KR100955044B1 - Apparatus and method for estimating a subpixel resolution - Google Patents

Abstract

In the optical flow estimation of the subpixel resolution according to the present invention, an optical flow vector of a pixel unit for an original image is estimated by receiving two temporally successive images and performing an iterative reliability transfer operation on an MRF of pixel resolution. The optical flow vector of the subpixel resolution of the subpixel resolution of the original image is estimated on the MRF of the subpixel resolution using the optical flow vector, and then the optical flow vector of the pixel resolution and the optical flow vector of the subpixel resolution are applied to the original image. Compute the final optical flow vector with sub pixel resolution for.

As described above, the present invention can significantly reduce the quantization error due to the limited number of displacement states that are common in the existing global methods, by measuring the optical flow up to the sub-pixel resolution unlike the conventional pixel-based optical flow.

Optical flow, subpixel, data cost

Description

Apparatus and method for optical flow estimation of subpixel resolution {APPARATUS AND METHOD FOR ESTIMATING A SUBPIXEL RESOLUTION}

The present invention relates to optical flow estimation, and more particularly, to an apparatus and a method for estimating an optical flow of sub-pixel resolution using an optical flow vector of pixel resolution.

Optical flow is a technique of estimating the temporal displacement vector of an image pixel by comparing two consecutive images in time, and is a computer vision technique widely used for motion estimation and image compression.

In order to estimate such an optical flow, a process of finding a correspondence point between two images is required. The problem of searching for a correspondence point is a kind of "ill-posed problem", which requires very high computational complexity. The problem of computational complexity in the optical flow estimation technique is represented by hardware high complexity and slow computation time in actual implementation. This is a major cause of difficulty in applying optical flow estimation techniques in applications where real-time processing is important, such as video surveillance or automotive vision systems.

For optical flow estimation, it is generally assumed that the brightness of the corresponding point of two consecutive images is constant. This assumes that corresponding points of the comparison image corresponding to the pixels of the reference image have the same brightness. Under the above assumptions, the difference in brightness of consecutive image pixels is purely caused by the relative movement of the object or camera.

This assumption reduces the complexity of the optical flow operation.

Existing correspondence points for optical flow estimation can be classified into two types, that is, a local method and a global method.

First, the local method is a technique of searching for a corresponding point by using information on a part of an area around an image pixel, and using a pixel mask-based technique or an image such as sum of absolute difference (SAD) or normalized cross correlation (NCC). A block matching technique for dividing into small blocks to find their correspondence is mainly used.

Representative local optical flow estimation techniques include "BDLucas and T.Kanade." An iterative image registration technique with an application to stereo vision. " Proceedings of Imaging understanding workshop , pp 121-130, 1981." and "BKPHorn, and BGSchunck," Determining optical flow. " Artificial Intelligence , vol 17, pp 185-203, 1981.". These techniques are characterized by relatively simple implementation and high computational speed.

Global methods for optical flow estimation are described in "J. Xiao and M. Shah." Motion layer extraction in the presence of occlusion using graph cuts. " IEEE Trans.on Pattern Analysis and Machine Intelligence , vol, 27, no. 10, pp. 1644-1659, 2005. ""PFFelzenswalb and DRHuttenlocher." Efficient belief propagation for early vision. " In Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition , no.1, pp.1261-1268, 2004." Described in academic papers.

This global method is a technique for estimating a globally optimized result using information of the entire image. For example, among the optimization algorithms based on the Markov Random Field (MRF) model, Graph cut and Belief propagation algorithms have low error rates and effectively cope with data discontinuities or obstructions, making them popular as next-generation technologies in computer vision, such as optical flow estimation and stereo vision. have.

In the conventional optical flow estimation method, it is assumed that the brightness of the corresponding point is constant due to a change in pose according to the movement of an object occurring in a continuous image, a change in shape, an illumination difference due to a shadow, and an occlusion phenomenon. Since a violation occurs, the estimated optical flow includes a large number of errors, and thus it is difficult to apply.

