KR101184588B1 - A method and apparatus for contour-based object category recognition robust to viewpoint changes - Google Patents

Abstract

The present invention relates to a method and apparatus for recognizing an object based on a contour, which is robust to a change in the viewpoint of an object. The present invention relates to a method and apparatus for contour-based category object recognition that is robust to change in the viewpoint of an object using a form descriptor vector that is robust to the viewpoint change of an object with respect to a pair of feature points without assuming a viewpoint in advance.
To this end, the contour-based category object recognition method robust to changing the viewpoint of an object according to the present invention includes (a) receiving image data, (b) recognizing a category object among the image data, and (c) a category object. Outputting the recognition result.
Accordingly, it is possible to recognize category objects at various viewpoints without assumptions about the viewpoints of the objects.

Description

A method and apparatus for contour-based object category recognition robust to viewpoint changes}

The present invention relates to a method and apparatus for recognizing an object based on a contour, which is robust to a change in the viewpoint of an object. The present invention relates to a method and apparatus for contour-based category object recognition that is robust to change in the viewpoint of an object using a form descriptor vector that is robust to the viewpoint change of an object with respect to a pair of feature points without assuming a viewpoint in advance.

Object recognition is one of the four core technologies of intelligent robots. It is a technology that finds three-dimensional spatial information such as the type, size, direction, and position of an object in real time based on the previously learned knowledge information. .

Such object recognition aims to be robust to various environments and to process in real time in order to realize a natural visual-based manipulation function having human-like performance. In particular, object recognition is a core technology such as data acquisition, recognition, position and attitude estimation, modeling, etc. of 3D objects and environments, and is widely studied in computer science as well as robots.

As part of this research, Lowe ( David G. Lowe , " Distinctive Image Features from Scale - Invariant Keypoints , " International Journal of Computer Vision , Vol. 60, no. 2, pp. 91-110, 2004. ) proposed a robust object recognition method based on the appearance information of the object. However, this method is suitable for recognition of individual objects having a specific texture component, and has a problem in that it is not suitable for recognition of category objects that belong to the same category but may have various color and texture characteristics.

In addition, Leordeanu and 2 others ( Marius Leordeanu , Martial Hebert , and Rahul Sukthankar, " Beyond Local Appearance : Category Recognition from Pairwise Interactions of Simple Features , " in Proceedings of IEEE International Conference on Computer Vision and Pattern Recognition , pp . 1-8, June 2007. ) proposed a method of recognizing category objects by extracting contours from images and using spectral matching of feature point pairs extracted from contours. However, this method has a limitation that can be applied only to category object recognition in a specific view direction by using view information of objects previously assumed in the matching process.

In addition, Ferrari and 3 others ( Vittorio Ferrari , Loic Fevrier , Frederic Jurie , and Cordelia Schmid , " Groups of Adjacent Contour Segments for Object Detection , " IEEE Transactions on Pattern Analysis and Machine Intelligence , Vol . 30, no. 1, pp . 36-51, Jan. 2008. ) proposed a method of extracting contours from the image and recognizing category objects using geometrical relations such as relative position, angle, and distance between connected contours close to a straight line. However, since this method also uses the viewpoint information of the object assumed in advance, it is difficult to apply if the viewpoint of the object changes.

SUMMARY OF THE INVENTION The present invention has been made to solve the above problems, and an object of the present invention is to provide a method and apparatus for recognizing a contour-based category object that operates robustly to change the viewpoint of an object without assuming a viewpoint of the object in advance.

In order to achieve the above object, an outline-based category object recognition method that is robust to a viewpoint change of an object according to the present invention includes the steps of: (a) receiving image data; (b) recognizing a category object among the image data; c) outputting the result of the category object recognition.

Step (b) includes: (d) detecting the contour of the image data; (e) extracting the feature point of the contour, calculating a shape descriptor vector; and (f) model data from which the feature point of the contour is extracted; The method may include performing spectral matching between image data, (g) improving the spectral matching result, and recognizing a category object by estimating a matching position among the image data.

In addition, in the step (d), the edge may be detected to merge the edge pixels within the preset range and the normal vector may merge the edge pixels within the preset range to detect the outline.

