JP2013069026A - Device, method, and program for restoring three-dimensional shape of subject - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a device for accurately restoring a three-dimensional shape by calculating many points on a surface of a subject from a multi-viewpoint image.SOLUTION: A device for restoring the three-dimensional shape of a subject, based on a multi-viewpoint image obtained by photographing a subject by a plurality of cameras and camera parameters of the respective cameras, extracts an occluding contour line of the subject in the multi-viewpoint image, calculates a tangent line corresponding to each point constituting the occluding contour line from the camera parameters, estimates the position of a tangent point with respect to a subject surface on the calculated tangent line, and restores the shape of the subject surface with the tangent point as a feature point on the subject surface.

Description

  The present invention relates to an apparatus, method, and program for obtaining a number of points on the surface of a subject from a multi-viewpoint image and restoring a three-dimensional shape with high accuracy.

  Conventionally, various proposals have been made on techniques for restoring a three-dimensional shape of a subject based on multi-viewpoint images taken by a plurality of cameras so as to surround the subject.

  In the conventional three-dimensional shape restoration method, a likely surface passing through these points and lines is restored using the points and lines that are highly likely to be on the surface of the subject. As points and lines that serve as clues, markers or the like pasted on the surface of the subject, characteristic points, lines, frontier points (Non-Patent Document 1), etc. on the surface of the subject are used.

Hiyama, Katayama, Orihara, Iwabuchi, “Reconstruction method of 3D shape constrained by local shape features and its real-time video display,” Journal of the Institute of Image Information and Television Engineers, vol. 61, no. 4, pp. 471-481 (2007)

  However, since there are generally a small number of points and lines on the surface of the subject that can be obtained using the above-described known technique, the accuracy of the restored three-dimensional shape is low.

  Therefore, an object of the present invention is to provide an apparatus, a method, and a program for obtaining many points on the surface of a subject from a multi-viewpoint image and restoring a three-dimensional shape with high accuracy.

  In order to achieve the above object, an apparatus for restoring the three-dimensional shape of a subject according to the present invention is a device for restoring the three-dimensional shape of a subject from a multi-viewpoint image obtained by photographing the subject with a plurality of cameras and camera parameters of each camera. Tangent calculating means for extracting a shielding contour of a subject in the multi-viewpoint image and obtaining a corresponding tangent at each point constituting the shielding contour from the camera parameter; and a subject surface on the calculated tangent Contact estimation means for estimating the position of the contact with respect to the surface, and surface restoration means for restoring the shape of the surface of the subject using the contact as a feature point on the subject surface.

  It is also preferable that the contact estimation means sets a tangent plane that contacts the subject surface through the tangent calculated by the tangent calculation means, and estimates the position of the contact while translating the tangent plane along the tangent. .

  In the contact estimation means, a function for projecting the pixels on the tangent plane onto each multi-viewpoint image and evaluating the degree of coincidence of the corresponding pixel values on each image is set, and the function is maximized. It is also preferable to use a point as a contact.

  Moreover, in the said contact estimation means, it is also preferable to use the function which decreases monotonously with respect to the square sum of the difference of the said applicable pixel value as a function which evaluates the said matching degree.

  In the contact point estimation means, it is also preferable to use a function that monotonously decreases with respect to the variance or standard deviation of the corresponding pixel value as a function for evaluating the degree of coincidence.

  In the contact point estimation means, it is also preferable to perform weighting according to the position of the camera that has captured the multi-viewpoint image when calculating the function for evaluating the degree of coincidence.

  In the contact estimation means, it is also preferable to set a weight that monotonously increases with respect to the inner product of the normal vector of the tangent plane and the optical axis vector of the camera that has captured the multi-viewpoint image.

  In the contact estimation means, it is also preferable to set a weight that monotonously decreases with respect to an angle formed by a normal vector of the tangent plane and an optical axis vector of a camera that has captured the multi-viewpoint image.

  It is also preferable to use a frustum surface that passes through the tangent calculated by the tangent calculation means and touches the subject surface instead of the tangent plane.

  In order to achieve the above object, the method of restoring the three-dimensional shape of the subject according to the present invention is a method of restoring the three-dimensional shape of the subject from a multi-viewpoint image obtained by photographing the subject with a plurality of cameras and camera parameters of each camera. A tangent calculation step of extracting a shielding contour of a subject in the multi-viewpoint image, obtaining a tangent corresponding to each point constituting the shielding contour from the camera parameter, and a subject surface on the calculated tangent A contact estimation step for estimating the position of the contact with respect to the surface, and a surface restoration step for restoring the shape of the surface of the subject using the contact as a feature point on the subject surface.