The local method among the corresponding point searching methods for optical flow estimation has a disadvantage in that an error due to an "aperture problem", an obstruction phenomenon, an obscure or periodic image pattern, etc. is severe.

Since the global method of the correspondence point search method for optical flow estimation covers all the information, not part of the image, even if the data to be estimated is increased little by little, the complexity of the image can be increased and only a limited number of discrete states can be estimated. This causes a large error in the tracking result of the optical flow.

Unlike conventional pixel-based optical flows, the present invention measures optical flows up to sub-pixel resolution, thereby greatly reducing quantization error due to a limited number of displacement states commonly found in conventional global methods.

In addition, the present invention effectively reduces the amount of computation greatly increased due to the subpixel resolution by using an efficient estimation method of the optical flow of the subpixel resolution.

In accordance with another aspect of the present invention, an apparatus for estimating an optical flow state at a subpixel resolution includes a data cost calculator configured to calculate a pixel cost and a data cost of a subpixel resolution of each node with respect to two consecutive input image data in time; A message operation unit that calculates a subpixel and a pixel resolution message between each node based on a pixel resolution data cost, and transfers the message to neighboring nodes; and after the repetition is finished, the data cost and message And a result output unit for estimating an optical flow state of a corresponding node based on the data cost calculator, the message calculator, and the result output unit, based on the optical flow state of the pixel resolution. Calculate, the calculated subpixel It based on the message and the data cost is characterized by estimating the optical flow state of the sub-pixel resolution.

The optical flow state estimation method of the subpixel resolution according to the present invention comprises the steps of estimating the optical flow vector per pixel for the original image by receiving two temporally consecutive images and performing an iterative reliability transfer operation on the MRF of the pixel resolution; Estimating the optical flow vector of the subpixel resolution of the subpixel resolution of the original image on the MRF of the subpixel resolution using the optical flow vector of the pixel resolution, and the optical flow vector of the pixel resolution and the subpixel Calculating a final optical flow vector having a sub-pixel resolution for the original image using the optical flow vector of the resolution.

In the present invention, since the global method is used, the error rate is reduced because the error due to the discontinuity caused by the obstruction or the boundary of the object can be reduced, and the complexity of the operation can be reduced by using parallel operation such as message passing. A system that can be minimized can be implemented.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. In addition, in describing the present invention, when it is determined that the detailed description of the related known configuration or function may obscure the gist of the present invention, the detailed description thereof will be omitted.

1 is a block diagram illustrating an optical flow estimation system according to an exemplary embodiment of the present invention, which includes a data cost calculator 100, a message operator 110, and a result outputter 120.

)에 대해 각 노드의 픽셀 해상도 및 서브 픽셀 해상도의 데이터 코스트를 계산하며, 계산된 데이터 코스트를 메시지 연산부(110)에 제공하는 수단으로서, 각 노드의 픽셀값을 기반으로 픽셀 해상도의 데이터 코스트를 계산하는 픽셀 단위 데이터 코스트 연 산기(102)와, 픽셀 해상도의 옵티컬 플로우 상태를 토대로 서브 픽셀 해상도의 데이터 코스트를 계산하는 서브 픽셀 단위 데이터 코스트 연산기(104)와, 픽셀 단위 및 서브 픽셀 단위 데이터 코스트 연산기(102, 104)의 출력 값을 저장하는 데이터 코스트 메모리(106)를 구비한다. 여기서, 데이터 코스트는 시간적으로 연속된 입력된 영상 데이터의 원본 픽셀과 비교 픽셀간 밝기 차이의 절대치이다.The data cost calculator 100 may input two input images successively in time.

