KR20130032990A - Method for detecting grasping points using category recognition and computer readable record medium thereof - Google Patents

Abstract

PURPOSE: A method for generating a nip point by range recognition and a computer-readable recording medium including a program for the method are provided to recognize and grip objects in various shapes in the same range by generating a nip point by recognizing the objects in the same range. CONSTITUTION: A CPU(Central Processing Unit) explores an object gripping direction after obtaining a 3D outline from image data(S10,S20). The CPU extracts a gripping cue for each set gripping cue and generates a candidate gripping area for each gripping cue(S30). The CPU selects one of candidate pairs in an optimal gripping area after the points of nip point candidate pairs are extracted(S40,S50). The CPU generates a nip point manifold(S60). [Reference numerals] (AA) Start; (BB) End; (S10) Step of obtaining a 3D outline from image data; (S20) Step of searching a basic or function grasping type; (S30) Step of generating a candidate grasping area; (S40) Step of extracting grasping point candidate pairs; (S50) Step of selecting a grasping point pair; (S60) Step of generating a grasping point manifold;

Description

TECHNICAL FIELD [0001] The present invention relates to a method for generating a wave point by category recognition, and a computer readable recording medium storing a program for the method. BACKGROUND OF THE INVENTION [0002]

More particularly, the present invention relates to a method of creating a wave point capable of holding an object by using a category recognition technique to generate a wave point of various objects, And a computer-readable recording medium storing a program for the method.

As one of technologies that are being studied in robot technology, there has been conventionally been studied a technology in which a robot recognizes an object by itself and moves an object to a desired position by using a robot arm.

This includes object recognition technology and techniques for analyzing recognized objects so that robot arms can grasp them. Object recognition technology is a core technology for data acquisition, recognition, position and attitude estimation, and modeling of three-dimensional objects and environments, and has been studied extensively in the field of computer science as well as robots.

As part of this research, Lowe ( David G. Lowe , " Distinctive Image Features (2), pp. 91-110, 2004.) proposed an object recognition method that is robust against scale and viewpoint changes based on appearance information of an object However, in the case of the object recognition method described above, there is a problem that it is not suitable for recognition of a categorical object which is suitable for recognizing an individual object having a specific texture component and belongs to the same category but has various texture characteristics.

In the conventional technique of gripping an object recognized using a robot arm, data on a point where a specific object is to be gripped is presented to the robot, and the robot is learned When an object similar to the input data is recognized, the robot merely generates a wave point in accordance with the input data. Therefore, even if an object belonging to the same category does not resemble the specific object provided with the data, the robot can not grasp the object by itself.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problems, and it is an object of the present invention to provide a robot that can recognize an object belonging to the same category by itself and generate a wave point from a recognized object, And a computer-readable recording medium on which a program for the method is recorded.

As a means for solving the above-mentioned problems,

(1) acquiring image data and 3D outline of a target object; (2) searching for a basic gripping type and a functional gripping type of the object using the image data and the 3D outline; (3) extracting a grasping cue for each of the set types of grasping, and creating at least one candidate grasping region through the grasping cue; (4) extracting wave point candidate pairs for each of the candidate finger regions; And (5) selecting one of the candidate regions from among the candidate regions, and selecting one of the pair of candidate regions in the selected region, The basic pager type is searched for an image pager type and a front pager type, a side pager type, and a lift-up pager type of the target object, The method according to claim 1, wherein, in the step (3), searching for the function type is performed by recognizing the category of the object by the category object recognition method A method of generating a wave point by recognizing a category is provided.

Here, after step (5), (6) combining the wave point pairs selected for each of the grasping types to generate one wave point manifold It provides a method of generating point.

Here, the categorized object recognition method may further include: (2-1) detecting a contour of the image data; (2-2) extracting feature points of the contour line and calculating a shape descriptor vector; (2-3) performing spectral matching between the extracted model data and the image data; (2-4) improving the spectral matching result; And (2-5) estimating a matching position from the image data to recognize a categorical object. According to the present invention, there is provided a method of generating a wave point by category recognition.