  In order to achieve the above object, a program according to the present invention causes a computer to function as the above-described apparatus.

  According to the contact point estimation means of the present invention, there is a geometrical feature that at least one contact point (a point on the surface of the object) always exists on a tangent line corresponding to a point constituting the shielding outline in the multi-viewpoint image. By using various conditions, the search range of points on the surface of the subject can be limited to the tangent line, and the position of the contact point with respect to the subject surface can be estimated. By restoring the three-dimensional shape of the surface of the subject using the set of contact points as a constraint, it is possible to obtain an effect that the restored three-dimensional shape can be highly accurate.

The system block diagram of the three-dimensional shape restoration apparatus by this invention is shown. The calculation of the tangent line drawn on the surface of the subject is shown. Fig. 3 shows contact estimation using a tangent plane according to the first embodiment. The contact estimation using the frustum surface by 2nd Embodiment is shown. The reconstruction of the three-dimensional shape of the surface of the subject is shown.

  EMBODIMENT OF THE INVENTION The form for implementing this invention is demonstrated in detail below using drawing. FIG. 1 is a system configuration diagram of a 3D shape restoration apparatus according to the present invention. As shown in FIG. 1, the 3D shape restoration apparatus 1 includes a multi-viewpoint image input unit 11, a tangent calculation unit 12, and a contact point estimation unit 13. , And a surface restoration unit 14.

  The multi-viewpoint image input unit 11 inputs image data (multi-viewpoint image) obtained by synchronously shooting a subject with a plurality of cameras arranged so as to surround the subject. It is assumed that camera parameters are already known. In each of the multi-viewpoint images, the subject and the background are separated and the silhouette of the subject is extracted. Here, a known technique may be used as a method for separating the subject and the background. For example, as a method for extracting the silhouette of the subject, chroma key, background difference or manual extraction is used. From the extracted silhouette, the occlusion outline which is the outline of the silhouette of the subject in the multi-viewpoint image is detected. A known technique may be used as a method for detecting the shielding contour. For example, a method of approximating with a polygonal line using the approximation algorithm of the Teh-Chin chain is used. You may use the cvFindContours function of OpenCV. The visual hull of the subject is obtained by applying the visual volume intersection method using the occlusion contour. In general, the Visual Hull acquired using the visual volume intersection method includes a true three-dimensional shape of a subject and circumscribes the three-dimensional shape.

  The tangent calculating unit 12 calculates a tangent drawn from the principal point (viewpoint) of the camera to the surface of the subject in each of the plurality of cameras. FIG. 2 shows the calculation of the tangent line drawn on the surface of the subject. 2a shows a view from the side and FIG. 2b shows a view from above. As shown in FIG. 2b, generally, a tangent drawn from the main point of the camera to the surface of the subject is projected onto one point on the boundary of the silhouette of the subject in the camera image. In general, the Visual Hull and the true three-dimensional shape are in contact with each other at the contact point between the tangent line and the subject.

  The contact point estimation unit 13 estimates the three-dimensional position of the contact point between the tangent line and the subject. As shown in FIG. 2, the contact always exists on the tangent line, but it is unclear at which position on the tangent line. Therefore, in the first embodiment, as shown in FIG. 3, a tangent plane that passes through the tangent line and touches the surface of the subject is set, and the tangent plane is translated on the tangent line while matching with images of other cameras. Estimate the contact position so that Alternatively, in the second embodiment, as shown in FIG. 4, a frustum surface that passes through the tangent line and touches the surface of the subject is set, and the frustum surface is translated on the tangent line and aligned with the images of other cameras. Estimate the contact position so that A known technique may be used as a method of matching with images from a plurality of cameras.

  As shown in FIG. 5, the surface restoration unit 14 estimates the positions of other points on the subject surface based on the fact that the set of contacts acquired by the contact estimation unit 13 is a set of points on the subject surface. Thus, the three-dimensional shape of the surface of the subject is restored. As a restoration method, a known technique such as depth estimation by graph cut may be used.

  Below, the tangent calculation means by the tangent calculation section 12 and the contact estimation means by the contact estimation section 13 will be described in detail. The following camera parameters of each camera that has acquired a multi-viewpoint image are known.

