KR101460517B1 - Method or providing visual tracking from video by learning and unlearning with dual modeling, and computer-readable recording medium for the same - Google Patents

Abstract

The present invention relates to a method for tracking an object in an image through dual modeling and learning and discard learning. More particularly, the present invention relates to various obstacles that may occur when an object is tracked in an input image through object tracking technology through dual modeling and learning and discard learning, such as obstruction of an object, background interference, And the like. According to the present invention, it is possible to solve background interference, masking, and drifting problems that are the conventional problems in continuously tracking an object region from an input image. In other words, it shows robust performance against background interference and blindness by utilizing the likelihood that the two models are combined, and provides advantages of robustness to drifting through learning / discarding learning of models using tracking results.

Description

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for tracking an object in an image through dual modeling, learning, and discard learning, and a computer-readable recording medium for the same.

The present invention relates to a method for tracking an object in an image through dual modeling and learning and discard learning. More particularly, the present invention relates to a method and apparatus for tracking objects in an input image by using dual-modeling and object tracking techniques through learning and discarding learning, including various obstacles that may occur when tracking an object in an input image, such as obstruction of an object, background interference, drifting ) At the same time.

Continuous tracking of a specific object in an image is a very important research topic in the field of computer vision. Image tracking is used as a basic processing step in various applications such as robot vision, video surveillance, video analysis, home automation and behavior recognition, and these applications require accurate tracking results. Various tracking methods have been continuously proposed for this purpose.

However, there are some obstacles in the object tracking process in the image. For example, if the tracking object is blocked by another object or if the tracking object is not noticeable compared to the background, the tracking performance is degraded. Also, when tracking an object over a long period of time, the problem of poor discrimination of the object model, which is referred to as a drifting problem, remains unresolved.

Accordingly, various methods have been proposed in relation to object tracking in a situation where the tracking object is obscured or background interference exists. A. Adam et al. "Robust fragments-based tracking using the integral histogram," In Computer Vision and Pattern Recognition, 2006 IEEE Computer Society Conference (hereafter, "Adam paper"), Methodology has been proposed.

In Mei and H. Ling, "Robust visual tracking using l1 minimization," In Computer Vision, 2009 IEEE 12th International Conference (hereafter, "Mei paper") deals with tracking problems in a conventional particle filter framework as a sparse approximation And to solve the problem of occlusion. In their method, a sparse representation of the object was generated using the tri-vial template and the l ₁ minimization technique, which showed robust characteristics over the conventional method.

To solve the background interference problem, H. Grabner et al. "Proceedings of the British Machine Vision Conference, 2006" (Grabner paper) proposed a method of learning a classifier that distinguishes between background and object on-line. This method falls within the category of classifying object tracking into two classes, which is described by S. Stalder et al. "Beyond semi-supervised tracking: Tracking should be as simple as detection," In Computer Vision Workshops (ICCV Workshops), 2009 ("Stalder"), .

However, in the case of the methods described above, only one of the occlusion situation and the background interference situation is considered, and if the obstacle occurs at the same time, the tracking performance is degraded.

Drifting is also a problem with conventional tracking methods. In order to cope with this, it is necessary to update the object model on-line while performing tracking.

Conventionally, in the tracking process, the tracking target is updated by using the result of the previous frame. Incorrect information may be included in the updating process. If such inaccurate information updates continue, the model will "drift" from the correct object model and create the wrong model.

To solve this drifting problem, B. Babenko et al. "Multi-instance learning is applied by extending the existing Grabner paper in" Pattern Analysis and Machine Intelligence, "IEEE Transactions, 2011 (" Babenko Paper "). Through this, the Babenko paper reduced the risk of learning false information by learning not only the tracking results of the previous frame but also various random changes around it. The Stalder thesis also avoided this drifting by using a multi-classification system that includes non-learning parts.

Z. Kalal et al. "Pn learning: Bootstrapping binary classifiers by structural constraints," In Computer Vision and Pattern Recognition (CVPR), 2010 IEEE Conference ("Kalal"), So that the two methods complement each other to prevent drifting. However, these methods have limitations because they do not consider the occlusion situation.

An object of the present invention is to provide an object tracking method in an image through dual modeling and learning and discard learning. More particularly, it is an object of the present invention to provide a method and apparatus for tracking an object in an input image by using an object tracking technology through dual modeling and learning and discarding learning, various obstacles that may occur when tracking an object in the input image, And to solve the problem of the problem simultaneously.