Here, in the step (2-1), contour lines are detected by detecting edges, and edge pixels within the predetermined range and the direction of the normal vector within the predetermined range are merged. The method comprising:

Here, the shape descriptor vector is 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. It provides a method of generating point.

Here, the step (2-3) may include measuring distortion between a pair of feature points of the model data and the image data.

In the step (2-3), a matrix is constructed by measuring geometric similarities of pairs of the minutiae points, and an eigenvector corresponding to the largest eigenvalue of the matrix is calculated to determine the feature points having large geometric similarities The method of generating a wave point by category recognition is provided.

Here, in the step (2-4), a RANSAC algorithm is applied to the corresponding points by the spectral matching, and a method of generating a wave point by category recognition is provided.

Here, the application of the RANSAC algorithm provides a method of generating a wave point by category recognition, in which a two-dimensional homography transformation is estimated and a contour is removed using a symmetric propagation error to search for a new correspondence point.

Here, the image data is acquired using a stereo camera. The method of generating a wave point by category recognition is provided.

Here, the extraction of the grasping cue may include generating a 2D projection image from the 3D contour, and forming the 2D projection image into a polyline of a peripheral portion and extracting the 2D projection image. It provides a method of generating point.

The generating of the candidate grasping region may include forming a skeleton segment in the grasping cue so as to surround a portion of the parchment located parallel to the skeleton segment, And generating a wave point based on the category recognition.

Here, the extracting of the candidate pair may include dividing the 3D contour into at least one grasping layer for each of the candidate grasping areas, and selecting, from among all the points belonging to the grasping layers, Of the robot arm are extracted as a candidate pair, and the structure of the robot arm is given by previously stored data or separately inputted data.

Here, the selecting the one of the gripping area in the step (5), G _Reg is the candidate gripping area, ht is the height or length values, std is gripping point candidate pair in the candidate gripping area of the candidate gripping area Std _th is an average value of std, and G ^* _Reg is a standard deviation indicating a distance between the two _grids ,

에Equation

on

The method comprising the steps of: (a) generating a wave point by a category recognition;

A computer-readable recording medium on which a program for a wave point generating method by any one of the above categories is recorded is provided.

According to the present invention, since the robot itself can recognize an object belonging to the same category and generate wave points, a separate learning step becomes unnecessary, thereby recognizing and holding various objects belonging to the same category but having different shapes There is an effect that can be done.

1 is a flowchart of a wave point generating method according to the category recognition of the present invention,
Figure 2 is a block diagram of a central processing unit that processes the method of Figure 1;
3 is a flowchart of a categorical object recognition method,
4 is an illustration of an example of a gripping type of a target object,
FIG. 5 is an example of a 2D projection of a 3D contour of a target object,
FIG. 6 is a diagram illustrating an example of parchyque extraction and generation of a grip region,
Fig. 7 is a view showing an example of generation of a parchment part using a skeleton,
8 is a diagram illustrating an example of a process of gripping a target object by a robot operation,
Fig. 9 is a graph showing various object grasping experiment data,
Fig. 10 shows various experimental data of gripping objects. Fig.

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.

FIG. 2 is a block diagram of a central processing unit for processing a wave point generating method by category recognition, and FIG. 3 is a flowchart of a category object recognizing method .

As shown in FIG. 1, a method of generating a wave point by category recognition according to an embodiment of the present invention includes the steps of (1) acquiring image data of a target object and obtaining a 3D outline (S10) and (2) Searching for a basic grasp type and a function grasp type of the target object using the 3D outline (S20); and (3) extracting a grasping cue for each of the set grasp types, A step (S30) of generating a candidate grasping region and (4) a step (S40) of extracting wave point candidate pairs for each of the candidate grasping regions and (5) (S50) selecting one of the pair of finger points from among the candidate pairs in the selected optimal finger area, and (6) combining the pair of finger points selected for each of the finger type The far point is configured to include a step (S60) for generating a manifold.

