a large error occurs in motion compensation, and the dual mode SGM gradually becomes inaccurate as the image input p1 progresses.

Accordingly, in the present invention, the motion of the camera is compensated by mixing the background model set at the previous time with respect to the adjacent grid area using the prediction result according to the background motion corresponding to the previous time t-1 and the current time t.

Hereinafter, the model mixing will be described in more detail.

을 획득할 수 있는데, 이는 배경 영상의 이동 방향에 대응한다.For the individual grids in which the input image is divided, RANSAC [MA Fischler and RC Bolles, " Random sample consensus: a paradigm for model fitting with applications of image analysis and automated cartography, "Commun. ACM, < RTI ID = 0.0 > 1981)

, Which corresponds to the moving direction of the background image.

의 센터를

라 하면 [도 2]에 나타낸 센터

의 이동은 원근변환(perspective transform)

로 표시할 수 있다. 그에 따라, 별다른 사이즈 변화가 없다면 중심이

인

영역

는 그리드

와 관련하여 비고정 카메라의 움직임이 보상된 배경 모델(motion compensated background model)로 간주된다.Specific grid

Center of

The center shown in Fig. 2

Is a perspective transform,

As shown in FIG. Accordingly, if there is no change in size,

sign

domain

The grid

The motion of the non-fixed camera is regarded as a compensated background model.

는 여러 그리드와 영역이 겹치는데, 모델 혼합을 시행할 때 이러한 겹침 영역의 면적에 비례하도록 혼합 가중치를 반영하는 것이 바람직하다. 영역

와 겹치는 그리드의 그룹을

라 하고 이들 그리드 그룹

에 속하는 그리드에 대한 혼합 가중치를

(

)라 하면, 본 발명에서 바람직하게 적용할 수 있는 모델 혼합은 다음의 [수학식 3]과 같이 나타낼 수 있다.
As shown in FIG. 2,

, It is desirable to reflect the mixed weights so as to be proportional to the area of the overlapping area when the model mixing is performed. domain

And groups of overlapping grids

These grid groups

The mixed weights for the grids belonging to

(

), The model mixture that can be preferably applied in the present invention can be expressed by the following equation (3).

에 대해서는 다음의 [수학식 4]를 설정한다.At this time,

The following equation (4) is set.

를 초과(즉,

)하는 경우에는 다음의 [수학식 5]에 따라 배경 모델의 에이지를 감소시키는 것이 바람직하다. 이처럼 에이지를 감소시킴으로써 배경 모델의 분산이 지나치게 증가하는 것을 방지할 수 있고, 이를 통해 에지 부근의 전경 이미지가 배경 모델에 영향을 미치지 않도록 해준다.
On the other hand, the motion compensation according to Equation (3) above has a problem that when the edge portion of the grid has many changes, an error that the dispersion value increases inappropriately after compensation may occur. Accordingly, after the motion compensation, the variance value is set to a predetermined threshold value

(I.e.,

), It is preferable to reduce the age of the background model according to the following equation (5). By reducing the age as described above, it is possible to prevent the background model from being excessively dispersed, thereby preventing the foreground image in the vicinity of the edge from affecting the background model.

는 미리 설정된 감쇄 파라미터)
(At this time,

Is a predetermined attenuation parameter)

Next, the dual mode SGM (p3) employed in the present invention will be described in detail. After performing the camera motion compensation (p2) according to the model blending on the input image as described above, one dual mode SGM is used for each grid. In the present invention, the dual mode SGM (p3) p3a and a candy dart background model p3b.

Each of the dual mode SGM model (p3a) and the candy dart background model (p3b) performs learning according to a general SGM model learning method, and is expressed as an average image, a distributed image, and an age image. At this time, in the individual SGM model constituting the dual mode SGM, the learning using the age, which means information observation memory for the observation of the information per pixel, proceeds with the variable learning rate. In the dual mode SGM model, the larger age is selected as the background background model p3a and the smaller age is selected as the candi dart background model p3b.

When we use the input image for the background model, we compare two models of the adherent background model (p3a) and the candy dart background model (p3b) to determine which model is more appropriate. If it is determined that both the background image model (p3a) and the background image model (p3b) are inappropriate for the input image, the existing background image model (p3b) is deleted and the background image (p3b) Create new and develop new learning.

The dual mode SGM employed in the present invention will be further described with reference to FIG. 3, which shows the characteristics of the dual mode SGM through comparison with a general SGM. Referring to the general SGM of FIG. 3 (a), a foreground (FG) image is included in the background model. However, if the learning rate is low and the input image is simple, the moving object can be detected without any problem. However, in the case of increasing the learning rate in order to achieve high-speed moving object detection on the complex input image provided by the non-fixed camera as in the present invention, when applying the general SGM of FIG. 3 (a) The problem of distortion becomes serious.