と表される。 With c as the camera ID, the internal parameter matrix A _{c of} the camera c, the rotation matrix R _{c of} the camera c, and the position coordinate Mc of the camera _c are:

It is expressed.

と表され、各カメラにおいて、式（１）の射影変換が成り立つ。

なお、ｍ_ｃ ^〜はｍ_ｃの斉次表現であり、λはスケール不定性を表すスカラーである。 Coordinates M of points in the real space, the coordinates m _c points in the camera c image,

In each camera, the projective transformation of Expression (1) is established.

Note that m _c ^~ is a homogeneous expression of m _c , and λ is a scalar representing scale indefiniteness.

A normal vector and a tangent in the image are obtained from the line elements constituting the polygonal line obtained when detecting the shielding contour. Focusing on line elements of polygonal lines constituting the occluding contour in camera c images, the end points each _{m s = (u s, v} s) T and _{m e = (u e, v} e) putting as ^T, the The line element is given by Equation (2) with t (0 ≦ t ≦ 1) as a parameter. Further, the normal vector n _{(m c)} at the point _{m c} of the line Motojo is represented by formula (3).

  From the above parameters, the tangent calculation unit 12 calculates a tangent corresponding to each point constituting the shielding contour and obtains the range of the common part with the Visual Hull as follows.

なお、Ｒ_ｃ ^ＴはＲ_ｃの転置行列であり、Ａ_ｃ ^−１はＡ_ｃの逆行列である。 Corresponding to the point m _c on the contour of the camera c images, tangent to the subject is given by Equation (4) s of (s> 0) as parametric.

R _c ^T is a transposed matrix of R _c , and A _c ⁻¹ is an inverse matrix of A _c .

なお、Ａ_ｃ’、Ｒ_ｃ’、Ｍ_ｃ’は、それぞれカメラｃ’の内部パラメータ行列、カメラｃ’の回転行列、カメラｃ’の位置座標を表し、ｍ_ｃ ^〜’、はカメラｃ’画像中の点の座標を表す。 In order to obtain the range of the common part with the Visual Hull, that is, the range of the parameter s, tangent lines are projected on all camera images (camera c ′ images) other than c. Specifically, by substituting Equation (4) into Equation (1 ′), Equation (5) representing a straight line projected onto the camera c ′ image is obtained.

A _c ′, R _c ′, and M _c ′ represent the internal parameter matrix of the camera c ′, the rotation matrix of the camera c ′, and the position coordinates of the camera c ′, respectively, and m _c ^to ′ represent the camera c ′ image. Represents the coordinates of the middle point.

式（５）と式（２’）を連立すれば、ｓの値は式（６）で与えられる。

全てのカメラ画像についてｓの範囲を算出し、その積集合（共通部分）を求める。これがｓの値の範囲になる。 In the camera c ′ image, the position of the intersection between the projected straight line and the line element constituting the shielding contour is obtained. Camera c 'line element constituting an occluding contour in the image, the equation (2), the end point, respectively m _s, m _e, when the parametric put is t, equation (2' given in).

If Equation (5) and Equation (2 ′) are combined, the value of s is given by Equation (6).

The range of s is calculated for all camera images, and the product set (common part) is obtained. This is the range of the value of s.

  Next, the contact point calculation means by the contact point estimation unit 13 will be described. For each tangent, the three-dimensional position of the contact point between the tangent and the subject is estimated. Hereinafter, a method for calculating the degree of coincidence by approximating the surface of the subject according to the first embodiment with a plane passing through a tangent will be described (FIG. 3).

ただし、格子点の個数を（２Ｎ＋１）×（２Ｎ＋１）（Ｎ：自然数）として、

とする。具体的には、式（１’）に式（１０）を代入すればよい。 First, a tangent plane is set on the tangent line, and lattice points are arranged on the tangent plane. Specifically, attention is focused on a point m _c on the extracted contour. Tangent line corresponding to the point m _c is represented by the formula (4), the range of s values is obtained using equation (6) and formula (2 '). Next, a tangent plane to the subject at a point M on the tangent line is obtained, and two-dimensional coordinates are set on the tangent plane to arrange grid points. Assuming that the normal vector of the tangent plane is N, N is expressed by Equation (7). On the tangent plane, if the coordinate axis in the tangential direction is the S axis and the coordinate axis in the direction perpendicular to the S axis is the T axis, the unit vector S in the S axis direction is expressed by Equation (8), and the unit in the T axis direction Vector T is expressed by equation (9). The coordinates of the lattice points on the tangent plane are expressed by equation (10).