Here, the step (S10) of receiving the image data and 3D outline of the object (1) receives the shape of the object transmitted from the image acquiring device such as a camera in the central processing unit 100 as image data , From which a 3D contour is obtained. The camera preferably uses a stereo camera capable of measuring the distance to each part of the object, thereby obtaining a 3D outline from the image data of the object. Specifically, a 3D contour can be obtained by extracting an edge portion from the image data and superimposing it on the edge portion of the filtered image of standard deviation. The 3D outline is obtained by calculating the 3D point from each pixel of the 2D contour using the distance information to each part obtained by the stereo camera.

Next, in step (2), the central processing unit 100 searches for a direction in which the object can be gripped by dividing the object into a basic grip type and a function grip type (S20).

In the case of the basic phage type, generally, an object is gripped and detected as an intact phage type, a front phage type, a side phage type, and a lift-up phage type. The recognition direction gripping type refers to a type of gripping in a direction when an image of a target object is acquired by a camera or the like as shown by a green line in Fig. The front grip type is of the type that grasps in the front direction as shown by the blue line in Fig. The side gripping type is a type of gripping from the side as shown by the black line in Fig. Although not shown in the opposite side, the side gripping type is naturally also possible on the opposite side, so that it can be composed of two gripping types. In the case of the lift-up phage type, as shown by the red line in Fig. 4, it is a grip type when the object is lifted. At this time, the lift-up phage type can be divided into lifting the object in the major axis direction and lifting it in the minor axis direction.

In the case of the functional phage type, it means that the category of the object is recognized and grasped by the category object recognition method. For example, if there is a cup with a handle, for a cup belonging to that category, the handle is searched as a functional part of the object and a function grip type for the handle is created separately from the basic grip type. To do this, a categorical object recognition method is applied.

As shown in FIG. 3, the category object recognizing method may include the steps of (2-1) detecting the contour of the image data (S210) and (2-2) extracting the characteristic points of the contour line to calculate a shape descriptor vector Steps (S220) and (2-3) Perform spectral matching between the extracted model data and the image data (S230) and (2-4) to improve the spectral matching result (S240) and (2-5) recognizing a category object by estimating a matching position from the image data (S250).

The categorical object recognition method may be implemented through a categorical object recognition unit 200 included in the central processing unit 100. The categorical object recognition unit 200 recognizes the categorical object recognition method using the outline (2-2) a feature extracting unit 220 for extracting feature points of a contour of the step and calculating a shape descriptor vector, (2-3) extracting the feature points of the outline of the model data A spectral matching unit 230 for performing spectral matching between the image data and the image data, a matching result improving unit 240 for improving a matching result of the spectral matching unit 230 in step (2-4) And a matching position estimating unit 250 for estimating a matching position among the image data of step 5) and recognizing a categorical object.

 In the following, each step of the categorical object recognition method will be described.

The step (2-1) of detecting the contour of the image data by the contour detecting unit 210 (S210) is a step of detecting the edge and detecting the edge within the preset range and the direction of the normal vector within the predetermined range The edge pixels are merged to detect the contour line. At this time, contour detection can use a Canny edge detector. Edge pixels within a predetermined range and normal vectors within a predetermined range can be found.

In step (2-2), the characteristic extraction unit 220 samples the contour detected by the contour detection unit 210, extracts the characteristic points, and calculates a shape descriptor vector for the pair of characteristic points (S220).

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 the following equation (1).

The distance between when the normal vector d of the pair of the feature point (i, j) to Θ _i and Θ _j, wherein Θ _ij is the absolute value, d _ij of Θi and Θj car feature point pair (i, j), k _i and k _j are curvatures of the i and j, and _ij is the directionality of the feature point i based on the position vector of the feature point i with respect to the feature point j, which is defined by the following equation (2).

Here, (x _i , y _i ) and (x _j , y _j ) represent positions of i and j, respectively. In the same way, we define σ _ji so that Θ _i , Θ _j, and σ _ij , σ _ji . Figure 2 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, Θ _j, and σ _ij , σ _ji .