로 계산된다고 할 때, 수학식

로 정의될 수 있다. The shape descriptor vector may be defined using at least one of a normal vector of the pair of feature points, a distance between the pair of feature points, a curvature of the feature point, or a position vector of the feature point, and the shape descriptor vector may be defined as e. _ij, the normal vectors of (i, j), which are pairs of the feature points, are respectively θ _i and θ _j , θ _ij is the absolute value of the difference between θ _i and θ _j , and the positions of i and j are respectively (x _i , y _i ) , (x _j , y _j ), the distance between the pairs of feature points (i, j) d _ij , the curvature of the i, j k _i , k _{j, for the} feature point j based on the position vector of the feature point i Measuring the directivity of i is called σ _ij , where σ _ij is

If it is calculated as

It can be defined as.

In addition, step (f) may include measuring deformation between the pair of feature points of the model data and the image data.

으로 정의되고, 상기 γ는 수학식

(δ는 0으로 나누어지는 경우를 방지하기 위한 작은 양의 상수값)에 의해 정의될 수 있다. The shape descriptor vector measures a deformation between (i, j), (a, b), and (i, j) and (a, b) for each pair of feature points of the model data and the image data. The descriptor vector is g _ij (a, b), i and the difference in the local shape histogram between a, j and b is the distance between C _ia , C _jb , and feature point pairs (i, j), respectively, d _{ij ,} ε ₁ Where ε ₆ is the element of the difference vector of the six-dimensional shape descriptor vector defined in (i, j) and (a, b),

Γ is expressed as

(δ can be defined by a small positive constant value to avoid dividing by zero).

Also, in the step (f), a geometrical similarity of the pairs of feature points is measured to construct a matrix, and a cluster of feature points having high geometric similarity is obtained by calculating an eigenvector corresponding to the largest eigenvalue of the matrix. Can be.

Also, the step (f) may give a penalty constant for the case where the pair of feature points of the model data and the image data exist on the same contour line when calculating the potential function for measuring the geometric similarity of the pair of feature points. Can be.

에 의해 정의될 수 있다. The potential function may be a pair of feature points of the model data and the image data of (i, j), (a, b), the penalty constant p, and the potential function of G _{ia : jb} , i, j reliability When v _ij is _expressed as

Can be defined by

에 의해 계산될 수 있다. In addition, the reliability of the i, j is v _ij is the equation

Can be calculated by

In addition, the step (f) may be configured to form an outline only for the pair of feature points given the penalty constant.

In addition, in step (g), a random SAmple consensus (RANSAC) algorithm may be applied to the corresponding points due to the spectral matching, and the application of the RANSAC algorithm estimates a two-dimensional homography transform and uses a symmetric propagation error. Can remove the contour and search for a new corresponding point.

On the other hand, in order to achieve the above object, a contour-based category object recognition apparatus robust to a change in viewpoint of an object according to the present invention includes a data input unit for receiving image data, a category object recognition unit for recognizing a category object among the image data, and the It includes an output unit for outputting the recognition result of the category object recognition unit.

The categorical object recognition unit may include a contour detection unit for detecting an outline of the image data, a feature extractor for extracting feature points of the outline, and a shape descriptor vector, and model data from which the feature points of the outline are extracted and the image data. The spectral matching unit performing the parallel matching, the matching result improvement unit for improving the matching result of the spectral matching unit, and a matching position estimation unit for recognizing a category object by estimating the matching position among the image data.

The contour detection unit detects an edge to detect an edge by merging edge pixels whose edges are within a preset range and a normal vector is within a preset range.

로 계산된다고 할 때, 수학식

로 정의될 수 있다. In addition, the shape descriptor vector may be defined using at least one of a position of the distance, the curvature or the feature point of the feature point between a pair of the normal of the feature point vector, pairs of the feature point vector, the shape descriptor vector e _{ij ,} The normal vectors of (i, j), which are pairs of feature points, are respectively θ _i and θ _j , θ _ij is the absolute value of the difference between θ _i and θ _j , and the positions of i and j are respectively (x _i , y _i ), (x _j , y _j ), the distance between the pair of feature points (i, j) d _ij , the curvature of the i, j k _i , k _{j, and the} feature point i with respect to the feature vector i Is measured as σ _ij , and σ _ij is expressed as

If it is calculated as

It can be defined as.

The spectral matching unit may measure deformation between pairs of feature points of the model data and the image data.