Data cost of the pixel resolution and the sub pixel resolution of each node, and providing the calculated data cost to the message operation unit 110. The data cost of the pixel resolution is calculated based on the pixel value of each node. A pixel-by-pixel data cost calculator 102, a sub-pixel data cost calculator 104 that calculates a data cost of sub-pixel resolution based on an optical flow state of pixel resolution, and a pixel- and sub-pixel data cost calculator ( And a data cost memory 106 that stores the output values of 102 and 104. Here, the data cost is an absolute value of the brightness difference between the original pixel and the comparison pixel of the input image data which is continuous in time.

을 받아들여 픽셀의 위치 u에 해당하는 도 2에 도시된 MRF 상의 노드 p의 변위 상태

에 대해 아래의 수학식 1을 사용하여 데이터 코스트를 계산한다. The pixel-by-pixel data cost calculator 102 performs a temporally continuous input image.

The displacement state of node p on the MRF shown in FIG. 2 corresponding to position u of the pixel

Calculate the data cost using Equation 1 below.

) 데이터에서 원본 영상의 픽셀 밝기값과 비교 영상의 픽셀 밝기값간 차의 절대치를 계산하여 픽셀 해상도의 데이터 코스트를 계산한다.As can be seen from Equation 1 above, the pixel-based data cost calculator 102 uses two consecutive input images, as described in Equation 1 above.

) The data cost of the pixel resolution is calculated by calculating the absolute value of the difference between the pixel brightness value of the original image and the pixel brightness value of the comparison image.

위치, 즉 원본 픽셀 좌표와 픽셀 해상도의 옵티컬 플로우 벡터를 합친 위치에 있는 비교 영상의 주변 픽셀값에 보간법(Interpolation)을 적용하여 영상을 오버샘플링하는데, 이때 오버샘플링의 정도는 서브 픽셀 해상도에 따라 결정된다. 오버샘플링된 비교 영상의 픽셀값

와 원본 영상의 픽셀값

에 대해서 변위 상태

일 때 서브픽셀 데이터 코스트는 아래의 수학식 2를 이용하여 계산된다. 여기서, 신뢰도 전달 알고리즘은 반복적 연산을 통해 MRF 상의 최적 해를 찾아내는 기법이다.As shown in FIG. 3, the sub-pixel unit data cost calculator 104 has an optical flow state, that is, an optical flow vector of pixel resolution and a corresponding original pixel, on an MRF of pixel resolution output from the result output unit 120. An oversampler 300 that oversamples the surrounding pixels of the comparison image at the sum of the coordinates to calculate pixel brightness values, and an absolute calculator 302 that calculates an absolute value between the calculated pixel brightness values and the pixel brightness values of the original image; It consists of. That is, the sub-pixel unit data cost calculator 104 uses information between pixels that do not exist in the original image. To obtain the subpixel data cost for node p corresponding to pixel u of the original image, sum the estimation results of the reliability transfer algorithm at pixel resolution.

The image is oversampled by applying interpolation to the surrounding pixel values of the comparison image at the position, that is, the original pixel coordinates and the optical flow vector of pixel resolution, and the degree of oversampling is determined according to the subpixel resolution. do. Pixel value of the oversampled comparison image

And pixel values of the original image

Displacement state with respect to

The subpixel data cost is calculated using Equation 2 below. Here, the reliability transfer algorithm is a technique for finding the optimal solution on the MRF through iterative operation.

)와 노드 p의 옵티컬 플로우 벡터인 변위 상태(

)를 입력으로 하여 비교 영상의 픽셀값

을 출력한다. That is, the oversampler 300 of the subpixel data cost calculator 104 stores the comparison image (

) And the displacement state (which is the optical flow vector of node p)

) As the input pixel value of the comparison image

Outputs

The absolute value calculator 302 calculates the data cost of the subpixel by calculating an absolute value of the difference between the pixel brightness value output from the oversampler 300 and the pixel brightness value of the original image.

The data cost calculated by the pixel-based data cost calculator 102 and the sub-pixel data cost calculator 104 is stored in the data cost memory 106.