According to another aspect of the present invention, there is provided a method for tracking an object through dual modeling, learning and discard learning, comprising: a first step of setting foreground models and background models using two individual single Gaussian models; A second step of updating the background model by compensating the background motion based on the feature in the input image; The third step is to calculate the likelihood from the foreground model and the background model by calculating the likelihood value inversely proportional to the error when reconstructing the foreground model, ; A fourth step of repeatedly executing an average moving process using random sampling to maximize likelihood to obtain a tracking result for a tracking object; And a fifth step of updating the foreground model by selectively applying learning and discard learning depending on whether the foreground model is reflected in the likelihood calculation or whether the background model is reflected in the likelihood calculation.

In the third step of the present invention, the background model and the foreground model use an image represented by one average RGB color, and an age image to remember how much information is observed per pixel in each model image, .

)로 설정하고, 변환(

)를 사용한 이미지(

)에 대한 호모그래피 변환을

라고 설정하는 경우, 변환(

)을 적용시킨 전경 모델을 수학식Also, in the present invention, the third step is to convert the vector representing the transformation applied to the foreground model

), And the conversion (

) (

) ≪ / RTI >

, The conversion (

) Is applied to the equation

및

And

와

를

와

에서 추적 후보 영역(

)을 잘라내어

및

와 동일한 크기로 확대/축소함으로써 연산하는 것이 바람직하다.And a background model corresponding to the foreground model

Wow

To

Wow

In the tracking candidate region (

) And cut

And

And enlarging / reducing the same size.

)에서 추적 후보 영역(

)을 잘라내어 확대/축소하여

를 구하는 경우, 추적 후보 영역(

)에서 변환(

)에 대한 우도

는 픽셀 인텍스 j에 대해In addition, in the present invention,

) To the tracking candidate region (

) To zoom in and out

, The tracking candidate region (

) To convert

) Likelihood

For pixel index j

및

각각은, Lt; / RTI >

And

Respectively,

. ≪ / RTI >

와

중 어떠한 모델이 우도 계산에 활용되었는지에 따라 전경 모델의 업데이트에서 학습을 적용할 것인지 폐기학습을 적용할 것인지를 결정한다.In addition, in the fifth step of the present invention,

Wow

Determines whether to apply learning or revocation learning in updating the foreground model, depending on which model is used for likelihood calculation.

In addition, in the fifth step of the present invention, when updating the foreground model

및

는 시간 t일 때 픽셀 j에 대한 전경 모델이며,

는 시간 t일 때 픽셀 j에 대한 확대/축소된 관측 값이며,

는 시간 t까지 관측된 프레임 수로 설정된다.Lt; RTI ID = 0.0 >

And

Is the foreground model for pixel j at time t,

Is the zoomed-in observation for pixel j at time t,

Is set to the number of frames observed until time t.

는 픽셀의 에이지로부터 구할 수 있으며, 픽셀 j의 시간 t일 때의 에이지를

라 하면

는 In the fifth step of the present invention,

Can be obtained from the age of the pixel, and the age at time t of the pixel j

If you say

The

는 업데이트 파라미터로 [-1, 1] 범위 안에 속하게 되며 학습과 폐기학습을 조율하는 것이 바람직하다.Lt; / RTI >

Is within the range of [-1, 1] as an update parameter and it is desirable to coordinate learning and discard learning.

, 배경 모델이 사용된 픽셀들을

라 할 경우, j ∈

의 경우

(학습 - learning) 사용하며, j ∈

의 경우

(폐기학습 - unlearning)을 사용하는 것이 바람직하다.Also, in the present invention, in the fifth step, pixels in which the foreground model is used

, Pixels using the background model

, Then j ∈

In the case of

(Learning-learning), and j ∈

In the case of

It is desirable to use unlearning.

을 사용하는 경우

가 [1 , ∞] 범위 안에 속하고,

로 정의할 경우

는 다시 [0, 1] 범위 안에 속하며,In the fifth step of the present invention,

If you use

Is in the range [1, ∞]

If you define

Is again in the range [0, 1]

.

Further, in the present invention, the second step is to update the background model

및

는And,

And

The

이면

, j ∈

이면

를 사용하는 것이 바람직하다., And in the case of the background model j < RTI ID = 0.0 >

If

, j ∈

If

Is preferably used.