으로 정의되고, 상기 γ는 수학식

(δ는 0으로 나누어지는 경우를 방지하기 위한 작은 양의 상수값)에 의해 정의될 수 있다. The shape descriptor vector measures a deformation between (i, j), (a, b), and (i, j) and (a, b) for each pair of feature points of the model data and the image data. The descriptor vector is g _ij (a, b), i and the difference in the local shape histogram between a, j and b is the distance between C _ia , C _jb , and feature point pairs (i, j), respectively, d _{ij ,} ε ₁ Where ε ₆ is the element of the difference vector of the six-dimensional shape descriptor vector defined in (i, j) and (a, b),

Γ is expressed as

(δ can be defined by a small positive constant value to avoid dividing by zero).

In addition, the spectral matching unit constructs a matrix by measuring geometric similarities with respect to the pair of feature points, and calculates a cluster of feature points having high geometric similarity by calculating an eigenvector corresponding to the largest eigenvalue of the matrix. have.

In addition, the spectral matching unit may give a penalty constant for the case where the pair of feature points of the model data and the image data exist on the same contour line when calculating a potential function for measuring geometric similarity with respect to the pair of feature points. have.

에 의해 정의되고, 상기 i, j의 신뢰도는 v_ij는 수학식

에 의해 계산될 수 있다. The potential function may be a pair of feature points of the model data and the image data of (i, j), (a, b), the penalty constant p, and the potential function of G _{ia : jb} , i, j reliability When v _ij is _expressed as

Where i and j are the reliability of v _ij

Can be calculated by

In addition, the spectral matching unit may configure an outline only for pairs of feature points given the penalty constant.

In addition, the matching result improvement unit may apply a RANSAC (RANdom SAmple Consensus) algorithm to the corresponding points by the spectral matching, and the application of the RANSAC algorithm estimates a two-dimensional homography transform and uses a symmetric propagation error. You can explore the new correspondence by removing the contour.

The invention also includes a computer readable recording medium having recorded thereon a program for executing the above method.

According to the contour-based category object recognition method and apparatus which is robust to the viewpoint change of the object according to the present invention, it is possible to recognize the category object at various viewpoints without assumptions about the object viewpoint.

1 is a block diagram of a contour-based category object recognition apparatus robust to a change in viewpoint of an object according to the present invention.
2 is a diagram illustrating an embodiment of recognizing a category object of a mobile phone by using a contour-based category object recognition device that is robust to a viewpoint change of an object according to the present invention.
3 is a flowchart of a contour-based category object recognition method that is robust to changing the viewpoint of an object according to the present invention.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. Prior to this, terms and words used in the present specification and claims should not be construed as limited to ordinary or dictionary terms, and the inventor should appropriately interpret the concepts of the terms appropriately The present invention should be construed in accordance with the meaning and concept consistent with the technical idea of the present invention.

Therefore, the embodiments described in this specification and the configurations shown in the drawings are merely the most preferred embodiments of the present invention and do not represent all the technical ideas of the present invention. Therefore, It is to be understood that equivalents and modifications are possible.

The present invention extracts the contour of the image data when the image data is input, extracts the feature points from the contour, and uses a shape descriptor vector that is robust to the change of the viewpoint of the object with respect to the pair of feature points without pre-assuming a view of the object in advance. The present invention relates to a contour-based category object recognition method and apparatus that is robust to a change in viewpoint of an object.

FIG. 1 is a block diagram of a contour-based category object recognition apparatus robust to a change in a viewpoint of an object according to the present invention. Referring to FIG.

The contour-based category object recognition apparatus that is robust to a viewpoint change of an object includes a data input unit 100 that receives image data, a category object recognition unit 200 that recognizes a category object among the image data, and the category object recognition unit 200. It includes an output unit 300 for outputting the recognition result.

The data input unit 100 may immediately obtain image data including an image input apparatus or receive image data from an image input apparatus including an external camera.

The category object recognition unit 200 automatically recognizes a category object to be recognized in the received image data. At this time, we define a shape descriptor vector that is robust to the point of view change of an object with respect to the pair of feature points without assumptions about the point of view of the object, detect the matched matching point using a spectral matching method, and run RANSAC (RANdom SAmple Consensus). After removing the contour using the algorithm, we search for a new corresponding point to recognize the category object.

Accordingly, the category object recognition unit 200 includes an outline detection unit 210 for detecting an outline of the image data, a feature extractor 220 for extracting feature points of the outline, and a shape descriptor vector, and a feature point of the outline. Matching among the spectral matching unit 230 for performing spectral matching between the extracted model data and the image data, the matching result improvement unit 240 for improving the matching result of the spectral matching unit 230, and the image data. And a registration position estimator 250 for estimating the position and recognizing the category object.

