KR20130095985A - Apparatus and method for tracking object - Google Patents

Abstract

PURPOSE: A device and a method for tracking objects are provided to stably track an object in an image regardless of the fast and irregular movement of a photographing device. CONSTITUTION: A movement estimator (110) obtains a local affine movement parameter for indicating the movement variations of a photographing device from a current image frame and a previous image frame which is ahead of the current image frame on the inputted image and estimates a movement model matrix which comprises the local affine movement parameter. An object predictor (130) predicts the position of an object on the current image frame through an object state prediction model and corrects the predicted position of the object using the movement model matrix. An object tracking unit (150) tracks the object on the current image frame based on the corrected position of the object. [Reference numerals] (110) Movement estimator; (130) Object predictor; (150) Object tracking unit

Description

Apparatus and method for tracking object}

The present invention relates to an object tracking apparatus and method, and more particularly, to an apparatus and method for tracking an object in an image in consideration of the movement of the photographing apparatus.

Object tracking technology is used in a variety of computer vision applications such as video surveillance, user interface, augmented reality, intelligent space, object-based video compression, and driver assistance.

In order to improve the performance of the object tracking technology, researches are actively conducted to overcome factors that negatively affect the performance such as overlapping objects, complex backgrounds, changes in illuminance, shape changes of objects, image noise, and camera movement.

KR 10-0849499 (Sungkyunkwan University Industry-Academic Cooperation Group) 2008. 7. 24. Patent Document 1 is a method and apparatus for tracking an object using an active camera. A method of generating a subtraction image between images compensated for a background change, and performing object tracking using the generated subtraction image is disclosed. KR 10-2011-0092781 (Samsung Techwin Co., Ltd.) August 18, 2011 Patent document 2 is an image processing method and apparatus for motion tracking. Patent document 2 discloses a shake of an input image and detects a shake of the detected input image. Disclosed is to extract the motion data from the compensated input image, and to perform motion tracking based on the extracted motion data.

An object of the present invention is to provide an object tracking apparatus and method for tracking an object in an image in consideration of the movement of the photographing apparatus.

Another object of the present invention is to provide a computer readable recording medium having recorded thereon a program for executing a method of tracking an object in an image in consideration of the movement of a photographing apparatus.

The object tracking device according to the present invention for achieving the above technical problem, the degree of change in the movement of the photographing device for capturing the input image from the current image frame and the previous image frame in time ahead of the current image frame of the input image; A motion estimator for obtaining a local affine motion parameter, and estimating a motion model matrix of the obtained local affine motion parameters; Predict the position of the object in the current image frame through an object state prediction model for predicting the state of the object in the current image frame from the state of the object in the previous image frame, and determine the position of the predicted object. An object predictor for correcting through a motion model matrix; And an object tracker that tracks an object in the current image frame based on the position of the object corrected by the object predictor.

The object tracking method according to the present invention for achieving the above technical problem, the degree of change in the movement of the photographing apparatus for capturing the input image from the current image frame and the previous image frame in time preceding the current image frame of the input image; Obtaining a representative local affine motion parameter; Estimating a motion model matrix of the obtained local affine motion parameters; Predicting a position of the object in the current image frame through an object state prediction model that predicts the state of the object in the current image frame from the state of the object in the previous image frame; Correcting the position of the predicted object through the motion model matrix; And tracking the object in the current image frame based on the corrected position of the object.

According to another aspect of the present invention, there is provided a computer readable medium storing a program for causing a computer to execute any one of the above methods.

According to the object tracking apparatus and method according to the present invention, even if there is a rapid movement or irregular movement of the photographing apparatus it is possible to stably track the object in the image. In addition, even when there is movement of a photographing device having a jello effect, the object may be continuously tracked in the image.

1 is a block diagram illustrating an object tracking apparatus according to a preferred embodiment of the present invention;
2 and 3 are views showing an example of actually applying a motion model matrix to an image according to a preferred embodiment of the present invention;
4 is a diagram illustrating an example of object tracking after correcting an object state using a motion model matrix according to an embodiment of the present invention;
5 is a flowchart illustrating an object tracking method according to a preferred embodiment of the present invention, and
6 to 9 are diagrams for explaining the performance of the object tracking operation according to a preferred embodiment of the present invention.

Hereinafter, exemplary embodiments of an object tracking apparatus and method according to the present invention will be described in detail with reference to the accompanying drawings.

1 is a block diagram illustrating an object tracking apparatus according to a preferred embodiment of the present invention.