Meanwhile, a computer-readable recording medium according to the present invention records an object tracking program for executing an object tracking method using dual modeling and learning and discard learning according to the present invention.

According to the present invention, it is possible to solve background interference, masking, and drifting problems that are the conventional problems in continuously tracking an object region from an input image. In other words, it shows robust performance against background interference and blindness by utilizing the likelihood that the two models are combined, and provides advantages of robustness to drifting through learning / discarding learning of models using tracking results.

1 illustrates an overall configuration of an object tracking algorithm through dual modeling and learning and discard learning according to the present invention.
FIG. 2 is a diagram exemplifying a dual model in the present invention. FIG.
FIG. 3 exemplarily shows reconfiguration of likelihood calculation in the present invention; FIG.
[FIG. 4] is a diagram for explaining the operation and effect of the present invention. [FIG.
FIG. 5 is a diagram showing an experimental image and a main tracking result frame for each algorithm to explain the operation and effect of the present invention. FIG.

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

1 is a diagram showing the overall configuration of an object tracking algorithm according to the present invention. The object tracking method using dual modeling and learning and discard learning according to the present invention presents an algorithm for simultaneously solving background interference, occlusion, and drifting problems through the entire configuration as shown in FIG.

In other words, it is noted that the human eye can easily distinguish the object from the background, thereby solving the problem of background interference, occlusion, and drifting easily. The object tracking method of the present invention includes two- (Dual Modeling), and construct the likelihood by using them at the same time.

By using the thus constructed likelihood, the present invention can exclude information provided from objects other than the tracking object, and thereby it is possible to obtain robust performance against background interference and occlusion.

Also, if the information does not come from the tracked object, it solves the drifting problem by unlearning it from the object model. Finally, the method of the present invention maximizes likelihood by iterating the average moving process using random sampling.

According to the present invention, according to the present invention, when a specific person is selected and automatically tracked in an input image obtained from, for example, a PTZ surveillance camera, two foreground and background images are depicted using two single Gaussian models Modeling), a likelihood combining these two models is formed, and the likelihood can be tracked by maximizing the likelihood efficiently through an average moving method using random sampling. According to the present invention, background interference, masking, and drifting can be overcome in continuous object tracking.

Hereinafter, the present invention will be specifically described with reference to FIG. 2 to FIG. 5 based on FIG.

First, a description will now be made of a dual modeling used in the present invention to represent a current input image (input image). The dual model consists of two models, each model expressed in a single Gaussian model (SGM) or the like. These two models are the foreground model and the background model, which combine the two models to calculate the likelihood of tracking.

We then use the likelihood to obtain the final position of the object through an average shift using random sampling.

The background model and the foreground model are preferably represented by an average RGB color, which means that the background and the object to be tracked are each represented as a single image. In addition, there is an age image to remember how much information was observed per pixel in each model image.

,

,

로 표기하고 배경에 대한 RGB 이미지와 에이지 이미지를

,

,

로 표기한다. 또한, 입력 이미지는

로 표기한다. In the following description, the RGB image and the age image for the foreground

,

,

And the RGB image and the age image for the background

,

,

. Also, the input image

.

,

,

,

,

,

의 예시이다. 즉, [도 2]는 본 발명에 따른 이중 모델의 예시로서 전경 모델에 대한 평균 이미지 (가), 분산 이미지 (나), 에이지 이미지 (다)로 나타내며, 배경 모델에 대한 평균 이미지 (라), 분산 이미지 (마), 에이지 이미지 (바)로 나타낸 것이다[Figure 2]

,

,

,

,

,

. 2 shows an example of a dual model according to the present invention. The average image (a), the distributed image (b), and the age image (c) Distributed image (e), age image (bar)

,

,

는 32×32로 고정하였고

,

,

는 실제 입력 프레임과 동일한 크기로 설정하였다.
In the dual modeling according to the present invention, regardless of the size of the actual object to be tracked

,

,

Was fixed at 32 x 32

,

,

Is set to the same size as the actual input frame.