The contour detector 210 detects an edge and detects an edge by merging edge pixels when the edge is within a preset range and the direction of the normal vector is within a preset range. In this case, the edge detection may use a canny edge detector. Edge pixels within a predetermined range and normal vectors within a predetermined range may find corner pixels having a close distance and similar directions to the normal vector.

The feature extractor 220 extracts feature points by sampling the contour lines detected by the contour detector 210, and calculates a shape descriptor vector for the feature point pair.

The shape descriptor vector may be defined using at least one of a normal vector of the pair of feature points, a distance between the pair of feature points, a curvature of the feature point, or a position vector of the feature point. For example, the shape descriptor vector e _ij for the feature point pair (i, j) is defined by Equation 1 below.

Here, when θ _i and θ _j are normal vectors of (i, j), which are pairs of the feature points, θ _ij is the absolute value of the difference between θ i and θ j, d _ij is the distance between pairs of feature points (i, j), k _i and k _j denote curvatures of i and j, and sigma _ij is measured by directionality of the feature point i with respect to the feature vector j based on the position vector of the feature point i.

Wherein (x _i , y _i ), (x _j , y _j ) means the respective positions of i, j. In the same way, we define σ _ji , where θ _i , The geometric relationship between θ _j , σ _ij , and σ _ji can be obtained. 2 is θ _i of the mobile phone that uses a strong contour-based object recognition category at the time of change of an object according to the present invention showing an embodiment in which a recognition object category of the cellular phone, and, Geometric relationships between θ _j , σ _ij , and σ _ji are shown.

The shape descriptor vector for a pair of feature points is independent of the direction of each feature point by using a normal vector or the like. In addition, since it is defined based on the relative physical quantity between pairs of feature points, it is invariant to parallel movement and in-plane rotation and robust to affine transformation. This allows us to recognize category objects at various points of view without making assumptions about the point of view of the object.

On the other hand, when one scene or object is photographed at different times or perspectives in computer vision, images are obtained in different coordinate systems. Image registration is a processing technique that transforms such different images and displays them in one coordinate system. Matching tells you how images from different measurement methods correspond.

The spectral matching unit 230 performs spectral matching using the shape descriptor vector for the pair of feature points calculated by the feature extraction unit 220. That is, a matrix is formed by measuring geometric similarities with respect to the pair of feature points, and a cluster of feature points having a large geometric similarity is obtained by calculating an eigenvector corresponding to the largest eigenvalue of the matrix.

At this time, the score function E _f for matching is defined as in Equation 3 below.

Where (i, j) and (a, b) denote feature point pairs of model data and image data, respectively, and t _ia and t _jb denote matching strengths of i, a, j, and b, respectively. It has a value of 1 and a value of 0 otherwise. G _{ia : jb} is a potential function that calculates the similarity between pairs of corresponding feature points based on the shape descriptor vector of the feature point pairs at (i, j) and (a, b), as defined in Equation 4 below. do.

When ρ is a penalty constant and v _ij is i and j reliability, v _ij is defined by Equation 5 below. (VAR here means variance.)

When the spectral matching unit 230 calculates the potential function, the spectral matching unit 230 provides a penalty constant ρ for the case where the pairs of feature points of the model data and the image data exist on the same outline. It has a value of 0.8 when (i, j) and (a, b) exist on the same contour, respectively, and a value of 1 otherwise. When feature points are extracted from contours having a constant curvature, many cases may have similar geometric relationships between feature point pairs, and thus, obscure feature point pairs may limit contributions to the potential function through this method. That is, only the pairs of feature points to which the penalty constant is given are configured to form an outline.

The spectral matching unit 230 may measure deformation between pairs of feature points of the model data and the image data. g _ij (a, b) is a vector measuring the deformation between (i, j) and (a, b), and is defined as in Equation 6 below based on the shape descriptor vector for the feature point pair.

Where d _ij is the distance between feature point pairs (i, j), and ε ₁ to ε ₆ are elements of the difference vector of the six-dimensional shape descriptor vector defined in (i, j) and (a, b). When γ is defined, γ is defined by Equation 7 below.

In addition, C _ia , C _jb means i and a, the difference in the local shape histogram between j and b is defined by Equation 8 below.