Referring to FIG. 1, the object tracking apparatus 100 tracks an object in an image in consideration of a movement of a photographing apparatus (not shown). To this end, the object tracking apparatus 100 includes a motion estimator 110, an object predictor 130, and an object tracker 150. Meanwhile, the object tracking device 100 may be integrated with a photographing device or may receive an image from another device such as a photographing device or a storage device (not shown).

The motion estimator 110 estimates a motion model matrix indicating the degree of motion of the photographing apparatus that captured the input image from the input image by using an elastic registration (ER) algorithm.

Here, an elastic registration (ER) algorithm is used to measure the degree of motion between two images taken in succession in time. The elastic registration (ER) algorithm obtains local affine motion, contrast change, and brightness change from two images that are continuous in time. Can be. That is, in addition to the local affine motion, the intensity variation of the contrast parameter and the brightness parameter is also considered.

On the other hand, the two consecutive video frames of the input video are shot in a scene with a short time difference, so the light change between the two video frames is small, so that the local affine except the contrast change degree is used. Modified elastic registration (ER) algorithms may be used to obtain a degree of local affine motion and brightness variations. Hereinafter, the present invention will be described using a modified elastic registration (ER) algorithm. Of course, an elastic registration (ER) algorithm may be applied to the present invention.

The affine transform is represented by Equation 1 below.

는 움직임 모델 행렬, 즉, 어파인 카메라 모델 행렬(afiine camera model matrix)을 나타내고,

,

,

및

는 선형 어파인 파라미터(linear affine parameter)를 나타내며,

및

는 트랜스레이션 파라미터(translation parameter)를 나타낸다. 그리고, 로컬 어파인 움직임 파라미터(local affine motion parameter)는 선형 어파인 파라미터(linear affine parameter)와 트랜스레이션 파라미터(translation parameter)를 말한다.here,

Denotes a motion model matrix, i.e., an afiine camera model matrix,

,

,

And

Represents a linear affine parameter,

And

Denotes a translation parameter. The local affine motion parameter refers to a linear affine parameter and a translation parameter.

The modified elastic registration (ER) algorithm is expressed as Equation 2 below.

가 입력 영상 중 현재 영상 프레임을 나타내고,

는 현재 영상 프레임에 시간적으로 앞서 있는 이전 영상 프레임을 나타내며,

는 밝기 파라미터(brightness parameter)를 나타낸다.here,

Represents the current video frame of the input image,

Indicates a previous video frame that is temporally ahead of the current video frame.

Denotes a brightness parameter.

,

,

및

와 트랜스레이션 파라미터(translation parameter)

및

와 밝기 파라미터(brightness parameter)

를 추정하기 위해, 다음의 [수학식 3]과 같은 수정된 에러 함수(error function)를 최소화한다.Linear affine parameter from two image frames that are continuous in time with each other

,

,

And

And translation parameters

And

And brightness parameters

In order to estimate, minimize the modified error function (Equation 3) below.

는

를 나타내고,

는 입력 영상을 나타낸다.here,

The

Lt; / RTI >

Indicates an input image.

의 비선형 함수(nonlinear function)이므로 최적의 답이 존재하지 않는다. 이에 따라, 이 에러 함수를 최소화하는 벡터

를 찾기 위해 뉴턴(Newton-Raphson) 알고리즘과 같은 반복적인 접근(iterative approach) 방법을 이용한다. 그러므로, 에러 함수

를 최소화하는 대신에 일차 단락 테일러 시리즈(first-order truncated Taylor series)을 이용하여 에러 함수

를 다음의 [수학식 4]와 같이 근사화한다.However, this error function

Since it is a nonlinear function of, there is no optimal answer. Thus, the vector that minimizes this error function

We use an iterative approach such as the Newton-Raphson algorithm. Therefore, the error function

Error function using first-order truncated Taylor series instead of minimizing

Is approximated by Equation 4 below.

는

의 수평 방향(horizontal direction)으로의 공간적 미분(spatial derivative)을 나타내고,

는

의 수직 방향(vertical direction)으로의 공간적 미분(spatial derivative)을 나타내며,

는

의 시간적 미분(temporal derivative)을 나타낸다.here,

The

Represents a spatial derivative in the horizontal direction of

The

Represents the spatial derivative in the vertical direction of

The

Represents the temporal derivative of.

Simplifying [Equation 4], the approximated error function is expressed by the following [Equation 5].