When the calculation of the data cost is finished, the message calculating unit 110 calculates a message value, which indicates probabilistic information including each displacement state of each node on the MRF and information of neighboring nodes, and delivers the message value to adjacent nodes. A pixel-by-pixel message operator 112 that calculates a message between each node on the MRF of pixel resolution and a sub-pixel-by-pixel message operator 114 that calculates a message between each node on the MRF of sub-pixel resolution after the message calculation of the pixel resolution is completed. It is provided.

에서 인접한 노드 q의 상태

로 전달되는 메시지를 평탄비용함수(smoothness cost)와 데이터 코스트를 통해서 계산한다. 픽셀 기반의 신뢰도 전달 알고리즘에서 메시지는 수학식 3을 통해서 계산된다.The message operation unit 110 is a state of the node p

Status of adjacent node q in

The message passed to is computed through the smoothness cost and the data cost. In the pixel-based reliability transfer algorithm, a message is calculated through Equation 3.

는 평탄 비용의 상한선이 되는 상수이며, 상태의 차이가 지나치게 클 경우 메시지 값이 지나치게 커지는 것을 방지하는 역할을 한다. Equation 4 refers to the flat cost function of the pixel-based reliability transfer algorithm. When each node has a discrete state, the flat cost function is expressed as the square of the difference of states.

Is a constant that is the upper limit of the flat cost, and prevents the message value from becoming too large when the state difference is too large.

As shown in FIG. 4, the sub-pixel message operator 114 includes an adder 400 for summing the data cost of each state of the node and the message of the previous step, and the MRF of pixel resolution to the message calculated at the sub-pixel resolution. Sums the flat cost function on each node to the result of the identification state assignment operator 402 and the dynamic state assignment operator 402 that assigns a message value to reflect the positional relationship between the node and the estimated optical flow vector of the neighboring node. And a message operator 404 for calculating the final message value of the subpixel resolution.

The sub-pixel message operator 114 of the message operator 110 is driven as the state estimation for the pixel resolution is completed, that is, the final message value of the sub-pixel resolution is calculated using the estimated state, and then the sub-pixel is performed based on the estimated state. Perform state estimation on the resolution.

의 데이터 코스트와 이전 단계에서 들어온 메시지 값을 아래의 수학식 5와 같이 합쳐준다. 여기서, 이전 단계에서 들어온 메시지 값은 메시지 값을 계산하고자 하는 노드의 주변 노드로부터 제공받는 메시지 값으로서, 예컨대 도 2의 q 노드에 대한 메시지 값을 계산한다는 과정에서 q 노드 가 전달하고자 하는 방향의 노드를 제외한 적어도 하나 이상의 이웃 노드로부터 전달받는 메시지 값을 의미한다.In this way, the state estimation for the sub-pixel resolution proceeds in three steps.

The data cost of and the message value from the previous step are combined as shown in Equation 5 below. Here, the message value received in the previous step is a message value provided from the neighboring node of the node for which the message value is to be calculated. A message value received from at least one neighbor node except for.

는 반복 연산에 의한 데이터의 오버플로우를 방지하는 정규화(normalization)를 위한 상수이다. 두 번째 단계에서는 첫 번째 단계에서 계산된 데이터 코스트와 메시지 합의 값의 영향을 픽셀 해상도에서 추정된 상태를 사용하여 계산한다. 즉, 두 노드 p와 q는 서로 다른 이산 상태

를 가지고 있기 때문에, 서브 픽셀 해상도에서 계산하는 메시지 값에 상기한 이산 상태의 차이를 반영해 줄 필요가 있다. 이는 수학식 6과 같은 형태로 수행된다.In Equation 5 above

Is a constant for normalization that prevents the overflow of data by iterative operations. In the second step, the influence of the data cost and the message sum value calculated in the first step is calculated using the state estimated at the pixel resolution. That is, two nodes p and q are different discrete states

Because of this, it is necessary to reflect the difference of the discrete state in the message value calculated at the subpixel resolution. This is performed in the form of Equation 6.

