JP2003085566A - Correspondent point searching method and matching device using the same - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a technology capable of reducing an error and a search error and performing a smooth correspondent point search. SOLUTION: This method of searching for correspondent point between the images specified in the other picked up image corresponding to every pixel of one picked up image of a pair of picked up images, comprises the step of specifying scanning lines in one picked up image, the step of setting a parallax vector value included within the search range of the other picked up image as a node of a state transition model, the step of setting a matching error in a block set previously and continuousness of vector between adjacent nodes as weight of pass between the nodes of the state transition model, and the step of presuming the correspondent point of each pixel of one scanning line of the other picked up image in block by Viterbi algorithm on the basis of the state transition model.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a corresponding point search method and a matching device using the same, and more particularly to a technique effectively applied to a corresponding point search method for stereo images and a motion vector search method for moving images. Is.

[0002]

2. Description of the Related Art The most basic method for searching for corresponding points between images is a block matching method, which has been put to practical use in motion vector detection applications in coding systems such as MPEG. Alternatively, it is also used in a distance measuring system using a stereo image, and some systems are commercially available.

Although this method has a simple algorithm,
If there are few errors due to noise or textures that can be used as clues for searching for corresponding points, there is a drawback that errors easily occur.

In order to remedy this drawback, "Park Jouichi, Inoue Seiki," Depth extraction using a multi-lens camera for image synthesis ", IEICE TECHNOLOGY, PRMU96-133, pp.33.
-40, 1997-01. There is a hierarchical block matching method capable of searching corresponding points with less failure by generating reduced images hierarchically and performing block matching sequentially from low resolution images. Also,
"W. Woo, N. Kimand Y. Iwadat
e, "Object Segmentation for
Z-keying Using Stereo Im
ages, "inProc. IEEE WCC-ICS
P'00, pp. 1249-1254, Aug. 200
0. , There is a research example of a method for searching for corresponding points in consideration of continuity with adjacent pixels. Also, "Yu
ichi OHTA, Takeo KANADE, “St
ereo by Intra-and Inter-S
cannline Search Using Dyna
mic Programming ”, IEEE Tr
a. On P. A. M. l. , Vol. PAM-7, N
o. 2, pp. 139-154, March 198
5. , There is also a research example on the corresponding point search of a stereo image using the Viterbi algorithm.
In this method, a state transition model in which an edge position on one scanning line in an image is used as a node is constructed, and a cost function having a distance from a neighboring edge as a parameter is used.

[0005]

The present inventor has found the following problems as a result of examining the above-mentioned prior art. It is known that the hierarchical block matching method, which is one of the conventional methods, can reduce the effects of noise and errors such as errors that occur in a portion having a small texture, as compared with a simple block matching method. However, the performance varies depending on the number of layers and the block size, and in some cases, a sufficient effect may not always be obtained.

Further, the method considering the continuity with the adjacent pixels has a problem that a sufficiently smooth corresponding point cannot be obtained because only the constraint condition with the adjacent pixels is used.

Further, in the conventional method using the Viterbi algorithm, although a high resolution parallax vector is obtained, there is a problem that an error occurs due to noise because only edge position information is used.

The object of the present invention is to reduce errors and search errors,
It is to provide a technique capable of smoothly searching for corresponding points.

The above and other objects and novel features of the present invention will be apparent from the description of this specification and the accompanying drawings.

[0010]

Among the inventions disclosed in the present application, a brief description will be given to the outline of typical ones.
It is as follows.

(1) In a corresponding point search method between images for identifying a pixel-to-pixel correspondence destination of one of the pair of captured images in the other captured image, a scanning line in the one captured image A step of specifying a parallax vector value included in the search range of the other captured image as a node of the state transition model, a matching error in a preset block and a vector continuity between adjacent nodes. The weight of the path between the nodes of the state transition model is set, and the Viterbi algorithm is used to collectively estimate the corresponding points of each pixel on one scan line of the other captured image based on the state transition model. And steps.

(2) In the method of searching for corresponding points between frames for specifying the correspondence destination for each pixel between consecutive frames of a moving image, the step of specifying the scanning line in the preceding moving image of the moving images And a step of setting a motion vector value included in a search range in a two-dimensional direction of a subsequent frame continuous to the previous moving image as a node of a state transition model, a matching error in a preset block and an adjacent node Setting the continuity of vectors between nodes as the weight of the path between the nodes of the state transition model, and determining the corresponding point of each pixel on one scan line of the subsequent frame by the Viterbi algorithm based on the state transition model. Collectively estimating.

