JPH04119475A - Three-dimensional shape identifying device - Google Patents

Abstract

PURPOSE:To enable stable recognition even when error is contained in input date by the noise of a measuring system or the like by handling a vector expressing the feature quantity of a three-dimensional object as a three- dimensional vector amount. CONSTITUTION:This device is equipped with a feature extraction part 10, feature vector corresponding part 20, standard pattern storage part 30, correlative coefficient decision part 40 and pattern decision part 50. First of all, the three- dimensional image of the object is observed and the three-dimensional coordinates of an object surface are calculated. According to obtained three- dimensional data, the feature quantity (such as the normal direction of the local part of the object surface or the like) to be expressed by a direction at the local part is extracted concerning the entire object. Since each feature quantity can be expressed by a unit vector, the object can be expressed by the set or unit vectors. The degree of similarity is calculated between the expression of the unit vectors obtained in this way and an expression to be obtained by a standard pattern recognizing the kind of the object in advance, and the object observed with the pattern having the maximum degree of similarity is recognized. Thus, the various objects can be stably identified.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention is a three-dimensional shape recognition method used when industrial robots, automatic monitoring systems, etc. recognize objects in space (parts, people, etc.). It is related to the device.

[Conventional technology]

Conventionally, in the recognition of three-dimensional objects, an extended Gaussian sphere representation is used as a feature representation of the object, and a matching measure is calculated between the extended Gaussian sphere representation obtained from input data and that obtained from a standard pattern. A method is used in which the object with the largest value is recognized (Reference 1).

When calculating the matching scale, directional data that originally has vector quantities is handled one-dimensionally.

That is, as shown in Figure 5 (A), input data (unit vectors) given as vector quantities are correlated radially from the center of one sphere, and the intersection point on the surface of one sphere is found. The distribution density is plotted on an axis that is a one-dimensional expansion of one spherical surface as shown in FIG. 5(B). Then, matching is performed based on the plotted results.

In addition, the wind direction, atmospheric pressure distribution, and the direction of magnetization of minerals are described by directional data (a set of unit vectors). That is, 7 These data are expressed as a point 1 on a unit circle or a distribution of points on a unit gem. Therefore, data such as weather data for a certain region or the direction of magnetization of minerals at multiple locations can be described by a set of unit vectors. Figure 6 (A)
(B) explains the situation in which the degree of similarity regarding the wind direction is determined, and the distribution in the unit beads is determined using the arrow shown in the figure as a unit vector, and the degree of similarity is determined.

Therefore, the wind direction at a certain point can be compared from season to season, or the similarities in atmospheric pressure distribution at multiple points can be investigated.

(Problem to be solved by the invention) In the conventional case, when the input data includes fluctuations, a large error occurs due to the limit of the matching range.Furthermore, when the object includes rotation, the same object This requires repeated calculations, increasing the calculation cost.Therefore, it cannot be said that the performance is sufficient for practical systems.

Furthermore, unlike data on a normal Euclidean space, directional data cannot be applied with a normal distance measure or correlation similarity (Reference 3). However, the 5 pherical mentioned below
This is possible using correlation. This scale treats directional data as having three-dimensional values, and can continuously express the degree of similarity or difference in directional data (Reference 2).

The present invention uses a camera and a computer to automate the identification of objects, especially the identification of parts in the assembly process, which previously had to be seen by humans.
The purpose of this invention is to realize a device that can stably identify two different objects by using the correlation coefficient according to n).

[References] 1), 1. J, Little, “Determi
ning 0bject Attitude from
Extended Gaussian Images
, Proc, of9th International
nal Joint Conference o
f8 tificial intelligence
, pp960-963 (1985) 2) N,
1. Fisher and A, J, 1. ee
Correlation Coefficients
for Random Variables o
n aUnit 5phere or Hypers
here, Biometricavol-73,
pp159-L64. (1986) 3) Yasutoshi Baba 2nd Masami Tanemura, “Correlation of Directional Data”, Materials for the 47th Japan Statistical Alloy National Conference, pp. 35-36
．． (1979, 7) Means for Solving 1 Problem] In the present invention, this scale is used for three-dimensional object recognition. There are various ways to represent three-dimensional objects, but we will focus on the fact that they are described using directional data. A typical example of this is a Gaussian sphere representation, which is a mapping of the normal image of each surface to a unit stock, but it is not limited to this. Therefore, the problem of three-dimensional object recognition can be reduced to the problem of constructing the similarity between directional data. Specifically, first, a three-dimensional image of the object is observed, and the three-dimensional coordinates of the object's surface are determined. From the obtained three-dimensional data 2
Features that are expressed in direction in a local portion (such as the normal direction of a local portion of the object surface) are extracted for the entire object. Since each feature can be represented by a unit vector, an object can be represented by a set of unit vectors. The similarity measure mentioned earlier is calculated between the unit vector representation obtained in this way and the representation obtained from the standard pattern whose type of object is known in advance, and the observed object is determined using the pattern that maximizes the similarity measure. be made to recognize.