Accordingly, the dual mode SGM of FIG. 3 (b) is applied in the present invention. Referring to FIG. 3 (b), since the foreground image is hardly included in the background model in the dual mode SGM, the above problem does not occur. In the dual mode SGM, an additional candy dirt background model is provided. While the age of the candy dart background model is smaller than the background of the attached background, it is not influenced by the input image. When the age becomes larger, Swap occurs.

That is, when the average image observed for the corresponding grid of the input image has an error within the critical range when compared with the current background image model, the input image of the corresponding grid is applied to the background background model, . On the contrary, if the average image observed for the corresponding grid of the input image exceeds the critical range when compared with the current appalant background model, if the candidate background model has an error within the critical range, The image is applied to the candy dart background model to develop learning according to [Equation 1]. In this way, the model learning is selectively performed on either the appalant background model or the candy dart background model based on the fitness for the corresponding grid of the input image.

On the other hand, if the average image observed for the corresponding grid of the input image exceeds the threshold range for both the current appalant background model and the candy dots background model, the existing candy background background model is deleted and the current input A candy turtle background model is newly created based on the image.

가 변화하게 된다. 본 발명에서는 학습 결과 캔디더트 배경모델의 에이지

가 어패런트 배경모델의 에이지

보다 크면 모델 스왑이 일어나며, 그에 따라 이전의 캔디더트 배경모델이 어패런트 배경모델로 설정되고 이전의 어패런트 배경모델이 캔디더트 배경모델로 설정된다. 바람직하게는, 캔디더트 배경모델은 모델 스왑 이후에는 초기화되어 새롭게 학습을 전개한다.
Referring to Equation (1), if model learning is performed,

. In the present invention, learning result Candidate background model age

Age of the apparel background model

The model swap takes place so that the previous candy dots background model is set as an adherent background model and the previous adventure background model is set as a candy dots background model. Preferably, the Candy Dot background model is initialized after the model swap, and new learning is developed.

Next, a process of acquiring moving object detection result from the attached background background model p3a will be described.

After acquiring the background model through the above procedure, in the present invention, a pixel having a distance equal to or greater than the threshold value from the center is regarded as a foreground pixel (moving object). To be perfectly theoretically, probability calculations must be performed, but probability calculations are inadequate for real-time applications because they require tremendous computation because they require square root calculations. Accordingly, in the present invention, the computation amount is greatly reduced by calculating the distance and determining whether pixels are foreground pixels. Experimentally, even if this method is adopted, the detection performance of the moving object is not insufficient.

Mathematically, for each pixel j of the grid i, if it satisfies the following equation (6), it is classified into foreground pixels. Referring to Equation (6), only the background model is used for foreground detection. Considering the above-described learning process, it is possible to avoid a situation in which the background is misjudged by the contamination of the background model .

는 미리 설정한 임계 파라미터)
(At this time,

Is a predetermined threshold parameter)

Meanwhile, the effect of improving the execution speed by the dual mode SGM-based moving object high speed detection method for the non-fixed camera image according to the present invention is shown in the table of FIG.

As shown in the table of FIG. 4, the present invention shows an improvement in operation speed at least twice as compared with the conventional methodology. Although the first comparison algorithm (Barnich paper, "CR1") and the second comparison algorithm (Kim dissertation, "CR2") are both methodologies focused on real-time property, the algorithm of the present invention is relatively superior Performance.

Next, an effect according to the result of the dual mode SGM (p3) will be described with reference to FIG. 5A shows an average image of an existing SGM, FIG. 5B shows an average image of an attached background model of a dual mode SGM, and FIG. 5C shows an average image of a dual mode Candidate Represents the average image of the background model. 5 (d) shows the object recognition result according to the conventional SGM, and FIG. 5 (e) shows the object recognition result according to the dual mode SGM of the present invention.

As shown in FIG. 5 (d), when a conventional SGM is used, it can be seen that a foreground automobile image is learned in a background model, and a part of the object is not properly detected. However, as shown in FIG. 5 (e), when the dual mode SGM of the present invention is used, the background model is protected from the foreground object and correctly learned, and thus excellent results can be obtained.

FIG. 6 is a diagram showing the effect of camera motion compensation (p2) through model mixing in the present invention. 6 (a) and 6 (b) show the existing pixel shift-based motion compensation results, and FIGS. 6 (c) and 6 (d) Based motion compensation results.

Referring to FIGS. 6A to 6D, the conventional method does not provide a proper result for object detection. Since modeling is performed in units of a grid, Because there are many errors. However, when the model mixing method proposed in the present invention is used, good results can be obtained even if modeling is applied in a grid unit.

FIG. 7 is a diagram illustrating a result of a dual-mode SGM-based moving object high-speed detection algorithm according to the present invention performed in a personal computer. The first column of FIG. 7 shows the input image, the second column shows the detection result according to the present invention, the third column shows the detection result according to the method of Barnich's paper and the fourth column shows the detection result according to the method of Kim's paper .

And FIG. 8 is a diagram illustrating a result of a dual-mode SGM-based moving object high-speed detection algorithm according to the present invention performed on a smartphone.