(3) In a matching device that specifies a correspondence destination for each pixel of one captured image of a pair of captured images in the other captured image, a pixel on one scanning line of the one captured image and the one A means for calculating a disparity vector value from the pixels in the search range of the other captured image corresponding to the scanning line, and the disparity vector value is set as a node of the state transition model, and a matching error within a preset block and Set the continuity of the vector between adjacent nodes as the weight of the path between the nodes of the state transition model,
A means for collectively estimating the corresponding points of each pixel on one scanning line of the other captured image based on the state transition model by the Viterbi algorithm is provided.

According to the above-mentioned means, the scanning line in one captured image is specified, the parallax vector value included in the search range of the parallax vector is set as the node of the state transition model, and the matching error and the adjacent node The continuity of the vector is set as the weight of the path between the nodes of the state transition model, and the Viterbi algorithm is used to collectively estimate the corresponding points of each pixel on one scanning line of the other captured image based on this state transition model. Therefore, it is possible to obtain a smooth parallax vector with few errors.

Further, in the case of searching for a corresponding point between frames for specifying a corresponding destination for each pixel between consecutive frames of a moving image, a scanning line in the previous frame of the moving image is specified, A motion vector value included in a search range in a two-dimensional direction of a moving image subsequent to the previous frame is set as a node of a state transition model, and a matching error in a preset block and a distance between adjacent nodes (continuous frames are set). The continuity of the vector in (between) is set as the weight of the path between the nodes of the state transition model, and the Viterbi algorithm is used based on this state transition model to group the corresponding points of each pixel on one scanning line of the subsequent moving image all together. Therefore, the motion vector of the moving image that is smooth and has few errors can be obtained.

[0016]

BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, the present invention will be described in detail with reference to the drawings together with an embodiment (example) of the invention. In all the drawings for explaining the embodiments of the invention, components having the same function are designated by the same reference numeral, and the repeated description thereof will be omitted.

(Embodiment 1) FIG. 1 is a diagram for explaining a method of searching for corresponding points between images according to Embodiment 1 of the present invention. In particular, state transition for applying the Viterbi algorithm will be described. FIG. However, the state transition for applying the well-known Viterbi algorithm to a pair of images is shown.

As shown in FIG. 1, in the first embodiment, the image starts at the left end side (indicated by 0 (zero)) and reaches the right end side (N).
-1)). Therefore,
For example, when a well-known television camera is used as the image collecting means, the scanning line starts at the left end and ends at the right end.

In this case, at each pixel position, there is a column consisting of nodes 101 corresponding to the search range 0 to S-1, and these node columns are connected by a path 102 shown by an arrow in FIG. . This path 102 is weighted by the matching error energy component indicating the degree of correspondence and the energy component of the parallax vector field indicating the smoothness of the parallax vector. At this time, as is clear from FIG.
There are S ^N combinations of 02 when the number of horizontal pixels is N, but since the Viterbi algorithm is applied in the first embodiment, this pass 1
From the combination of 02, it is possible to efficiently estimate the optimum path 102 with a small matching error and a smooth disparity vector. At this time, the disparity vector of each pixel is obtained as the node number through which the optimum path passes. However, in the first embodiment, when estimating the optimum path by the Viterbi algorithm, as described above, as shown in FIG.
The calculation is started from the left end of, and the surviving path is obtained at each node 101 at each pixel position.

For example, in the case of two stereo images taken by moving the camera in parallel as shown in FIG. 2, the corresponding points of the pixels on a predetermined scanning line of the R image must be on the same scanning line of the L image. Exists in.

Therefore, in the corresponding point searching method of the first embodiment, the state transition in which the disparity vector value for each pixel on one scanning line is the node 101 is configured as shown in FIG. Disparity vectors corresponding to pixels are collectively obtained using the Viterbi algorithm.

Next, FIG. 3 shows a diagram for explaining the procedure for estimating the survivor path by the Viterbi algorithm, and the procedure for estimating the survivor path by the Viterbi algorithm will be described below with reference to FIG. In the following description, a procedure for determining a surviving path that is one path from the S paths 102 from the pixel position i-1 to the pixel position i shown in FIG. 1 will be described. However, the determination method at this time is determined based on the matching error and the smoothness of the disparity vector.