는

를 나타내고,

는

를 나타낸다. 근사화된 에러 함수

는

의 이차 함수(quadratic function)임으로 이 에러 함수를 최소화하는 벡터

는

에 대하여 미분하고 그 결과를 0으로 설정하는 것에 의해 찾을 수 있다. 여기서, 그 결과는

를 구하는데 이용되는 다음의 [수학식 6]과 같은 선형 방정식(linear equation)의 집합(set)이다.here,

The

Lt; / RTI >

The

Indicates. Approximate error function

The

A vector that minimizes this error function by being a quadratic function of

The

Can be found by differentiating against and setting the result to zero. Where the result is

It is a set of linear equations as shown in Equation 6 below.

는

를 나타내고 7x7 행렬이다. 행렬

의 역(inverse)이 존재하는 보장은 없지만, 입력 영상

가 충분히 크고 충분한 콘텐츠(content)가 있다면, 행렬

의 역(inverse)은 일반적으로 존재한다.here,

The

Represents a 7x7 matrix. procession

There is no guarantee that inverses exist, but the input image

Is large enough and has enough content, the matrix

The inverse of is usually present.

2 and 3 are diagrams showing an example of actually applying a motion model matrix to an image according to a preferred embodiment of the present invention.

2 illustrates a result of affine transformation of an input image using the estimated camera movement. That is, FIG. 2A shows the current image among the input images, FIG. 2B shows the previous image among the input images, and FIG. 2C shows the affine transformation of the current image. Results (d) of FIG. 2 show a difference image between (a) and (b), and (e) of FIG. 2 shows a difference image between (b) and (c). ).

FIG. 3 shows a result of affine transformation of an input image using the estimated motion along with a jello effect caused by a rolling shutter. That is, FIG. 3A illustrates the current image among the input images, FIG. 3B illustrates the previous image among the input images, and FIG. 3C illustrates affine transformation of the current image. 3D shows a difference image between (a) and (b), and FIG. 3E shows a difference image between (b) and (c). ).

As described above, it can be seen that the difference between the current image and the previous image is large, whereas the difference between the image obtained by the affine transformation of the current image and the previous image is small.

The object predictor 130 corrects the position of the object predicted through the object state prediction model with respect to the input image by using the motion model matrix estimated by the motion estimator 110. That is, the object predictor 130 uses the motion model matrix, i.e., the afiine camera model matrix, to calculate the position of the object predicted through the object state prediction model through Equation 7 below. Correct.

는 보정된 객체의 위치를 나타내고,

는 보정전 객체의 위치를 나타낸다.here,

Indicates the position of the calibrated object,

Indicates the position of the object before correction.

에 따른 측정 또는 관측 모델

을 통해 시간

에서의 객체

의 상태(state) 추정을 반복적으로 갱신한다. 상태 벡터(state vector)는 객체의 2D 또는 3D 위치

, 객체의 크기(scale)

등과 같은 복수의 상태로 구성된다. 시간

까지 측정된 모든 주어진 시간

에서의 상태(state)의 확률 밀도 함수(probability density function)

를 알고 있다면, 시간

에서의 상태(state)는 다음의 [수학식 8]과 같은 채프먼-콜모고로프 방정식(Chapman-Kolmogorov equation)을 이용하여 예측된다.In more detail, the Bayesian approach to tracking objects is given a given time.

Measurement or observation model

Through time

The object at

Iteratively updates the state estimate of. A state vector is a 2D or 3D position of an object

, The scale of the object

It consists of several states, such as these. time

All given time measured up to

Probability density function of states in

If you know, time

The state at is predicted using the Chapman-Kolmogorov equation as shown in Equation 8 below.

는 사전 밀도(prior density)를 나타낸다. 그러면, 시간

에서의 주어진 새로운 측정 모델

는 베이즈 법칙(Bayes' rule)을 이용하여 다음의 [수학식 9]와 같이 갱신된다.here,

Denotes the prior density. Then time

Given new measurement model

Is updated using Equation (9) using Bayes' rule.

는 사후 밀도(posterior density)를 나타낸다.