Here, h _i (k) and h _a (k) mean bins of shape context histograms aligned with respect to the direction of the normal vector at i and a, respectively, for measurements independent of in-plane rotation.

In addition, the score function E _f (Equation 3) may be expressed in the form of a vector and a matrix having the corresponding pairs i and a of feature points as indices as shown in Equation 9 below.

Here, t is defined as an index vector as shown in Equation 10 below.

In addition, G is a potential matrix and is defined as in Equation 11 below.

In the present invention, a spectral matching method is applied to quickly find an index vector t that maximizes the matching score function of Equation (9).

The matching result improvement unit 240 applies a random SAmple consensus (RANSAC) algorithm using the corresponding points calculated by the spectral matching unit 230.

In order to create a theoretical model that describes a phenomenon, we need to gather observation data from the phenomenon. Observation data, however, often contains various types of noise that result from incorrect assumptions about the model or errors in measurement equipment. Since this noise prevents the model from accurately predicting, the RANSAC algorithm provides the answer to the question of how to build the model from observational data that contain data that prevents the prediction of the model parameters.

By applying this RANSAC algorithm to the present invention, the two-dimensional homography transform is estimated and the contour is removed using the symmetry propagation error. In addition, to search for a new corresponding point, for this purpose it is defined by the following equation (12) using a symmetric transfer error as a measure for measuring the error of the corresponding point i, a of the model data and the image data.

Here, p _i and q _a are position vectors of i and a in the homogeneous coordinate system, H denotes a two-dimensional homography transformation matrix, and d (?) Denotes L2-norm.

When the symmetric transfer error satisfies Equation 13 below, the corresponding point is determined as an outline and removed.

Here, d _transfer is defined as a threshold sampling interval equal to the contour sampling interval when the feature extracting unit 220 extracts the feature point. In addition, a new corresponding point is searched by applying a two-dimensional homography transformation matrix calculated by the RANSAC algorithm to feature points of the model data having no corresponding relationship. When the feature point of the model data having no correspondence relation does not have the correspondence relation and the symmetric transfer error with the feature point of the image data does not correspond to Equation 13, it is determined as a new correspondence point. When a plurality of feature points satisfy Equation 13 at the same time, the point having the smallest symmetric transformation error is determined as the corresponding point.

The matching position estimating unit 250 calculates the matching position at the corresponding points determined by the matching result improvement unit 240. The position of the recognized category object is determined as a rectangular area including corresponding points as shown in Equations 14 and 15 below.

Where (x _tl y _tl ) and (x _br , y _br ) are the upper left and lower right vertices of the rectangle, respectively, (x _m and y _m ) are the positions of the corresponding corresponding points, and l is the The feature extraction unit 220 is a constant defined to be equal to the contour sampling interval when the feature point is extracted.

Through this process, the output unit 300 outputs the recognition result of the category object recognition unit 200. The recognition result is characterized in that it is robust to viewpoint changes with respect to other objects belonging to the same category.

3 is a flowchart of a contour-based category object recognition method that is robust to a viewpoint change of an object according to the present invention. Referring to FIG. 3, a contour-based category object recognition method that is robust to a viewpoint change of an object will be described.

When receiving the image data (S100), the category object is recognized from the image data (S200), and the step of recognizing the category object (S200) detects an outline of the image data (S210) and detects a feature point of the outline. After extracting the shape descriptor vector (S220), the spectral matching is performed between the extracted model data and the image data (S230), and the spectral matching result is improved (S220). S240), the method may be subdivided into a step S250 of recognizing a category object by estimating a matching position among the image data. Thereafter, the category object recognition result is output (S300).

In this case, performing spectral matching between the extracted model data and the image data (S230) may further include measuring deformation between the pair of feature points of the model data and the image data. Can be. Further, the detailed description of each of the above steps is the same as that of the contour-based category object device that is robust to the viewpoint change of the above-described object.

The invention also includes a computer readable recording medium having recorded thereon a program for executing the above method.

As described above, although the present invention has been described by way of limited embodiments and drawings, the present invention is not limited thereto and is intended by those skilled in the art to which the present invention pertains. Of course, various modifications and variations are possible within the scope of equivalents of the claims to be described.