Node 101 at arbitrary position v of pixel position i
Let E _v be the path metric of the surviving path in
_{When u} is the path metric of each node 101 at the pixel position i−1, E _v is determined by the following equation (1). However, the path metric in the specification of the present application indicates the total energy of the paths included in the surviving paths. Also, u is the vertical node 1 at pixel position i-1.
01 and v corresponds to the vertical node 101 at the pixel position i, so u and v are 0 ≦ u, respectively.
≤S-1, 0≤v≤S-1.

[0024]

[Equation 1] Here, ssd _v is a matching error energy _{when v} is the parallax vector, D _{u, v} is an energy component representing the smoothness of the parallax vector, and the following equation (2),
It is given in (3).

[0025]

[Equation 2] For the matching error energy, as shown in equation (2), a rectangular block 402 centered on pixel i and indicated by Xi
The integral value in is used. It should be noted that X centered on the pixel i
The integral value in the rectangular block 402 indicated by i will be described later.

Further, the smoothness is as shown in the equation (3),
The norm of u and v. The initial value of the equation (1) is i
When _u = 0, E _u = 0.

In this way, one path is determined from the S paths 102 input to the node 101 at the pixel position i, for example. Since one path is left out of all the nodes 101 corresponding to the pixel i, there are always S paths between each node 101 at the pixel position i−1 and each node 101 at the pixel position i. survive. When the pixel position reaches the end point on the right end, the path having the smallest path metric is selected from the S paths, and the node number through which the path has passed is set as the parallax vector of each pixel. This process can be performed for all scan lines to determine the disparity vector for the entire image.

Here, the corresponding points of the pixels on the predetermined scanning line of the R image always exist on the same scanning line of the L image, that is, the continuity in the vertical direction of the parallax vector and the corresponding scanning line at the same vertical position. Based on the feature that it exists, the energy component representing the smoothness of the parallax vector according to the above equation (3) is calculated as shown in the following equation (4) to correspond to each pixel on one scanning line. The parallax vector is calculated collectively using the Viterbi algorithm.

[0029]

[Equation 3] However, the first term of the equation (4) is a constraint condition on the epipolar line, and it is a condition that the corresponding point of the predetermined pixel in the R image always exists on the right side of the corresponding point of the preceding pixel. . This condition holds as long as it is not occluded, ie it is visible on one side but hidden on the other.

The second term is a condition in the vertical direction. d
_p is a parallax vector at the pixel position one line before, and is multiplied by ssd _v so that this constraint condition becomes small when ssd _v at the current pixel position is small, that is, when the reliability is high.

FIG. 4 is a diagram for explaining the rectangular block according to the first embodiment. As shown in FIG. 4, the first embodiment
Then, the matching error energy in the above equation (2) is 3 with a predetermined pixel (for example, pixel i) as the center.
The integral value in the rectangular block 402 of × 3 pixels is used. The number of pixels of the rectangular block 402 is not limited to 3 × 3, and needless to say, it may be another number of pixels.

FIG. 5 is a parallax vector image calculated by using the corresponding point searching method of the first embodiment, and FIG. 6 is a parallax vector image calculated by the block matching method which is a conventional corresponding point searching method.

However, FIGS. 5 and 6 are obtained by obtaining the parallax vector image from the pair of stereo images shown in FIG. In the parallax vector image shown in FIG. 5, the block size of the rectangular block 402 indicated by Xi is 3 × 3 pixels, λ
_L is 3000, λ _S is 1000, λ ₂ is 0.6, and the search range is 0 to 31. In addition, in FIGS. 5 and 6, the gradation value obtained by the calculation is multiplied by 8 in order to easily understand the difference in gradation.

As is apparent from FIGS. 5 and 6, in the parallax vector image calculated by using the corresponding point search method of the first embodiment, the parallax vector obtained by the block matching method is larger than that of the parallax vector image obtained by the block matching method. There are few mistakes. As a result, it is clear that an overall smooth image can be obtained while maintaining the sharpness of the edge portion of the image.

As described above, in the corresponding point search method according to the first embodiment, all parallax vector values included in the search range are set as the node 101 for the pixel row on one scanning line.
A state transition model is configured with the matching error and the continuity of the vector between adjacent nodes as the weight of the path 102 between the nodes, and the parallax vector for one scanning line is collectively obtained by the Viterbi algorithm. The disparity vector can be obtained.

Further, in the corresponding point search method of the first embodiment,
Since all the image information on one scanning line is used and the corresponding point search is performed in consideration of the continuity in the vertical direction of the screen, it is possible to make it difficult to be affected by noise or lack of texture and to perform stable corresponding point search. .