는 관측 모델을 나타낸다. [수학식 8] 및 [수학식 9]는 최적의 베이스 해(optimum Bayes' solution)을 기초로 형성된다. 일반적으로 이러한 밀도들의 재귀적인 전파(recursive propagation)는 다루기 힘들다. 이에 따라, 이 해를 근사화기 위해 다른 방법들이 이용된다. 이러한 다른 방법들 중의 하나가 순차적인 중요 샘플링(sequential importance sampling)을 사용하는 몬테 카를로(Monte Carlo : MC) 방법과 같은 파티클 필터(particle filter)이다. 파티클 필터(particle filter)에서 사후 밀도

는 중요 가중치(importance weight)

를 가지는

개의 샘플들

의 유한 집합(finite set)에 의해 근사화된다. 후보 샘플들(candidate samples)

은 중요 분포(importance distribution)

로부터 샘플링에 의해 독립적으로 선택되고, 후보 샘플들의 가중치는 다음의 [수학식 10]과 같다.here,

Denotes the posterior density.

Represents an observation model. Equations 8 and 9 are formed based on an optimal Bayes' solution. In general, recursive propagation of these densities is difficult to handle. Accordingly, other methods are used to approximate this solution. One of these other methods is a particle filter, such as the Monte Carlo (MC) method, which uses sequential importance sampling. Post Density in Particle Filters

Is the import weight

Having

Samples

Approximated by a finite set of. Candidate samples

Is an import distribution

Are independently selected by sampling, and the weights of candidate samples are given by Equation 10 below.

는

과 동일하고 가중치는 관측 모델

가 된다. 파티클 필터(particle filter)의 기본 아이디어는 연관된 가중치(associated weight)를 가지는 랜덤 샘플들의 집합에 의해 사후 밀도(posterior density)로 대표되는 것이고, 이러한 가중치들과 샘플들을 이용하여 예상된 값(expected value)과 같은 이러한 상태들의 추정을 계산하는 것이다.Candidate samples are resampled to produce an unweighted particle set according to their import weight to avoid degeneracy. For bootstrap filter,

The

Is the same as

. The basic idea of a particle filter is represented by the posterior density by a set of random samples with associated weights, and the expected value using these weights and samples. Is to calculate an estimate of these states.

는 이계 자기 회귀(second-order autoregressive process)와 가우시안 함수(gaussian function)에 의해 정의되는 잡음 모델(noise model)을 이용한다. 관측 모델은 카메라 움직임에서 안정적인 조명 변화를 위해 음영 효과(shading effect)로부터 색채 정보(chromatic information)를 분리시키는 색상-채도-명도 컬러 히스토그램(Hue-Saturation-Value color histogram)으로 활용된다. 색상-채도-명도 컬러 히스토그램(HSV histogram)은

빈스(bins)로 구성되고,

는 시간

의 위치

에서 빈 인덱스(bin index)로 정의된다.Prediction model

Uses a noise model defined by a second-order autoregressive process and a Gaussian function. The observation model is used as a Hue-Saturation-Value color histogram that separates the chromatic information from the shading effect for stable lighting changes in camera movement. The color-saturation-brightness color histogram (HSV histogram)

Consisting of bins,

Time

Location of

Is defined as the bin index in.

는 다음의 [수학식 11]과 같은 시간

에서의

를 나타낸다.Color distribution

Is the same time as [Equation 11]

In

Indicates.

는 크로네커 델타 함수(kronecker delta function)를 나타내고,

는 정규화 상수(normalization constant)를 나타내며,

는 샘플링된 객체 영역을 나타낸다. 관측 모델과 참조 모델 사이의 유사도 측정을 위해 바타챠랴 거리 측정(Bhattacharyya distance measurement)

를 사용하고 다음의 [수학식 12]와 같이 정의된다.here,

Represents the kronecker delta function,

Represents a normalization constant,

Represents the sampled object area. Bhattacharyya distance measurement to measure the similarity between the observation model and the reference model

It is defined as in Equation 12 below.

는 참조 컬러 모델(reference color model)을 나타낸다. 참조 분포(reference distribution)는 초기 시간

에서 획득된다.here,

Denotes a reference color model. Reference distribution is the initial time

Lt; / RTI >

The similarity measure as an observation model is defined by the Bhattacharyya distance as shown in Equation 13 below.

에서의 상태(state)의 최대 사후(maximum a posteriori : MAP) 추정은 다음의 [수학식 14]와 같이

개의 파티클들(샘플들)로부터 얻어진다.For example, time

The maximum a posteriori (MAP) estimation of the state in is given by Equation 14 below.

From the two particles (samples).

As the number of particles increases, the particle filter approaches an optimal Bayesian estimate. The sampling method is used to generate particles comprising sequential Monte Carlo (Market Chain Monte Carlo (MCMC) and Wang-Landau MC).