As shown in FIG. 7 and FIG. 8, the dual mode SGM-based moving object high speed detection method for non-fixed camera images can be driven to a resolution of 5.8 ms for a 320x240 image without assistance of a parallelization operation in a personal computer , And real-time performance is guaranteed even in an embedded wireless terminal such as a smart phone.

In order to ensure superior performance in addition to real-time performance, the method of the present invention uses a dual-mode SGM and age to model the background and to compensate for the camera's own motion by mixing adjacent models. That is, the dual mode SGM newly proposed in the present invention prevents the background model from being damaged by foreground pixels and noise, and at the same time makes it possible to learn the change appropriately in the background.

The present invention can also be embodied in the form of computer readable code on a computer readable recording medium. At this time, the computer-readable recording medium includes all kinds of recording apparatuses in which data that can be read by a computer system is stored.

Examples of the computer-readable recording medium include a ROM, a RAM, a CD-ROM, a magnetic tape, a floppy disk, an optical data storage, and the like, and may be implemented in the form of a carrier wave . The computer-readable recording medium can also be stored and executed by a computer-readable code in a distributed manner on a networked computer system. And functional programs, codes, and code segments for implementing the present invention can be easily deduced by programmers skilled in the art to which the present invention belongs.

As described above, the embodiments of the present invention have been disclosed in the present specification and drawings, and although specific terms have been used, they have been used only in a general sense to easily describe the technical contents of the present invention and to facilitate understanding of the invention. And is not intended to limit the scope of the invention. It will be apparent to those skilled in the art that other modifications based on the technical idea of the present invention are possible in addition to the embodiments disclosed herein.

Claims (8)

A first step of setting a dual mode SGM comprising an apparent background model and a candidate background model using two individual single Gaussian models having average, variance and age as background models;
A second step of dividing the input image of the non-fixed camera into a plurality of grid areas;
A third step of predicting a moving direction of the background motion for each of the divided grids and compensating the camera motion for the input image by mixing the background model set at the previous time with respect to the adjacent grid area;
A fourth step of applying the dual mode SGM to each grid area compensated for camera motion;
A fifth step of examining the fitness of the dual mode SGM with respect to the input image and selectively learning the input image with respect to any one of the adherent background model and the candi dart background model according to the review result;
A sixth step of swapping the dual mode SGM background model and the candy dots background model based on the age of the background model;
A seventh step of detecting a moving object from the input image of the non-fixed camera from the attached background model;
Wherein the moving speed of the non-stationary camera is determined based on the moving speed of the moving object.
The method according to claim 1,
If it is determined that the input image is not appropriate for all of the apparent background model and the candy dots background model of the dual mode SGM, the Candidate background model is deleted, and a Candidate background model is newly generated according to the input image, An eighth step of developing the learning by the learning step;
The method according to claim 1, further comprising the steps of:
The method of claim 2,
In the fifth step,
If the input image has an error within a critical range with respect to the aparental background model, determining that the aparental background model is appropriate and performing the learning of the input image on the aparental background model;
If the input image has an error exceeding a critical range for the attached background background model but has an error within a critical range for the candidate background model, the candidate background model is determined to be suitable, Performing learning on the Candidate background model;
The method of claim 1, further comprising the step of:
delete The method of claim 3,
In the third step,
A homography matrix for the grid

And estimating a moving direction of the background image;
In accordance with the predicted movement direction,

A motion compensation background model region

;
The region

And overlapping grid groups

Lt; RTI ID = 0.0 > weighted < / RTI &

(

);
For the grid, the following equation

(Where t is the time; i is the grid;

A group of pixels of grid i at time t;

,

,

Lt; RTI ID = 0.0 > i < / RTI >

The average, variance, and age of the SGM model applied to the SGM model;

,

,

Denotes the average, variance, and age of the SGM model in which motion compensation is performed through model mixing at time (t-1);

Lt; RTI ID = 0.0 > i < / RTI >

Center)
(In this case,

The

Lt; / RTI >
And compensating movement of the background image according to the movement of the non-fixed camera;
The method of claim 1, further comprising the step of:
The method of claim 5,
In the third step,
After the motion compensation, the variance value is set to a threshold value

, The following equation

(At this time,

Is a predetermined attenuation parameter)
Decreasing the age of the background model through < RTI ID = 0.0 >
The method of claim 1, wherein the moving speed of the non-stationary camera is determined based on the moving speed of the moving object.
The method of claim 6,
The single Gaussian model constituting the background model is expressed by the following equation

(At this time,

The average of grid i;

Is the variance of the grid i;

The age of the grid i;

A group of pixels of grid i;

Is the number of pixels in grid i;

Is the pixel intensity of pixel j)
Wherein the moving speed of the moving object is set in accordance with the moving speed of the moving object.
A computer-readable recording medium having recorded thereon a program for causing a computer to execute a dual mode SGM based moving matter high speed detection method according to any one of claims 1 to 3 and 5 to 7.

Images