However, the number of lattice points is (2N + 1) × (2N + 1) (N: natural number),

And Specifically, equation (10) may be substituted into equation (1 ′).

  Next, the grid points are projected onto each camera image while the tangent plane is translated on the tangent line. Each lattice point represented by Expression (10) is projected onto each camera image while changing the value of s within the range obtained using Expression (2 ′). A known technique may be used for the interpolation. For example, a Lanczos filter or the like is used. Matching is performed at a grid point projected on each camera image, a function for evaluating the degree of coincidence of pixel values at the grid point is set, and a value of s that maximizes the function is obtained. A point corresponding to the value of s at this time is a contact point between the subject and the tangent line in the real space.

Here, there are the following two examples of functions for evaluating the degree of coincidence.
A sum of squares of pixel value differences of pixels corresponding to each grid point is used. The smaller the sum of squares, the higher the degree of coincidence.
Use the variance or standard deviation of the pixel values of the pixels corresponding to each grid point. The smaller the variance or standard deviation, the higher the degree of coincidence.

を用いる。角度が大きいほど、重み付けは小さい。
さらに、ダイクストラ法を用いて、隣接する接点が連続となるように調整してもよい（図５）。 Further, when calculating the function for evaluating the degree of coincidence, weighting can be performed according to the position of the camera that has captured the multi-viewpoint image. Each camera image has the following three weights. However, a camera image in which no grid points on the tangent plane are projected by occlusion is not used. Here, a vector from the contact point toward the principal point of the camera is set as a line-of-sight vector V = M _c −M.
・ Each camera image is handled equally.
Use the inner product (V · N) of the normal vector and the line-of-sight vector of the tangent plane. The larger the inner product, the greater the weight.
・ An angle between the normal vector of the tangent plane and the line-of-sight vector

Is used. The greater the angle, the smaller the weight.
Furthermore, you may adjust so that an adjacent contact may become continuous using the Dijkstra method (FIG. 5).

  Hereinafter, a method for calculating the degree of coincidence by approximating the surface of the subject according to the second embodiment with a frustum surface passing through a tangent will be described (FIG. 4).

ただし、格子点の個数を（２Ｎ＋１）×（２Ｎ＋１）（Ｎ：自然数）として、

とする。具体的には、式（１’）に式（１０’）を代入すればよい。 First, a frustum surface is set on the tangent line, and lattice points are arranged on the frustum surface. Specifically, attention is focused on a point m _c on the extracted contour. Tangent line corresponding to the point m _c is represented by the formula (4), the range of values of s is obtained by using equation (6) and formula (2 '). Next, a frustum surface having the main point of the camera as a vertex and a shielding contour as a bottom surface is obtained, and lattice points are arranged on the frustum surface by setting two-dimensional coordinates with a point M on the tangent as an origin. If the coordinate axis in the tangential direction is set as the S axis on the frustum surface, the unit vector S in the S axis direction is expressed by the equation (8 ′). If a point that is T apart along the shielding contour on the camera image is set as m _c ″ = (u _c ″, v _c ″), the coordinates of the lattice points on the frustum surface are expressed by Expression (10 ′).

However, the number of lattice points is (2N + 1) × (2N + 1) (N: natural number),

And Specifically, the formula (10 ′) may be substituted into the formula (1 ′).

  Next, the grid points are projected onto each camera image while the frustum surface is translated on the tangent line. While changing the value of s within the range obtained using Expression (2 '), each lattice point represented by Expression (10') is projected onto each camera image. A known technique may be used for the interpolation. For example, a Lanczos filter or the like is used. Matching is performed at a grid point projected on each camera image, a function for evaluating the degree of coincidence of pixel values at the grid point is set, and a value of s that maximizes the function is obtained. A point corresponding to the value of s at this time is a contact point between the subject and the tangent line in the real space.

There are the following two examples of functions for evaluating the degree of coincidence.
A sum of squares of pixel value differences of pixels corresponding to each grid point is used. The smaller the sum of squares, the higher the degree of coincidence.
Use the variance or standard deviation of the pixel values of the pixels corresponding to each grid point. The smaller the variance or standard deviation, the higher the degree of coincidence.