If the movement of the object is smooth, the particle filter works well when tracking the object. However, in the case of irregular object movement caused by sudden movement of the object or camera movement, it is difficult to accurately predict the state of the object, and particle filters tend to fail tracking.

Accordingly, the present invention uses another approach to correct the object state prediction model using the estimated motion model matrix. Equation 8 above represents the predicted state of the object and the previous state of the object using the prediction model. The sampling method implemented in the object state where the position of the object is corrected can effectively track the object even under camera movement.

4 is a diagram illustrating an example of object tracking after correcting an object state using a motion model matrix according to an exemplary embodiment of the present invention.

4 shows the result of the corrected particles. That is, FIG. 4A illustrates the current image among the input images, FIG. 4B illustrates the previous image among the input images, and FIG. 4C illustrates the corrected particles. Here, blue represents original particles, and purple represents corrected particles. As such, as a result of comparing the corrected particle with the original particle, it can be seen that the corrected particle is closer to the object than the original particle.

The object tracker 150 tracks the object in the input image based on the object state prediction model corrected by the object predictor 130.

5 is a flowchart illustrating an object tracking method according to a preferred embodiment of the present invention.

The object tracking device 100 uses a motion registration matrix, that is, an affine camera model matrix, indicating a degree of motion of a photographing device that captures an input image from an input image using an elastic registration (ER) algorithm. ) Is estimated (S510). That is, the object tracking apparatus 100 obtains a linear affine parameter, a translation parameter, and a brightness parameter from two image frames that are continuous in time with each other, and obtains the obtained linear A motion model matrix consisting of a linear affine parameter and a translation parameter is estimated.

Thereafter, the object tracking apparatus 100 corrects the position of the object predicted through the object state prediction model with respect to the input image using the motion model matrix (S530). Then, the object tracking apparatus 100 tracks the object in the input image based on the object whose position is corrected (S550).

6 to 9 are diagrams for explaining the performance of the object tracking operation according to a preferred embodiment of the present invention.

에서 초기 객체 영역은 수동적으로 선택된다. 본 발명의 성능 평가에 사용되는 파라미터

,

및

은 각각 10으로 설정되고

는 20으로 설정된다.The performance of the object tracking operation according to the present invention is tested using an image captured by a handheld device such as a smart phone equipped with a camera. Here, the captured image is 29 frames per second and the resolution is 640x480. The object to be photographed is outdoor as shown in FIG. 8 and a fish as shown in FIG. 6. In particular, in the open air, fast and irregular movements are generated by photographing surrounding scenes while a person walks with the camera. Conventional particle filters and Markov Chain Monte Carlo (MCMC) methods are used to compare the performance of the present invention. 300 particles are used for particle filter, initial time

At initial object area is selected manually. Parameters used in the performance evaluation of the present invention

,

And

Are set to 10 each

Is set to 20.

To assess the performance of the present invention, the Euclidian distance between a manually-labeled ground truth position and an estimated center position of the object relative to the center of the object Is calculated by the following [Equation 15].

및

는 각각 객체 영역의 추정된 센터(center)의 수평 좌표(horizontal coordinate) 및 수직 좌표(vertical coordinate)를 나타내고,

및

는 각각 그라운드 참 센터(ground truth center)의 수평 좌표(horizontal coordinate) 및 수직 좌표(vertical coordinate)를 나타낸다.here,

And

Denotes the horizontal and vertical coordinates of the estimated center of the object area, respectively,

And

Denotes the horizontal and vertical coordinates of the ground truth center, respectively.

6 shows an experimental result of tracking an object in an image of a fish. That is, FIG. 6 (a) shows an experimental result of tracking an object in an image using a conventional particle filter, and FIG. 6 (b) shows Markov Chain Monte Carlo: An experiment result of tracking an object in an image using the MCMC method is shown, and FIG. 6 (c) shows an image using a particle filter based on affine camera motion according to the present invention. 6 shows an experimental result of tracking an object, and FIG. 6 (d) shows an image using an Markov Chain Monte Carlo (MCMC) method based on affine camera motion according to the present invention. Shows the results of experiments that tracked objects. Here, the green part represents the object area.

In the conventional method using a conventional particle filter, as shown in (a) of FIG. 6, it can be seen that the estimation of the object region is incorrect due to camera movement. Referring to FIG. 7, it can be seen that the conventional particle filter method (①) fails to track an object after the 321 th frame. In addition, it can be seen that the Markov Chain Monte Carlo (MCMC) method (2) may miss an object in general. On the other hand, the particle filter method (③) or Markov Chain Monte Carlo (MCMC) method (④) based on the affine camera motion according to the present invention is used for object tracking. You can see that the overall performance is stable. That is, the object tracking according to the present invention enables continuous object tracking even with irregular camera movement.