)에 대해 추적 절차에서 사용될 우도는 배경 모델과 전경 모델을 복합적으로 사용하여 계산된다. 우도는 두 모델을 이용하여 현재 후보 영역의 관측 값을 잘 표현할수록 높은 값을 갖도록 설계한다. 이는 다시 말해 전경 모델에 변화를 준 후 배경 모델과 결합하여 현재의 관측 값(입력 영상)과 가장 유사하도록 재구성하고, 재구성할 때의 오류가 적을수록 높은 우도 값을 갖도록 설계하였음을 의미한다.Next, the likelihood calculation of the dual modeling used for object tracking in the present invention will be described. Track candidate region (

), The likelihood used in the tracking procedure is calculated using a combination of the background model and the foreground model. Likelihood is designed to have a higher value with better representation of the observed values of the current candidate region using the two models. In other words, it means that the change is made to the foreground model and then reconfigured to be similar to the current observation value (input image) in combination with the background model, and the higher the likelihood value, the smaller the error in reconstruction.

3 is a diagram exemplarily showing reconstruction of likelihood calculation in the present invention. FIG. 3 (a) shows the observed image of the candidate region, FIG. 3 (b) shows the reconstructed image of the candidate region, FIG. 3 3 (D)] represents the average value of the background model. On the other hand, in these drawings, the part of the age less than a certain value indicating the information observation memory for the information observation made by the pixel is displayed in sky blue.

)는 전경 모델에 적용되는 변환을 표현하는 벡터이다. 변환(

)를 사용한 이미지

에 대한 호모그래피 변환을

라 할 때, 이 변환을 적용시킨 전경 모델을 수학식 Any transformation used in the present invention (

) Is a vector representing the transform applied to the foreground model. conversion(

)

Homography transformation for

, The foreground model to which this transformation is applied is expressed by Equation

및

And

.

와

를

와

에서 추적 후보 영역(

)을 잘라내어

및

와 동일한 크기로 확대/축소함으로써 구한다. 이후 마찬가지로 입력 이미지(

)에서 추적 후보 영역(

)을 잘라내어 확대/축소하여

를 구한다. 이 경우 추적 후보 영역(

)에서 변환(

)에 대한 우도

은 다음과 같이 [수학식 1]에 의해 정의된다.The corresponding background model

Wow

To

Wow

In the tracking candidate region (

) And cut

And

And enlarging / reducing the same size. Likewise,

) To the tracking candidate region (

) To zoom in and out

. In this case,

) To convert

) Likelihood

Is defined by Equation (1) as follows.

및

은 하기의 [수학식 2]에 의해 연산된다.
here

And

Is calculated by the following equation (2).

또는 변환 절차에 의해

가 존재하지 않는 픽셀들의 경우(경계를 벗어난 픽셀들의 경우) 이는 특별한 경우로 취급하며

만을 사용하였다.Where j is the pixel index. These equations imply that foreground and background models are compared for each pixel j and the lesser of the two values is included in the likelihood value. But if

Or by a conversion procedure

In the case of pixels that do not exist (for out-of-bounds pixels) this is treated as a special case

Respectively.

와

중 어떠한 모델이 사용되었는지에 기초하여 이를 전경 모델 업데이트에서 학습(learning)을 적용할 것인지 아니면 폐기학습(unlearning)을 적용할 것인지를 결정한다.
Thus, in the present invention, it is determined which model is used in the likelihood calculation,

Wow

Based on which model is used, determines whether to apply learning or unlearning in the foreground model update.

: 이 경우는 추적 결과 영역)에 대하여 다음과 같은 [수학식 3]에 의해 업데이트가 수행된다.
Next, we will look at updating the foreground model of dual modeling in detail. The foreground model update proceeds similar to the general online SGM (see Stalder article), but applies learning or revocation learning based on which model is used. Given a candidate region (

: In this case, the tracking result area) is updated by the following equation (3).

및

는 시간 t일 때 픽셀 j에 대한 전경 모델이며

는 시간 t일 때 픽셀 j에 대한 확대/축소된 관측 값이다.

는 시간 t까지 관측된 프레임 수, 즉 픽셀의 에이지로부터 구할 수 있다. 픽셀 j의 시간 t일 때의 에이지를

라 하면

는 [수학식 4]와 같이 정의할 수 있다.
here

And

Is the foreground model for pixel j at time t

Is the zoomed-in observation for pixel j at time t.

Can be obtained from the number of frames observed until time t, i.e., the age of the pixel. The age at time t of pixel j

If you say

Can be defined as shown in Equation (4).