Further, since it is resistant to noise, the size of the block for calculating the cost function can be reduced, and high-resolution corresponding point search can be performed. Furthermore, since the Viterbi algorithm has been put to practical use in the decoding of convolutional codes, it can be easily implemented in hardware.

In the corresponding point searching method between images of the first embodiment, the corresponding point searching method of the parallax vector for two stereo images forming a pair has been described, but the present invention is not limited to this. The node 101, which is the disparity vector value included in the vector search range, is made two-dimensional, and the motion vector of each pixel on one scanning line is collectively estimated to be applied to the corresponding point search method of the motion vector of the moving image. It is possible.

However, the motion vector of the moving image has a horizontal direction component and a vertical direction component. Therefore, when the present invention is applied to the corresponding point search method of the motion vector of the moving image, for example, the node 101 shown in FIG. 1 described above is converted into a plane (hereinafter referred to as “node plane”) as shown in FIG. It is applicable by replacing with the complex number node 701 arranged in (notation). However, the horizontal vector in FIG. 7 shows the real part, and the vertical vector shows the imaginary part.

That is, as shown in FIG. 7, the search range in the horizontal direction is ± H and the search range in the vertical direction is ± H with reference to the complex number node 701 indicated by the coordinate position (0,0).
The complex number node 701 of V is the node 10 shown in FIG.
Provide for each one. At this time, the adjacent node 101 in FIG.
The inter-path is a path as a combination of all node states from each complex number node 701 arranged on one node surface to each complex number node 701 arranged on the other node surface as a path between adjacent node surfaces. Becomes

In this way, a two-dimensional state transition model is constructed by defining a path between adjacent node faces, and the optimum path is selected using the Viterbi algorithm to thereby determine the movement at each node position. Since it is possible to estimate the vector value, it is possible to obtain a motion vector of a moving image that is smooth and has few errors.

Therefore, the present invention can be applied to the motion vector search in the MPEG encoding device,
For example, as shown in FIG. 9, by applying the corresponding point search method of the motion vector of the moving image of the present invention described above to the conventional motion vector searching unit, the conventional motion vector searching unit is used for the motion of the moving image of the present invention. It can be replaced with the motion vector search unit 901 using the vector corresponding point search method. That is, as shown in FIG. 9, the output of the motion vector search unit 901 to which the invention of the present application is applied is a known MP.
By inputting them to the EG video coding unit 902 and the well-known motion vector coding unit 903, and multiplexing the outputs from the MPEG video coding unit 902 and the motion vector coding unit 903 by the well-known multiplexing means 904. , MPEG code will be obtained. At this time, the motion vector search unit 901 according to the method for searching the corresponding points of the motion vector of the moving image according to the present invention.
, It is possible to improve the estimation accuracy of the motion vector, so that it is possible to improve the coding efficiency of the inter-frame predictive coding used in the MPEG video coding.

In MPEG, motion vector information is encoded by DPCM and multiplexed and transmitted. That is, only the difference from the motion vector value of the previous pixel is transmitted. This difference value is further Huffman coded, and information having a small difference value is assigned a short code as information having a high frequency. Therefore, the smaller the difference value, the more efficient transmission can be expected. On the other hand, in the conventional motion vector estimation method, the estimation error is large, and the DPCM-coded difference value may be large. On the other hand, by applying the present invention, a smooth vector sequence can be obtained by the Viterbi algorithm, so that the DPCM-encoded difference value can be reduced, and the motion vector can be efficiently transmitted.

(Second Embodiment) FIG. 8 is a diagram for explaining the schematic configuration of a matching device according to a second embodiment of the present invention. However, 801 to 803 are storage means, and 804 to 8
Reference numeral 07 denotes a matching error energy calculation unit, 808 a Viterbi calculation unit, and 809 a minimum value selection unit. Further, in the matching device of the second embodiment, the corresponding points of pixels on a predetermined scanning line of the R image always exist on the same scanning line of the L image, that is, the continuity of the parallax vector in the vertical direction and the corresponding point are the same vertical lines. The condition is that the position exists on the scanning line. Furthermore, the matching device according to the second embodiment has a configuration opposite to that of the first embodiment, and is a device that searches for a corresponding destination of the L image that is the left image from the R image that is the right image.

In the matching apparatus of the second embodiment shown in FIG. 8, the right image (R image) is input from the R image input terminal, and the left image (L image) is input from the L image input terminal. Has become. At this time, the image input from the R image input terminal and the L image input terminal is configured such that the pixel values of the same one scanning line are sequentially input. That is, the R image and the L image are synchronously read in the same pixel order on the same scanning line and input to the corresponding input terminals (the R image input terminal and the L image input terminal), respectively. .