The shape descriptor vector for the feature point pair is independent of the orientation of the individual feature points by using a normal vector or the like. Also, since it is defined based on the relative physical quantity between the pair of feature points, it is invariant to parallel movement and rotation in the plane, and operates robustly to the ape transform. Thus, it enables recognition of categorical objects at various viewpoints without assumptions about the viewpoint of the object.

On the other hand, when a scene or an object is photographed at a different time or viewpoint in the computer vision, the image is obtained in a different coordinate system. Image registration is a processing technique that transforms these different images into one coordinate system. Through matching, we can see how images obtained through different measurement methods correspond.

In the step (2-3), spectral matching is performed using the shape descriptor vector for the pair of feature points calculated by the feature extraction unit 220 in the spectral matching unit 230 (S230). That is, a matrix is constructed by measuring the geometric similarity to the pair of feature points, and an eigenvector corresponding to the largest eigenvalue of the matrix is calculated to obtain a cluster of feature points having a large geometric similarity.

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

Here, (i, j) and (a, b) represent feature point pairs of the model data and the image data, respectively, and t _ia and t _jb denote the matching strengths of i and a, j and b, When the value is 1, the value is 0 otherwise. G _{ia : jb} is a potential function for calculating a similarity between pairs of feature points corresponding to each other based on a shape descriptor vector for a pair of feature points in (i, j) and (a, b) do.

Is a penalty constant, and v _ij is the reliability of i, j, v _ij is defined by Equation (5). (Where VAR means dispersion).

When calculating the potential function, the spectral matching unit 230 gives a penalty constant p for a case where the pair of feature points of the model data and the image data exist on the same contour line, respectively. (i, j) and (a, b) are on the same contour line, they have a value of 0.8 and the other cases have a value of 1. If a feature point is extracted from a contour having a constant curvature, a plurality of geometric relationships between the pair of feature points may occur in a similar manner. Thus, the contribution of the ambiguous feature point pairs to the potential function can be limited by this method. That is, contour lines are formed only for pairs of feature points to which the penalty constant is assigned.

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

Here, d _ij denotes a distance between the pair of feature points (i, j), and ε ₁ to ε ₆ denote the distance between each element of the difference vector of the 6-dimensional shape descriptor vector defined by (i, j) , The gamma is defined by the following equation (7).

In addition, C _ia and C _jb denote the difference of the regional form histogram between i and a, j and b, and are defined by the following equation (8).

Where h _i (k) and h _a (k) denote the bins of the 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.

Further, the score function E _f (Equation 3) can be expressed by the following equation (9) in the form of a matrix and a vector having the corresponding pair i and a of the minutiae as indexes.

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

The above G is a potential matrix and is defined by Equation (11) below.

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

The step (2-4) is a step (S240) of applying a RANSAC algorithm consensus (RANSAC) algorithm using the corresponding points calculated by the spectral matching unit 230 in the matching result improving unit 240.

If you want to create a theoretical model that describes a phenomenon, you have to collect observational data from the phenomena. Observation data, however, contain various types of noise caused by incorrect assumptions about the model and errors in measuring equipment. This noise prevents accurate prediction of the model, so the RANSAC algorithm provides the answer to the question of how to construct a model from observed data containing data that interferes with the prediction of the model parameters.

By applying the RANSAC algorithm to the present invention, a two-dimensional homography transformation is estimated and a contour is removed using a symmetric transfer error. Also, a new correspondence point is searched. For this purpose, a symmetric transfer error is used as a scale for measuring an error with respect to the corresponding point i, a of the model data and the image data, and is defined by the following equation (12).

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

When the symmetric transfer error satisfies Equation (13), the corresponding point is determined to be a contour line and removed.