5A to 5C illustrate the above technique on the one-dimensional MRF. In the present invention, the above technique is called a dynamic state allocation technique.

In the third step, the final message value is calculated using Equation 7, Equation 8, using the value calculated in the second step and the flat cost function.

After the repetitive delivery and update of the message are all completed, the result output unit 120 estimates and outputs an optical flow state of pixel and sub pixel resolution based on the message value and data cost output from the message operation unit 110. . That is, the result output unit 120 outputs a pixel-by-pixel result output unit 122 that estimates an optical flow vector result in units of pixels using a data cost and a message value of each node in an MRF of pixel resolution, and an MRF of sub-pixel resolution. Subpixel result output 124 for estimating the optical flow vector results in subpixels using the data cost and message value of each node and the final optical flow vector results for each node using the pixel and subpixel results. The final output unit 126 for outputting the.

는 해당 노드의 데이터 코스트와 주변 노드의 메시지 값을 취합하여 수학식 9와 같이 추정될 수 있다. That is, the final state of each node p

May be estimated by Equation 9 by combining the data cost of the node and the message value of the neighboring node.

As such, the present invention uses a two-step iterative algorithm to estimate the state of the final sub-pixel resolution, i.e., as shown in FIGS. 6A and 6B, the time sequence is the same as the conventional reliability transfer algorithm. After estimating the displacement state in pixel units for the two image data, as shown in FIG. 6C, the obtained sub-pixel unit information is obtained using the estimated displacement state information in pixel units, and the estimated pixel unit displacement is obtained. Finally, the displacement state of the subpixel resolution is estimated using the state information and the information of the subpixel unit.

The operation of the optical flow estimating apparatus having the above configuration will be described with reference to FIG. 7.

7 is a flowchart illustrating an optical flow estimating process according to an exemplary embodiment of the present invention.

) 데이터가 입력(S700)됨에 따라 데이터 코스트 계산부(100)의 픽셀 단위 데이터 코스트 연산기(102)는 픽셀의 위치 u에 해당하는 MRF 상의 노드 p의 변위 상태(

)에 대한 데이터 코스트(

)를 계산(S702)한다. 이때, 데이터 코스트(

)는 상기의 수학식 1에 의거하여 산출될 수 있다. First, as shown in FIG. 7, a temporally continuous image (

As the data is input (S700), the pixel-based data cost calculator 102 of the data cost calculator 100 determines the displacement state of the node p on the MRF corresponding to the position u of the pixel.

Data cost for

) Is calculated (S702). In this case, the data cost (

) May be calculated based on Equation 1 above.

)를 이용하 여 메시지를 계산(S704)하는데, 즉 상기 수학식 4의 평탄화 비용 함수와 데이터 코스트(

)를 수학식 3에 적용하여 노드 p의 상태

에서 인접한 노드 q의 상태

로 전달되는 메시지를 계산한다.The pixel unit message operator 112 of the message operator 110 may calculate the data cost of the pixel resolution calculated by the pixel unit data cost calculator 102.

In step S704, a message is calculated using the flattening cost function and data cost (Equation 4).

) Is applied to equation (3) to state node p

Status of adjacent node q in

Calculate the message passed to.

When the message calculation process for each node on the MRF is finished (S706), the pixel-by-pixel result output unit 122 calculates an optical flow state estimation, that is, an optical flow vector for the pixel resolution (S708).

와 원본 영상의 픽셀값

에 대해서 변위 상태

에 대한 서브픽셀 데이터 코스트를 수학식 2를 이용하여 계산한다(S710).When the optical flow state estimation for the pixel resolution is completed, the sub-pixel data cost calculator 104 of the data cost calculator 100 calculates the peripheral pixel at the position of the sum of the optical flow vector calculated in S708 and the original pixel coordinates corresponding thereto. Pixel values of the oversampled image after oversampling the image by applying interpolation to the values

And pixel values of the original image

Displacement state with respect to

The subpixel data cost for is calculated using Equation 2 (S710).