Further, the matching device of the second embodiment is
For example, it is configured to include well-known storage units 801 to 803 that sequentially store pixel values for each pixel input from the input terminal of the R image. As shown in FIG. 8, each storage means 8
01 to 803 have a configuration in which a plurality of storage units 801 to 803 are connected in series, and the pixel value input from the input terminal of the R image is the first storage unit (first storage unit) 801. To the third storage means 803 are sequentially shifted and stored. However, the number of storage units 801 to 803, that is, the number of stages is appropriately determined depending on the number of pixels to be the search range of the R image. Further, the pixel value of the same pixel as the pixel input to the first storage unit 801 is input to the 0th matching error energy calculation unit 804.

Further, the matching device of the second embodiment is
Matching error energy calculating means for inputting the pixel value for each pixel input from the input terminal of the L image to one input and for inputting the pixel value output from the input terminal of the R image or the storage means 801 to 803 to the other input 804-807
It is configured to have. The matching error energy calculation means 804 to 807 are the same as those in the first embodiment described above.
As shown in, it is a means for calculating the matching error energy ssd _v when the parallax vector is v from two inputs, based on the equation (2), and is realized by a program operating on a well-known information processing device, for example. It Also,
The matching device according to the second embodiment is configured to have a plurality of matching error energy calculating means 804 to 807, and each matching error energy calculating means 804.
˜807 are connected in parallel to the input of the L image. However, the matching error energy calculation units 804 to 807 have a number obtained by adding 1 to the number of storage units 801 to 803.

Each matching error energy calculation means 80
The outputs of 4 to 807 are configured to be input to the Viterbi operation unit 808, and the Viterbi operation unit 808 is configured to sequentially perform the path metric operation based on the well-known Viterbi algorithm. However, the Viterbi operation unit 80
The calculation of the path metric by 8 is reset when the input pixel reaches the left side of the scanning line. With such a structure, the path metric of the search range for each scanning line is calculated. To do.

The output of the Viterbi operation unit 808 is input to the minimum value selection unit 809. When the path metric operation by the Viterbi operation unit 808 reaches the rightmost node on the scanning line, the minimum value is output. The selecting unit 809 selects the path having the smallest value among the path metrics input from the Viterbi calculating means, and sets 1 as the minimum path of the paths.
It is configured to output parallax vectors for scanning. In addition, the minimum path selection unit 809 selects the minimum path when the pixel value reaches the right end,
With such a configuration, the path metric for each scanning line is calculated. In addition, the matching device according to the second embodiment is configured to sequentially perform the above-described operation for all scanning lines.

As described above, the matching apparatus according to the second embodiment uses 1 when searching for corresponding points of a pair of images.
Since the corresponding points of the pair of images are searched by sequentially repeating the operation of collectively obtaining the parallax vectors of the scanning line segments, it is possible to obtain a smooth parallax vector with few errors.

Further, since all the image information on one scanning line is used and the corresponding points are searched in consideration of the continuity in the vertical direction of the screen, it is possible to suppress the influence of noise and lack of texture, and to stabilize. Corresponding point search can be performed.

Furthermore, since it is resistant to noise, the size of the block for calculating the cost function can be reduced, and high-resolution corresponding point search can be performed.

Further, as shown in FIG. 10, by providing a well-known distance calculating means 1002 for calculating a distance from a parallax vector which is an output of the matching device 1001 using the corresponding point searching method of the present invention, the accuracy is high. It goes without saying that it is possible to configure a distance measuring device using a stereo image from which distance information can be obtained. However, L image and R
In order to obtain the distance from the camera to the subject based on the parallax vector (parallax value) obtained by searching the corresponding points between the images, the distance between the cameras that photographed the L image and the R image is B, and the camera is The focal length of F is F, the physical size of the pixel of the camera is ps, and the parallax value (the number of pixels) is d.

[0054]

[Equation 4]

As described above, the invention made by the present inventor is
Although the specific description has been given based on the embodiment of the invention, the invention is not limited to the embodiment of the invention and can be variously modified without departing from the scope of the invention. .

[0056]

The effects obtained by the typical ones of the inventions disclosed in the present application will be briefly described as follows.