Here, d _transfer is defined as a threshold value equal to the contour sampling interval when the feature extracting unit 220 extracts feature points. In addition, a new correspondence point is searched by applying a two - dimensional homography transformation matrix computed by the RANSAC algorithm to minutiae points of model data that do not have a corresponding relationship. When the symmetric transfer error between the feature points of the model data that do not have the corresponding relationship and the feature points of the image data that do not have the corresponding relationship satisfies Equation (13), the new correspondence point is determined. When a plurality of minutiae points simultaneously satisfy Equation (13), a point having the smallest symmetric conversion error is determined as a corresponding point.

The step (2-5) is a step of calculating a matching position at corresponding points determined by the matching result improving unit 240 in the matching position estimating unit 250. [ The position of the recognized category object is determined as a rectangular area including corresponding points as shown in the following equations (14) and (15).

Here, (x _tl y _t1 ), (x _br , y _br ) means the upper left vertex and the lower right vertex of the rectangle, (x _m , y _m ) Is a constant defined in the same manner as the contour sampling interval when the feature extracting unit 220 extracts feature points.

Through this process, objects of different shapes belonging to the same category can be recognized. As an example of the handle of the cup, it is possible to calculate the axis through PCA (Principal Component Analysis) using the above-described recognition of the handle and the 3D contour point of the handle.

Next, (3) a grasping cue is extracted for each of the set grasping types, and a candidate grasping region for each of the grasping cues is generated (S30). (3) step (S30) because the basic grasp types intact, front, side, and lift-up have been searched through (1) S20 and (2) The candidate finger area is created through the above process.

First, in the central processing unit 100, a grasping cue is extracted for each of the detected types. For this purpose, a 2D projection image is obtained from the 3D contour as shown in FIGS. 4 and 5 by the grip type. 6 (a) and 6 (b) illustrate image data in the case of the image recognition direction (Intact) type, as shown in FIG. 6 as an embodiment of a specific type, (C) shows a 2D projection image obtained therefrom. (C) In the case of the intact type, the projection direction becomes the optical axis of the camera. A grasping cue is extracted from the 2D projection image by simplifying the 2D projection image into a polyline at the periphery as shown in (e) of FIG. 6, by fitting and covering the outer portion as shown in (d) . In the case of FIG. 6, only the image recognition direction (Intact) is shown. However, the front grasping type, the side grasping type, the lift-up grasping type, and the function grasping type are respectively extracted as described above. The extracted grasping cue represents the outline of the edge portion in place of the detailed image data in which the curvature is revealed as it is in the object.

Then, the central processing unit 100 acquires a 2D skeleton image as shown in FIG. 7 (a) from the extracted parcquired image and converts the skeleton image into a skeleton image of each skeleton having no node as shown in FIG. 7 (b) Decompose into a skeleton segment. Except for very short or trivial portions of these skeleton segments, this simplifies them. This simplification of the skeleton segment is described in the paper <Dougla, D. and T. Peucker (1973). Algorithm for the reduction of number of points required to represent a digitized line or its caricature. The Canadian Cartographer, 10 (2), 112-122.

Then, as shown in Fig. 7 (c), the simplified simplified skeleton segment is used to select a pair of roughly parallel Grasping Cue portions (red portions) from the simplified skeleton segment. When a rectangular area surrounding the selected parquiche pair is generated, a candidate grasping region as shown in Figs. 5 (f) and 5 (g) is generated in this area.

On the other hand, when there are a plurality of simplified skeleton segments, a plurality of candidate grasping regions can be generated in one of the grasping types.

The subsequent step (4) is a step (S30) of extracting points of the wave point candidate pair which are likely to be gripping points for each of the candidate grip regions generated in the step (S30) in the central processing unit 100. [

Various methods are available for this purpose. In the present invention, a grasp plane dividing the 3D contour into a plurality of grasping layers is defined and extracted. The 3D outline of the 3D outline of the image data is generated by one or more candidate grasping regions for each of the respective types of grasping, and the 3D outline is divided into a plurality of grasping layers according to the respective grasping planes, . That is, the gripping plane is set for each of the image recognition direction (Intact) type, the front side type, the side side type, the lift type, and the function type, and is divided into several different grasping layers. Therefore, in the case of the other grasping types except for the lift-up grasping type, the grasp plane is perpendicular to the longitudinal axis of the grasping region. In the case of the lift-up grasping type, Is perpendicular to the horizontal axis of the grasping region. This can be identified through Principal Component Analysis (PCA).