When the calculation of the data cost of the sub pixel resolution is completed, the adder 300 of the sub pixel unit message operator 114 adds the calculated data cost and the message of the previous step (S712).

Then, the dynamic state assignment calculator 302 assigns a value to the data output from the adder 300 by reflecting the positional relationship between the estimated optical flow vector of the corresponding node and the neighboring node in the MRF of pixel resolution (S714). .

Thereafter, the message operator 304 calculates the final message value by applying a flat cost function between each node to the value output from the dynamic state assignment operator 302 (S716).

When the final message value for all nodes at the subpixel resolution is calculated (S718), the subpixel unit result output unit 124 may determine the subpixel unit for each node based on the data cost of the node and the final message value of the neighboring nodes. The optical flow is estimated (S720).

According to a preferred embodiment of the present invention, first, the displacement state of the pixel unit is estimated, and then information, such as data cost and message value, of the sub-pixel unit is calculated using this, and based on this, the displacement state of the sub-pixel unit is finally estimated. By doing so, the computation amount can be effectively reduced due to the sub pixel resolution.

It has been described so far limited to one embodiment of the present invention, it is obvious that the technology of the present invention can be easily modified by those skilled in the art. Such modified embodiments should be included in the technical spirit described in the claims of the present invention.

1 is a block diagram illustrating an optical flow estimating apparatus having a sub pixel resolution according to an exemplary embodiment of the present invention.

2 is a view showing an MRF structure applied to the present invention,

3 is a block diagram showing the structure of a data cost calculator according to the present invention;

4 is a block diagram showing the internal structure of a message operation unit according to the present invention;

5A through 5C are exemplary diagrams illustrating a dynamic state allocation scheme according to the present invention.

6A to 6C are diagrams for explaining a data cost calculation process and a displacement state estimation process according to the present invention;

7 is a flowchart illustrating an optical flow estimation process at a subpixel resolution according to an exemplary embodiment of the present invention.

<Description of the code | symbol about the principal part of drawing>

100: data cost calculator 110: message calculator

120: final output unit

Claims (14)