는 업데이트 파라미터로서 일반적인 응용에서는

= 1로 설정된다. 하지만 본 발명의 경우에 업데이트 파라미터

는 [-1, 1] 범위 안에 속하게 되며 학습과 폐기학습을 조율한다.here

Is an update parameter,

= 1. However, in the case of the present invention,

Falls within the range [-1, 1] and coordinates learning and disposal learning.

, 배경 모델이 사용된 픽셀들을

라 할 경우, j ∈

의 경우

을 사용하며(학습 - learning), j ∈

의 경우

을 사용한다(폐기학습 - unlearning).Learning and discard learning is determined based on which model is used from Equation (1). The foreground model uses pixels

, Pixels using the background model

, Then j ∈

In the case of

(Learning-learning), j ∈

In the case of

(Discard learning - unlearning).

을 사용하는 경우

가 [0 , 1] 범위 안에 속하게 되어 일반적인 온라인 SGM과 동일하다. 하지만

을 사용하는 경우

가 [1 , ∞] 범위 안에 속하게 된다.

If you use

Is in the range [0, 1] and is the same as a normal online SGM. However

If you use

Is in the range [1, ∞].

로 정의할 경우

는 다시 [0, 1] 범위 안에 속하게 되며 [수학식 3]은 다음과 같이 [수학식 5]로 다시 정의할 수 있다.here

If you define

Is again in the range [0, 1], and [Equation 3] can be redefined as Equation (5) as follows.

로 표현하면

가

를

에 학습함으로써 얻어진 결과, 즉

가

로부터 를 r^-에 비례한 양 만큼 제거한 폐기학습된 결과임을 의미한다. [수학식 3]의 경우에도 마찬가지로 해석이 가능하다. 이를 통하여 모델로부터 물체가 아닌 정보를 제거하는 동시에 지속적으로 물체의 변화를 학습하는 것이 가능하다. 따라서 본 발명에 따른 알고리즘은 드리프팅 문제에 강인한 성능을 갖게 된다.
this is

In terms of

end

To

The results obtained by learning

end

from Is removed and learned by removing the amount proportional to r ^- . Equation (3) can also be interpreted in the same way. Through this, it is possible to continuously learn the change of the object while removing non-object information from the model. Therefore, the algorithm according to the present invention has a robust performance against the drifting problem.

Next, the background model update of dual modeling will be discussed in detail. Generally, when tracking an object, the object is tracked under a moving background rather than a fixed background. Therefore, it is essential to compensate the background motion in order to use the background model.

를 32×24의 균등 구역으로 나눈 후, 각 구역의 귀퉁이 점마다

로의 KLT [C. Tomasi and T. Kanade, "Detection and tracking of point features," Technical report, Carnegie Mellon University, Apr. 1991 참조]를 수행한다.To this end, the present invention uses a conventional feature-based motion compensation method. First, input image

Is divided into 32 x 24 equally divided zones, and then each corner of each zone

KLT [C. Tomasi and T. Kanade, "Detection and tracking of point features," Technical report, Carnegie Mellon University, Apr. 1991).

로부터

의 호모그래피 변환행렬

를 구한다. 이후 호모그래피 변환행렬

를 이용하여

,

,

를 현재 관측 값에 대응되도록

,

,

로 변환한다.
Then, based on these point tracking results, RANSAC [MA Fischler and RC Bolles, "Random sample consensus: a paradigm for model fitting with applications of image analysis and automated cartography," Commun. ACM, 1981]

from

Homography transformation matrix

. Then the homography transformation matrix

Using

,

,

To correspond to the current observation value

,

,

.

,

,

를 바탕으로 전경 모델과 유사한 방법을 사용하여 하기의 [수학식 6] 내지 [수학식 7]을 통해

,

,

를 구한다. 다만 배경 모델의 경우 전경 모델과는 반대로 전경 모델에서 학습이 수행된 부분에 폐기학습이 수행되어야 한다. 따라서 배경 모델의 경우 j ∈

일 경우

, j ∈

일 경우

를 사용한다.
These transformed models

,

,

(6) to (7) using the method similar to the foreground model on the basis of the following equations

,

,

. However, in the case of the background model, as opposed to the foreground model, the discard learning should be performed on the part where the learning is performed in the foreground model. Therefore, for the background model, j ∈

If

, j ∈

If

Lt; / RTI >

및

는 하기의 [수학식 7]에 의해 연산된다.
here,

And

Is calculated by the following equation (7).