A set of 3D contour points between two parallel grasp planes forms a grasping layer.

Here, in order to extract the wave point candidate pair, the robot mechanic factor such as the feasible width and depth of the robot finger is considered. That is, as shown in FIG. 4 or 8, since the width of the object to be picked up by the robot is limited and the length of the robot finger is limited, the robot can not actually pick up all points inside the gripping layer to be. Therefore, considering the robot finger width and the like mechanically, points that can be picked up are extracted as a pair of point candidates. The mechanistic information of such a robot can be inputted or stored as separate data.

(5) selecting one of the grasping regions from among the candidate grasping regions for each of the grasping types in the central processing unit 100, and selecting one of the pair of grasps among the candidate pairs in the selected optimal grasping region (S50).

For this purpose, a method of selecting a grasping region and a pair of wave points may be variously presented, such as selecting randomly. However, as a preferred method in the present invention, a grasping region is selected and a pair of wave point candidates existing within the grasping region Provides a way to select one wave point pair.

Since a plurality of candidate grasping regions may be generated in one of the grasping types (for example, the front grasping type) as described in (3) step S30, Select one finger area.

As one of the methods, i) the length (height) of the candidate gripping region is the longest, and ii) the gripping region having the smallest standard deviation of the candidate pairs in the candidate gripping region is selected.

Specifically, in the case of the intact gripping type, the front gripping type, the side gripping type, or the function gripping type among the gripping types, the vertical axis lengths of the respective candidate gripping areas are compared. Specifically, in the case of the lift-up gripping type, It is possible to compare the lengths and to search for each of the holding types having the longest length by each of the holding types. For example, in the case of the frontal grip type, the most stable gripping area is obtained when the gripping area having the longest vertical axis (height) is actually picked up by the robot.

And ii) is a condition for searching for the smallest standard deviation of the candidate pairs belonging to the candidate gripping region. Here, the standard deviation represents the distance between the candidate pairs, and the smallest standard deviation means that the robot is less likely to slip off the object when picked up.

Therefore, the optimum grasping region is selected among the candidate grasping regions in consideration of the conditions of i) and ii).

This optimal method of selecting a phage region can be expressed by the following equation.

The _Greg represents a holding region candidate including a parameter ht, which is a height or a length value indicating the number of gripping layers. and std is the standard deviation, which represents the distance between pairs of wave point candidates in the grip region. std _th represents the average value of the standard deviation. The G ^* _Reg represents an optimal grip area that can be stably and efficiently gripped.

Then, one pair of wave points is extracted from among a plurality of candidate pairs existing in the optimal grasp region selected in the above. However, in the present invention, the candidate pair of the gripping layer positioned at the very middle of the gripping layers belonging to the gripping area is finally selected as one pair of the pointing point.

Next, (6) combining the wave-point pairs selected for each of the gripping types in the central processing unit 100 to generate one wave-branch manifold S60.

Since the wave point pairs are extracted one by one for each of the basic grasp types (intact, front, side, lift-up) and function grasp type, and finally five wave point pairs can be extracted, Thereby generating a branch manifold.

Meanwhile, the present invention provides a computer-readable recording medium on which a program for a wave point generation method by category recognition by the above-described (1) step (S10) to (6) step (S60) is recorded.

FIG. 8 shows test results by controlling the robot according to the method of generating a wave point by category recognition according to the present invention. FIG. 9 shows test results thereof. FIG. 10 shows test results . As shown in FIGS. 9 to 10, according to the present invention, there is an effect that the robot itself can recognize and grasp the shape of an object having various shapes, even if the grip is not learned about the shape of the object.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is to be understood that the invention is not limited to the disclosed exemplary embodiments. It is to be understood that various modifications and changes may be made without departing from the scope of the appended claims.