(1) Since the parallax vector for one scanning line can be collectively obtained by the Viterbi algorithm,
It is possible to obtain a smooth disparity vector with few errors. (2) Since all the image information on one scanning line is used and the corresponding point search is performed in consideration of the continuity in the vertical direction of the screen, it is possible to make it difficult to be affected by noise or lack of texture, and to obtain a stable corresponding point. You can search.

(3) Since it is resistant to noise, the size of the block for calculating the cost function can be reduced, and high-resolution corresponding point search can be performed. (4) Since the Viterbi algorithm has been put to practical use in the decoding of convolutional codes, it can be easily implemented in hardware.

[Brief description of drawings]

FIG. 1 is a diagram for explaining a corresponding point searching method between images according to a first embodiment of the present invention.

FIG. 2 is an example of two stereo images taken by moving the camera in parallel and is a halftone image displayed on a display.

FIG. 3 is a diagram for explaining a procedure for estimating a survivor path by the Viterbi algorithm.

FIG. 4 is a diagram for explaining a rectangular block according to the first embodiment.

FIG. 5 is a parallax vector image calculated by using the corresponding point search method according to the first embodiment, and is a halftone image displayed on the display.

FIG. 6 is a parallax vector image calculated by a block matching method which is a conventional corresponding point search method, and is a halftone image displayed on a display.

FIG. 7 is a diagram for explaining a schematic operation for applying the present invention to a corresponding point search method for a motion vector of a moving image.

FIG. 8 is a diagram for explaining a schematic configuration of a matching device according to a second embodiment of the present invention.

FIG. 9 is a diagram for explaining a schematic configuration of an MPEG device to which the present invention is applied.

FIG. 10 is a diagram for explaining a schematic configuration of a distance measuring device using a stereo image to which the present invention is applied.

[Explanation of symbols]

101 ... Node 102 ... Path 401 ... Pixel 402 ... Rectangular block 701 ... Complex node 801 to 803
Storage units 804 to 807 Matching energy calculation unit 808 Viterbi calculation unit 809 Minimum value selection unit 901 Motion vector search unit 902 MPE
G video coding unit 903 ... Motion vector coding unit 904 ... Multiplexing means 1001 ... Matching device 1002 ... Distance calculation means

Continued front page    (72) Inventor Miwa Katayama             1-10-11 Kinuta, Setagaya-ku, Tokyo, Japan             Broadcasting Association Broadcast Technology Institute F term (reference) 5C059 MA00 NN03 NN28 PP04 PP26                       UA02                 5L096 FA69 GA19 HA01 JA03

Claims (4)

[Claims] 1. A method of searching a corresponding point between images for identifying a corresponding destination for each pixel of one captured image of a pair of captured images with respect to the other captured image, in which scanning is performed on the one captured image. A step of specifying a line, a step of setting a disparity vector value included in the search range of the other captured image as a node of a state transition model, a matching error in a preset block and a vector between adjacent nodes A step of setting continuity as a weight of a path between nodes of the state transition model, and 1 of the other captured image by the Viterbi algorithm based on the state transition model.
A step of collectively estimating a corresponding point of each pixel on a scanning line. 2. A method of searching for corresponding points between images according to claim 1, wherein the parallax vector value of a pixel in any one scanning line before and after the scanning line in the one captured image is changed to the state transition. A method of searching for corresponding points between images, comprising a step of setting a model constraint condition. 3. A method of searching a corresponding point between frames for identifying a corresponding destination for each pixel between consecutive frames of a moving image, wherein a scanning line in a previous frame image of the moving image is specified. A step of setting a motion vector value included in a search range in a two-dimensional direction of a moving image subsequent to the preceding moving image as a node of a state transition model, a matching error in a preset block, and an adjacent node Setting the continuity of vectors between nodes as the weight of the path between the nodes of the state transition model, and determining the corresponding point of each pixel on one scan line of the subsequent frame by the Viterbi algorithm based on the state transition model. A method of searching for corresponding points between frames, which comprises a step of collectively estimating. 4. A matching device that specifies a corresponding destination for each pixel of one captured image of a pair of captured images in the other captured image, and a pixel on one scanning line of the one captured image and the one scan. Means for calculating a parallax vector value from the pixels in the search range of the other captured image corresponding to the line, and setting the parallax vector value as a node of the state transition model, and matching error and adjacency in a preset block The continuity of the vector between the nodes is set as the weight of the path between the nodes of the state transition model, and the corresponding point of each pixel on one scanning line of the other captured image is determined by the Viterbi algorithm based on the state transition model. A matching device comprising: means for collectively estimating.