100 Data input part 200 Category Object recognition part
210 Contour Detection Unit 220 Feature Extraction Unit
230 Spectral Matching Unit 240 Matching Result Improvement Unit
250 Registration Position Estimation 300 Output

Claims (29)

delete (a) receiving image data;
(b) recognizing a category object among the image data; And
(c) outputting the result of category object recognition;
The step (b)
(d) detecting an outline of the image data;
(e) extracting feature points of the contour to calculate a shape descriptor vector;
(f) performing spectral matching between the model data from which feature points of an outline are extracted and the image data;
(g) improving the spectral matching result; And
(h) recognizing a category object by estimating a matching position among the image data;
The step (d)
And detecting edges by merging edge pixels whose edges are within a preset range and the direction of the normal vector is within a preset range. delete The method of claim 2,
The shape descriptor vector is
Contour-based category object recognition robust to a viewpoint change of an object, characterized by being defined using at least one of a normal vector of the pair of feature points, a distance between the pair of feature points, a curvature of the feature points, or a position vector of the feature points. Way. The method of claim 4, wherein
The shape descriptor vector is
The shape descriptor vector is e _ij, the normal vectors of (i, j) which are pairs of the feature points are respectively θ _i and θ _j , θ _ij is the absolute value of the difference between θ _i and θ _j , and the positions of i and j are respectively (x _i , y _i ), (x _j , y _j ), the distance between feature point pairs (i, j) is d _ij , the curvature of i, j is k _i , k _j, the location of feature point i with respect to feature point _j, respectively The directionality of the feature point i based on the vector is measured as σ _ij , and σ _ij is expressed as

Is calculated as
Equation

Contour-based category object recognition method robust to the change of viewpoint of the object, characterized in that defined by. The method of claim 2,
The step (f)
Measuring deformation between the pair of feature points of the model data and the image data; and a contour-based category object recognition method robust to a change in a viewpoint of an object. The method according to claim 6,
The shape descriptor vector is
A pair of feature descriptors of the model data and the pair of feature points of the image data are (i, j), (a, b), and a shape descriptor vector measuring a deformation between the (i, j) and (a, b), respectively, g _ij (a b), i and a, the difference in the local shape histogram between j and b is the distance between C _ia , C _jb , and feature point pairs (i, j), respectively, d _{ij ,} ε ₁ to ε ₆ is (i, Given each element of the difference vector of the six-dimensional shape descriptor vector defined in j) and (a, b),
Equation

Defined as
Γ is an equation

(δ is a small positive constant value for preventing the case of dividing by zero). The method of claim 2,
The step (f)
A viewpoint change of an object is characterized by obtaining a cluster of feature points having a large geometric similarity by measuring a geometric similarity of the pair of feature points and calculating a eigenvector corresponding to the largest eigenvalue of the matrix. Robust contour-based category object recognition method. The method of claim 8,
The step (f)
When calculating a potential function for measuring geometric similarity with respect to the pair of feature points, changing the viewpoint of the object characterized in that a penalty constant is given to the case where the pair of feature points of the model data and the image data are on the same contour, respectively Robust contour-based category object recognition method. The method of claim 9,
The potential function is
The pairs of feature points of the model data and the image data are (i, j), (a, b), the penalty constant is ρ, and the potential functions are G _{ia: jb} , (i, j) and (a, b) When the vector measuring the deformation between g _ij (a, b), the weight vector w, and the reliability of the i, j is v _ij ,
Equation

Contour-based category object recognition method robust to the change of viewpoint of the object, characterized in that defined by. The method of claim 10,
The reliability of the i, j is v _ij ,
If θ _i , θ _j are normal vectors of feature points i and j,
Equation