100: Central processing unit 200: Category object recognition unit
210: contour detection unit 220: feature extraction unit
230: spectral matching unit 240: matching result improving unit
250: matching position estimating unit

Claims (15)

(1) acquiring image data and 3D outline of a target object;
(2) searching for a basic gripping type and a functional gripping type of the object using the image data and the 3D outline;
(3) extracting a grasping cue for each of the set types of grasping, and creating at least one candidate grasping region through the grasping cue;
(4) extracting wave point candidate pairs for each of the candidate finger regions; And
(5) selecting one of the candidate regions from among the candidate regions in each of the plurality of candidate regions, and selecting one of the candidate regions in the selected region,
In the step (2), the basic phage type may include an Intag phage type, a front phage type, a side phage type, and a lift-up phage type, And the function type is detected on the basis of the function information of the previously stored object,
The method of claim 1, wherein, in the step (3), searching for the functional grip type is performed by recognizing the category of the object by the category object recognition method. The method according to claim 1,
After the step (5)
(6) combining the wave point pairs selected for each of the plurality of grasping types to generate one wave point manifold. The method according to claim 1,
The categorical object recognition method includes:
(2-1) detecting a contour of the image data;
(2-2) extracting feature points of the contour line and calculating a shape descriptor vector;
(2-3) performing spectral matching between the extracted model data and the image data;
(2-4) improving the spectral matching result; And
(2-5) A method of generating a wave point by category recognition, comprising the steps of: estimating a matching position from among the image data and recognizing a category object. The method of claim 3,
The step (2-1)
Wherein edge detection is performed by detecting an edge and merging edge pixels whose edge is within a predetermined range and direction of a normal vector is within a predetermined range. The method of claim 3,
The shape descriptor vector,
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. The method of claim 3,
The step (2-3)
And measuring deformation between a pair of feature points of the model data and the image data. The method of claim 3,
The step (2-3)
A matrix of feature points is calculated by measuring the geometric similarity of the pair of feature points and an eigenvector corresponding to the largest eigenvalue of the matrix is calculated to obtain a cluster of feature points having large geometric similarities, How to create wave points. The method of claim 3,
The step (2-4)
And a RANSAC algorithm consensus is applied to the corresponding points by the spectral matching. 9. The method of claim 8,
The application of the RANSAC algorithm,
Dimensional homography transformation and removing a contour using a symmetric propagation error to search for a new correspondence point. The method according to claim 1,
Wherein the image data is acquired using a stereo camera. &Lt; RTI ID = 0.0 &gt; 11. &lt; / RTI &gt; The method according to claim 1,
In extracting the grasping cue,
Generating a 2D projection image from the 3D contour, and forming the 2D projection image into a polyline of a periphery and extracting the 2D projection image. 12. The method of claim 11,
The generating of the candidate grasping area comprises:
Wherein a skeleton segment is formed in the grasping cue to create a rectangular area so as to surround a portion of the parchment located parallel to the skeleton segment. A method of generating a wave point by means of. 13. The method of claim 12,
The step of extracting the candidate pair
Dividing the 3D contour into at least one grasping layer for each of the candidate grasping regions and extracting only pairs of points that can be actually grasped in the structure of the robot arm among all the points belonging to the grasping layers as candidate pairs With features,
Wherein the structure of the robot arm is given by stored data or separately inputted data. 14. The method of claim 13,
In the step (5), selecting one of the grip areas may include:
G _Reg is the candidate gripping area, ht is the height or length of the said candidate gripping area, std is the standard deviation, std _th is the average value of the std that is the distance between the gripping point candidate pair in the candidate gripping area, G ^* _Reg Is the one holding region,
Equation

on
And selecting one of the plurality of wave points by category recognition. A computer-readable recording medium recording a program for a wave point generating method according to any one of claims 1 to 14.

Images