8 shows an experiment result of tracking an object in an outdoor image. That is, FIG. 8 (a) shows an experimental result of tracking an object in an image using a conventional particle filter, and FIG. 8 (b) shows Markov Chain Monte Carlo: An experiment result of tracking an object in an image using the MCMC method is shown, and FIG. 8 (c) shows an image using an particle camera based on an affine camera motion according to the present invention. An experiment result of tracking an object is shown, and FIG. 8 (d) shows an image using an Markov Chain Monte Carlo (MCMC) method based on affine camera motion according to the present invention. Shows the results of experiments that tracked objects. Here, the green part represents the object area.

Conventional particle filter method and Markov Chain Monte Carlo (MCMC) method is a camera having a jello effect as shown in (a) and (b) of Figure 8 It can be seen that because of the movement, the object is missing or the estimation of the object area is inaccurate. Referring to FIG. 9, it can be seen that the conventional particle filter method (①) fails to track an object after the 81 st frame, and the Markov Chain Monte Carlo (MCMC) method ( ②) shows that object tracking fails after the 121st frame. On the other hand, the particle filter method (③) or Markov Chain Monte Carlo (MCMC) method (④) based on the affine camera motion according to the present invention is a gel phenomenon ( Continuous object tracking is possible even with camera movement with jello effect.

The present invention can also be embodied as computer-readable codes on a computer-readable recording medium. A computer-readable recording medium includes all kinds of recording apparatuses in which data that can be read by a computer apparatus 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 device, and the like in the form of a carrier wave . In addition, the computer-readable recording medium may be distributed to computer devices connected to a wired / wireless communication network, and a computer-readable code may be stored and executed in a distributed manner.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is to be understood that the invention is not limited to the disclosed exemplary embodiments, but, on the contrary, It will be understood by those skilled in the art that various changes may be made and equivalents may be substituted without departing from the scope of the appended claims.

100: object tracking device, 110: motion estimation unit,
130: object prediction unit, 150: object tracking unit

Claims (7)

Acquiring a local affine motion parameter representing a degree of change in the motion of the photographing apparatus which captured the input image from the current image frame and the previous image frame temporally preceding the current image frame among the input images, A motion estimator for estimating a motion model matrix including the obtained local affine motion parameters;
Predict the position of the object in the current image frame through an object state prediction model for predicting the state of the object in the current image frame from the state of the object in the previous image frame, and determine the position of the predicted object. An object predictor for correcting through a motion model matrix; And
And an object tracker for tracking an object in the current image frame based on the position of the object corrected by the object predictor. The method of claim 1,
The motion estimator is a motion model matrix as shown in Equation A below.

An object tracking device, characterized in that for estimating:
[Mathematical formula A]

Here,

, remind

, remind

And

Denotes a linear affine parameter,

And

Denotes a translation parameter, and the local affine motion parameter refers to the linear affine parameter and the translation parameter. The method of claim 2,
The object predicting unit corrects the position of the object through the following Equation B, wherein the object tracking device:
[Mathematical expression B]

Here,

Denotes the position of the corrected object,

Indicates the position of the object before correction. Acquiring a local affine motion parameter indicating a degree of change in motion of a photographing apparatus that has captured the input image from a current image frame among the input images and a previous image frame temporally preceding the current image frame;
Estimating a motion model matrix of the obtained local affine motion parameters;
Predicting a position of the object in the current image frame through an object state prediction model that predicts the state of the object in the current image frame from the state of the object in the previous image frame;
Correcting the position of the predicted object through the motion model matrix; And
Tracking an object in the current image frame based on the corrected position of the object. 5. The method of claim 4,
In the motion model matrix estimation step, a motion model matrix as shown in Equation A below.

An object tracking method comprising estimating a:
[Mathematical formula A]

Here,

, remind

, remind

And

Denotes a linear affine parameter,

And

Denotes a translation parameter, and the local affine motion parameter refers to the linear affine parameter and the translation parameter. 6. The method of claim 5,
In the object position correction step, the object tracking method characterized in that for correcting the position of the object through the following [Equation B]:
[Mathematical expression B]

Here,

Denotes the position of the corrected object,

Indicates the position of the object before correction. A computer-readable recording medium having recorded thereon a program for executing the object tracking method according to any one of claims 4 to 6.

Images