의 양 만큼 폐기학습을 적용한다. 이러한 업데이트 과정을 통하여 추적 대상이 또한 배경에 학습되는 것을 방지하는 것이 가능하다.
In the actual implementation, after applying the learning to the whole background model,

The discard learning is applied. Through this update process, it is possible to prevent the tracking object from being learned in the background as well.

을 찾는 과정이며 하기의 [수학식 8]로 표기할 수 있다.
Next, a detailed description will be given of tracking using random sampling and average movement. The object tracking problem is an estimate of the area in which the object is located with respect to time t

And can be represented by the following equation (8).

값을 갖는

을 찾는 것이다. 이를 위하여 R을 고정시킨 후 평균이동 벡터

를 구하고, 다시 이 벡터를 활용하여

을 업데이트시킨다.In other words, without changing any of the foreground models,

Value

. For this purpose, after fixing R,

, And again using this vector

.

The method of obtaining the average motion vector is similar to that of the conventional technique (A. P. Leung and S. Gong, "Mean shift tracking with random sampling," In Proc. BMVC 2005). However, in the case of the present invention, each random conversion is sampled rather than a method of sampling an existing pixel.

, 시간 t에서의 추적 결과를

라 할 경우,

로 시작한다. 이후 매 반복실행 마다 M개의 샘플을 이용할 경우 첫 M/2개는 다음의 [수학식 9]와 같이 선택된다.
The candidate region in the k < th >

, The tracking results at time t

In other words,

≪ / RTI > Then, when M samples are used for every repetition, the first M / 2 is selected as shown in the following equation (9).

로 대칭이 되도록 한다. 이후 각

에 대하여 [수학식 1]을 이용하여

를 계산하고, 다음의 [수학식 10]에 의해 평균이동 벡터

를 계산한다.
Here, N (0, σ ² ) is a Gaussian distribution in which 0 is an average and σ ² is a dispersion, and i is a sample index. To effectively use these samples for average shifting, the remaining M / 2 cases

To be symmetrical. Then each

Using Equation (1)

And calculates the average motion vector < RTI ID = 0.0 >

.

Where l denotes an element of a vector. The candidate region is updated according to the following Equation (11) using the average motion vector of Equation (10).

Such repetitive execution is performed until convergence is achieved, but as a result of experiments, it is found that there is not a large difference in the results even if it is fixedly executed by performing a predetermined number of times, for example, 10 times.

For comparison and evaluation of the effect of the present invention, the method of the present invention is implemented in C ++. In implementation, KLT uses openCV library. The maximum age of the model is limited to 30 so that the model has a learning rate of at least 0.03. Th is set to 5 and r + = 1 and r- = -0.05. In the case of sampling σ, 10% of the model size was set in the x and y directions (5% in case of FaceOcc). The algorithm according to the object tracking method through dual modeling and learning and discard learning of the present invention showed a performance of more than 10 frames per second.

Six different methods and comparative experiments were performed to evaluate the tracking performance of the present invention. The methods of FRAG (see Adam's paper) and L1 (see Mei's paper) are mainly related to masking. In the case of SEMI (Stalder paper) and OAB (Grabner paper), background interference and drifting are discussed simultaneously. MIL Thesis) and TLD (Kalal thesis) are the latest methods focusing on drifting.

Experiments were performed on nine images, seven of which are obscured, background interference, and the other two are long images produced by repeating two of the seven images back and forth. Caviar image is C. project page. Caviar test case scenarios. It was used in http://homepages.inf.ed.ac.uk/rbf/CAVIARDATA1/, visited on 9/15/2010. FaceOcc and Woman were used in Adam paper. Sylvester and Tiger were used in the Babenko paper, and CupOcc and Cheetah were made in-house. In this comparative experiment, we used the program code provided by the authors of each paper.

Here, for the quantitative analysis, the average error of the four corner points of the tracking result box was used. As a result of the actual tracking, a person 's own point was used. Except when the tracking method did not return a result when calculating the average error (that is, when the tracking failure was recognized). In addition to the average error, we considered the number of correctly tracked frames. For correctly tracked frames, we assumed that the average error was correctly tracked when the current width and height of the object set by the person were smaller than the smaller value. 4 shows the numerical experimental results (average error, tracking success rate) of each algorithm, and the best results are based on the correctly tracked frame rate, with the darker letters and underlines, and the second best result is the darker ones Respectively.