を用いる。角度が大きいほど、重み付けは小さい。
さらに、ダイクストラ法を用いて、隣接する接点が連続となるように調整してもよい（図５）。 Further, when calculating the function for evaluating the degree of coincidence, weighting can be performed according to the position of the camera that has captured the multi-viewpoint image. Each camera image has the following three weights. However, a camera image in which no grid point on the frustum surface is projected by occlusion is not used. Here, a vector from the contact point toward the principal point of the camera is set as a line-of-sight vector V = M _c −M.
・ Each camera image is handled equally.
Use the inner product (V · N) of the normal vector and the line-of-sight vector. The larger the inner product, the greater the weight.
・ An angle between normal vector and line-of-sight vector

Is used. The greater the angle, the smaller the weight.
Furthermore, you may adjust so that an adjacent contact may become continuous using the Dijkstra method (FIG. 5).

  As described above, according to the present invention, since a contact point (a point on the surface of the object) is always obtained for each tangent line, a large number of points and lines with known three-dimensional position coordinates can be obtained. Can be made easier. Further, in a known three-dimensional shape restoration method, the three-dimensional shape restoration can be made highly accurate by using together with characteristic points and lines used in the method.

  Moreover, all the embodiments described above are illustrative of the present invention and are not intended to limit the present invention, and the present invention can be implemented in other various modifications and changes. Therefore, the scope of the present invention is defined only by the claims and their equivalents.

DESCRIPTION OF SYMBOLS 1 3D shape restoration apparatus 11 Multi-viewpoint image input part 12 Tangent calculation part 13 Contact point estimation part 14 Surface restoration part

Claims (11)

An apparatus for restoring a three-dimensional shape of a subject from a multi-viewpoint image obtained by photographing the subject with a plurality of cameras and camera parameters of each camera,
A tangent calculating means for extracting a shielding contour of a subject in the multi-viewpoint image and obtaining a corresponding tangent at each point constituting the shielding contour from the camera parameter;
Contact estimation means for estimating the position of the contact with respect to the subject surface on the calculated tangent line;
Surface restoration means for restoring the shape of the surface of the subject using the contact as a feature point on the subject surface;
An apparatus for restoring a three-dimensional shape of a subject characterized by comprising:   The contact estimation means sets a tangent plane that contacts the subject surface through the tangent calculated by the tangent calculation means, and estimates the position of the contact while translating the tangent plane along the tangent. The apparatus of claim 1.   In the contact point estimation means, a pixel on the tangent plane is projected onto each multi-viewpoint image, a function for evaluating the degree of coincidence of corresponding pixel values on each image is set, and the point at which the function is maximized is set. The device according to claim 2, wherein the device is a contact.   The apparatus according to claim 3, wherein the contact estimation unit uses a function that monotonously decreases with respect to a sum of squares of the difference between the corresponding pixel values as the function for evaluating the degree of coincidence.   The apparatus according to claim 3, wherein the contact estimation unit uses a function that monotonously decreases with respect to a variance or a standard deviation of the corresponding pixel value as a function for evaluating the degree of coincidence.   The weighting according to any one of claims 3 to 5, wherein, in the contact estimation means, when calculating the function for evaluating the degree of coincidence, weighting is performed according to a position of a camera that has captured the multi-viewpoint image. The device according to item.   The weight of the contact point estimation unit is set to a monotonically increasing weight with respect to an inner product of a normal vector of the tangent plane and an optical axis vector of a camera that has captured the multi-viewpoint image. 6. The apparatus according to 6.   The contact estimation means sets a weight that monotonously decreases with respect to an angle formed by a normal vector of the tangent plane and an optical axis vector of a camera that has captured the multi-viewpoint image in the weighting. Item 7. The apparatus according to Item 6.   9. The apparatus according to claim 2, wherein instead of the tangent plane, a frustum surface that passes through a tangent calculated by the tangent calculation unit and is in contact with a subject surface is used. A method for restoring a three-dimensional shape of a subject from a multi-viewpoint image obtained by photographing the subject with a plurality of cameras and camera parameters of each camera,
Extracting a shielding contour of a subject in the multi-viewpoint image and calculating a tangent corresponding to each point constituting the shielding contour from the camera parameters; and
A contact estimation step for estimating the position of the contact with respect to the subject surface on the calculated tangent line;
A surface restoration step for restoring the shape of the surface of the subject using the contact point as a feature point on the subject surface;
A method for reconstructing a three-dimensional shape of an object.   A program that causes a computer to function as the apparatus according to claim 1.

Images