A data cost calculator for calculating the data cost of the pixel resolution and the subpixel resolution of each node with respect to two input image data which are consecutive in time; A message operation unit which repeatedly calculates a subpixel and pixel resolution message between nodes based on the data cost of the subpixel and pixel resolution, and transfers the message to neighboring nodes; A result output unit for estimating an optical flow state of a corresponding node based on the data cost and the message after the repetitive execution is completed; The data cost calculator, the message operator, and the result output unit calculate a data cost and a message of the sub-pixel resolution by using the optical flow state of the pixel resolution, and perform the sub based on the calculated message and data cost of the sub-pixel. And an optical flow estimating apparatus having a subpixel resolution. The method of claim 1, The data cost calculation unit, A pixel data cost calculator for calculating a data cost of pixel resolution based on the pixel value of each node; A subpixel data cost calculator for calculating a data cost of the subpixel resolution based on an optical flow state of the pixel resolution Optical flow estimation apparatus of the sub-pixel resolution comprising a. The method according to claim 1 or 2, And wherein the data cost is an absolute value of a difference in brightness between the original pixel and the comparison pixel of the continuous input image data. The method of claim 2, The subpixel data cost calculator, An oversampler that calculates a pixel brightness value by oversampling a pixel periphery of a comparison image at a position where an optical flow vector and corresponding original pixel coordinates are summed on the pixel resolution MRF; An absolute value calculator for calculating an absolute value of the difference between the calculated pixel brightness value and the pixel brightness value of the original image. Optical flow estimation apparatus of the sub-pixel resolution comprising a. The method of claim 2, The pixel unit data cost calculator may calculate a data cost of the pixel resolution by calculating an absolute value of a difference between a pixel brightness value of an original image and a pixel brightness value of a comparative image from the two consecutive input image data. Optical flow estimation device with pixel resolution. The method of claim 1, The message operation unit, A pixel-by-pixel message operator for calculating a message between each node on the pixel resolution MRF; Sub-pixel message operator for calculating a message between each node on the MRF of the sub-pixel resolution after the message resolution of the pixel resolution is completed Optical flow estimation apparatus of the sub-pixel resolution comprising a. The method of claim 6, The sub-pixel unit message operator, An adder for adding the data cost of any node to which the message is to be calculated and the message received from neighboring nodes except for the neighbor node to which the node is to deliver the message; A dynamic state allocation operator for allocating a message value to the message at the subpixel resolution by reflecting a positional relationship between the estimated optical flow vector of the node and the neighboring node in the MRF of the pixel resolution; A message operator that adds a flat cost function between each node to the result of the dynamic state allocation operator to produce a final message value of the subpixel resolution Optical flow estimation apparatus of the sub-pixel resolution comprising a. The method of claim 7, wherein The dynamic state allocation operator, An operator for calculating a difference between the optical flow vector estimated in the pixel unit of the corresponding node and the optical flow vector estimated in the pixel unit of the neighboring node; Allocation operator for dynamically allocating the message value according to the magnitude of the optical flow vector difference Optical flow estimation apparatus of the sub-pixel resolution comprising a. The method of claim 1, The result output unit, A pixel-by-pixel result output unit for estimating an optical flow vector result in pixel units using a data cost and a message value of each node in the pixel resolution MRF; A subpixel unit result output unit for estimating an optical flow vector result in units of subpixels using a data cost and a message value of each node in the MRF having the subpixel resolution; Final output unit for outputting the final optical flow vector results of each node using the pixel-by-pixel and sub-pixel results Optical flow estimation apparatus of the sub-pixel resolution comprising a. Estimating an optical flow vector in units of pixels for the original image by receiving two temporally consecutive images and performing an iterative reliability transfer operation on an MRF having a pixel resolution; Estimating an optical flow vector of the subpixel resolution of the subpixel resolution of the original image on an MRF of the subpixel resolution using the optical flow vector of the pixel resolution; Calculating a final optical flow vector having a subpixel resolution for the original image by using the optical flow vector of the pixel resolution and the optical flow vector of the subpixel resolution Optical flow estimation method of the sub-pixel resolution comprising a. The method of claim 10, Estimating the optical flow vector of the sub-pixel resolution, Calculating a data cost of subpixel resolution for each pixel of the original image; Calculating a message value of each node based on the calculated data cost; Estimating an optical flow of subpixel resolution for the node using the calculated message value and data cost Optical flow estimation method of the sub-pixel resolution comprising a. The method of claim 11, Computing the data cost of the sub pixel resolution, Oversampling the comparison image by applying an interpolation method to the coordinates of the original image corresponding to each node to the neighboring pixel values at the sum of the optical flow vector values of the pixel resolutions of the nodes; Calculating a data cost of the sub-pixel resolution by using the pixel value of the oversampled comparison image and the pixel value of the original image Optical flow estimation method of the sub-pixel resolution comprising a. The method of claim 12, And the data cost of the subpixel resolution is an absolute value of the difference between the pixel value of the oversampled comparison image and the pixel value of the original image. The method of claim 12, Computing the message value, Receiving a message value from a neighboring node except for a neighbor node to which a node wants to transmit a message value, and summing the received message value and the data cost of the node; Allocating a dynamic message value using a message of the arbitrary node and an optical flow vector difference in pixels of the neighboring node; Calculating a flat cost function indicating a state continuity between the arbitrary node and a neighbor node, and calculating a final message value using the calculated flat cost function and the dynamic message value Optical flow estimation method of the sub-pixel resolution comprising a.

Images