Caviar, FaceOcc, Woman, CupOcc images are images for performance evaluation of occlusion. As shown in FIG. 4, the method of the present invention shows superior performance over all compared methods. It should be noted that in the case of Woman image, the FRAG was reported to perform very well in the Adam paper, but the performance was low when the entire image was not used. Also, in the case of CupOcc image, it is confirmed that the method of the present invention succeeds while the other algorithm fails because the object is obscured from the beginning. In the case of the Cheetah image, the object is largely invisible from the background. In this case, all of the methods except L1, OAB, MIL and DM of the present invention fail.

We used Sylvester, Tiger, Sylv_rep, and Tiger_rep images to evaluate the performance of drifting. The Sylvester and Tiger images are images with various changes. The images are repeatedly stretched back and forth so that drifting occurs for a long time. By comparing the results obtained by repeating the method of the present invention and the existing results, Can be confirmed. As a result of the experiment, it was confirmed that the method of the present invention did not cause extra drifting. In the case of SEMI and FRAG, drifting did not occur largely, but overall tracking performance deteriorated. In the case of OAB and MIL, the Sylvester image is good, but the Tiger shows large drift. The performance of TLD is improved for Tiger, but it is difficult to use it as a tracking result when checking actual experimental results. Finally, it was confirmed that the performance of L1 is severely degraded by drifting.

FIG. 5 is a reference diagram showing an experimental image and a main tracking result frame for each algorithm to explain the operation and effect of the present invention.

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 (11)

A first step of setting foreground models and background models using two individual single Gaussian models, respectively;
A second step of updating the background model by compensating a background motion based on a feature in an input image;
A change is made to the foreground model and then combined with the background model to reconstruct the most similar to the current observation value and a likelihood value is calculated from the foreground model and the background model by a likelihood value calculation in inverse proportion to the error at the time of reconstruction A third step of calculating;
A fourth step of repeatedly executing an average moving process using random sampling to obtain a tracking result for an object to be tracked by maximizing the likelihood;
A fifth step of updating the foreground model by selectively applying learning and discard learning according to whether the foreground model is reflected in the likelihood calculation or whether the background model is reflected in the likelihood calculation;
And a method for tracking an object through learning and discard learning.
delete The method according to claim 1,
The third step is to convert the vector representing the transform applied to the foreground model

), And the conversion (

) (

) ≪ / RTI >

, The conversion (

) Is applied to Equation

And

And a background model corresponding to the foreground model

Wow

To

Wow

In the tracking candidate region (

) Is cut out,

And

And the object is tracked by the dual modeling and learning and discarding learning.
The method of claim 3,
The third step is a step

) In the tracking candidate region (

) To zoom in and out

, The tracking candidate region (

Lt; RTI ID = 0.0 >

) Likelihood

For pixel index j

, And

And

Respectively,

And the object tracking method using learning and discard learning.
The method of claim 4,
Wherein the fifth step comprises:

And

And determining whether to apply learning or discard learning in updating the foreground model according to whether any of the models is utilized in the likelihood calculation.
The method of claim 5,
The fifth step is a step of updating the foreground model

Lt; / RTI >

And

Is the foreground model for pixel j at time t,

Is the zoomed-in observation for pixel j at time t,

Is set to the number of frames observed until time < RTI ID = 0.0 > t. ≪ / RTI >
The method of claim 6,
In the fifth step

Can be obtained from the age of the pixel, and the age at time t of the pixel j

If you say

The

, And

Is a value belonging to the range of [-1, 1] as an update parameter, and the learning and discard learning are coordinated.
The method of claim 7,
In the fifth step, pixels in which the foreground model is used

, Pixels in which the background model is used

, Then j ∈

In the case of

(Learning-learning), and j ∈

In the case of

(2) a method of tracking objects through learning and discarding learning.
The method of claim 8,
In the fifth step,

In case of using

Is in the range [1, ∞]

When defined as above,

Is again in the range [0, 1]

Wherein the object tracking method is based on dual modeling and learning and discard learning.
The method of claim 6,
The second step is to update the background model

, And

And

The

, And in the case of the background model j < RTI ID = 0.0 >

If

, And j ∈

If

And a method for tracking an object through learning and discard learning.
A computer-readable recording medium recording an object tracking program for executing an object tracking method through dual modeling and learning and discard learning according to any one of claims 1 and 3 to 10.

Images