Contour-based category object recognition method robust to the viewpoint change of the object, characterized in that calculated by. 12. The method of claim 11,
The step (f)
Contour-based category object recognition method robust to the change in the viewpoint of the object, characterized in that the contour is configured only for the pair of feature points given the penalty constant. The method of claim 2,
The step (g)
A contour-based category object recognition method robust to a change in viewpoint of an object, wherein a random SAmple consensus (RANSAC) algorithm is applied to the corresponding points by the spectral matching. The method of claim 13,
Application of the RANSAC algorithm,
An outline-based category object recognition method that is robust to changes in the viewpoint of an object, by estimating a two-dimensional homography transform and searching for a new corresponding point by removing the contour using a symmetric propagation error. delete A data input unit for receiving image data;
A category object recognition unit recognizing a category object among the image data; And
And an output unit for outputting a recognition result of the category object recognition unit.
The category object recognition unit,
An outline detection unit detecting an outline of the image data;
A feature extraction unit for extracting feature points of the contour to calculate a shape descriptor vector;
A spectral matching unit for performing spectral matching between the extracted model data and the image data of the contour points;
A matching result improvement unit for improving a matching result of the spectral matching unit; And
And a registration position estimator for recognizing a category object by estimating a registration position among the image data.
The contour detection unit,
An edge-based category object recognition apparatus robust to a change in viewpoint of an object, characterized by detecting edges and merging edge pixels whose corners are within a preset range and a normal vector is within a preset range. delete 17. The method of claim 16,
The shape descriptor vector is
Contour-based category object recognition robust to a viewpoint change of an object, characterized by being defined using at least one of a normal vector of the pair of feature points, a distance between the pair of feature points, a curvature of the feature points, or a position vector of the feature points. Device. The method of claim 18,
The shape descriptor vector is
The shape descriptor vector is e _ij, the normal vectors of (i, j) which are pairs of the feature points are respectively θ _i and θ _j , θ _ij is the absolute value of the difference between θ _i and θ _j , and the positions of i and j are respectively (x _i , y _i ), (x _j , y _j ), the distance between feature point pairs (i, j) is d _ij , the curvature of i, j is k _i , k _j, the location of feature point i with respect to feature point _j, respectively The directionality of the feature point i based on the vector is measured as σ _ij , and σ _ij is expressed as

Is calculated as
Equation

Contour-based category object recognition device robust to the change of viewpoint of the object, characterized in that defined by. 17. The method of claim 16,
The spectral matching unit,
A contour-based category object recognition apparatus robust to change of viewpoint of an object, characterized by measuring deformation between pairs of feature points of the model data and the image data. 21. The method of claim 20,
The shape descriptor vector is
A pair of feature descriptors of the model data and the pair of feature points of the image data are (i, j), (a, b), and a shape descriptor vector measuring a deformation between the (i, j) and (a, b), respectively, g _ij (a b), i and a, the difference in the local shape histogram between j and b is the distance between C _ia , C _jb , and feature point pairs (i, j), respectively, d _{ij ,} ε ₁ to ε ₆ is (i, Given each element of the difference vector of the six-dimensional shape descriptor vector defined in j) and (a, b),
Equation

Defined as
Γ is an equation

Contour-based category object recognition apparatus robust to a change in viewpoint of an object, characterized by δ being a small positive constant value to prevent dividing by zero. 17. The method of claim 16,
The spectral matching unit,
A viewpoint change of an object is characterized by obtaining a cluster of feature points having a large geometric similarity by measuring a geometric similarity of the pair of feature points and calculating a eigenvector corresponding to the largest eigenvalue of the matrix. Robust contour-based category object recognition device. The method of claim 22,
The spectral matching unit,
When calculating a potential function for measuring geometric similarity with respect to the pair of feature points, changing the viewpoint of the object characterized in that a penalty constant is given to the case where the pair of feature points of the model data and the image data are on the same contour, respectively Robust contour-based category object recognition device. 24. The method of claim 23,
The potential function is
The pairs of feature points of the model data and the image data are (i, j), (a, b), the penalty constant is ρ, and the potential functions are G _{ia: jb} , (i, j) and (a, b) When the vector measuring the deformation between g _ij (a, b), the weight vector w, and the reliability of the i, j is v _ij ,
Equation

Contour-based category object recognition apparatus robust to the change of viewpoint of the object, characterized in that defined by. 25. The method of claim 24,
The reliability v _ij of i and j is
If θ _i , θ _j are normal vectors of feature points i and j,
Equation

Contour-based category object recognition apparatus robust to the change of viewpoint of the object, characterized in that calculated by. The method of claim 25,
The spectral matching unit,
Contour-based category object recognition apparatus robust to changing the viewpoint of the object, characterized in that the contour is configured only for the pair of feature points given the penalty constant. 17. The method of claim 16,
The matching result improvement unit,
A contour-based category object recognition apparatus robust to changing the viewpoint of an object, wherein a random SAmple consensus (RANSAC) algorithm is applied to the corresponding points due to the spectral matching. The method of claim 27,
Application of the RANSAC algorithm,
A contour-based category object recognition apparatus robust to change in the viewpoint of an object characterized by estimating a two-dimensional homography transform and searching for a new corresponding point by removing the contour using a symmetric propagation error. delete

Images