[Operation] When using the similarity measure in Reference 2, the feature quantity expressions of the object and standard pattern are expressed by χ and Y (where X= (X+,
,,,,X,,),Y= (Y,,,,,Y,)χ8
．． Let Y be a unit vector), then the order of n vectors in X and Y is X, , , , X, , and Y.

, , Yr - it is necessary to decide at will. this is.

This corresponds to determining which vector in a set of vectors obtained as feature quantities of an object corresponds to which vector in a set of feature vectors of a standard pattern calculated in advance. Therefore, here we pay attention to the fact that the end points of a unit vector are distributed on a spherical surface with a radius of 1.
While rotating one of the data, the feature values are associated so that the sum of the distances between the respective corresponding end points becomes the smallest.

Thereby, the maximum correlation coefficient calculated between the input pattern and the standard pattern can be found.

〔Example〕

FIG. 1 is a block diagram showing an embodiment of the present invention.

In the figure, reference numeral 10 represents a feature extraction section, 20 a feature vector association section, 30 a standard pattern storage section 40 a correlation coefficient determination section, and 50 a pattern determination section.

When three-dimensional data of an object (such as a distance image of the object) is input, the feature extraction unit 10 stores the data and calculates the three-dimensional coordinates of each point on the object. Then, as an example of a feature amount expressed in direction, the normal vector of the surface is determined for each certain area.

FIG. 2 shows that when the results of applying unit vectors in the normal direction to the surfaces of two cones are associated with a sphere, a set of unit vectors unique to one cone can be obtained as a feature quantity.

Arrow a in the figure is a unit vector on the side of the cone,
The arrow mark is the unit vector of the base of the cone. This representation corresponds to an extended Gaussian sphere representation in which the length of each vector corresponds to the number of unit vectors. For these unit vectors, find the coordinates of the end point of the unit vector,
be made to remember. The results are sent to the feature vector matching section 20.

Since the end points are distributed on the unit spherical surface, the feature vector matching section 20 rotates one of the spherical surfaces to minimize the sum of the distances between the corresponding points. A correspondence is made between the standard data stored in the database and the human data vector.

The standard pattern storage unit 30 stores feature amounts expressed in the same direction as those determined by the feature extraction unit 10 for objects whose types are known in advance.

The correlation coefficient determining unit 40 uses the vector data representing the feature amounts sent from the feature vector matching unit 20 and also uses the correspondence between the five-person cover turn and the standard pattern, and based on them, A similarity measure based on 5pherical correlation is calculated between the input pattern and the standard pattern.

The pattern determining section 50 accumulates the degrees of similarity sent from the correlation coefficient determining section 40 for each standard pattern and the input pattern, and determines the maximum similarity as the category to which the human kavakoon belongs.

Here, the correlation coefficient determining section 40 will be described.

Now, the feature amount obtained from the input data is χ= (X, ,
X2. ,,,,X,l) The feature amount of the standard data is Y= (Y,,Y2.,,,,Y, )(X,,Y,(
i=1.22. , , n) is a unit vector), then
Various correlation amounts between the two direction data X and Y have been proposed. - As an example, use the following definition by Fisher (Reference 2).

Now the two correlation matrix S xv When.

It has been proposed that ρ calculated as ρ = S xy/ (S xx S yy) ''' is used as a measure of correlation.
tA = 1), it becomes ρ-1. In other words,
When the data is rotated and the end points of vectors representing all feature quantities match, ρ=1, which is rotation-invariant and independent of the coordinate system.

2>-1<ρ/, 1 3) It has been confirmed that if X and Y are independent, the property of ρ-〇 is not satisfied, and it can be seen that the property required for a measure of the amount of correlation is satisfied.

Here, if the subscripts of X and Y are determined based on the characterization determined by the feature vector matching unit 20.

The correlation coefficient determined earlier can be determined.

Next, the processing performed by the feature vector matching unit 20 will be explained in detail. As mentioned above, the feature amount expressed in the direction calculated by the feature extractor 10 can be expressed as a unit vector, so the end point of the unit vector exists on the unit stock.

When determining the correlation coefficient, it is necessary to make a correspondence between each vector in the input data and the vector in the standard data.

FIG. 3 shows a manner in which the end point of a unit vector corresponding to input data is associated with the end point of a unit vector corresponding to standard data. Either unit stock is rotated to obtain a more accurate correspondence overall.

Considering that the correlation coefficient calculated by the correlation coefficient determination unit 40 is invariant to rotation, the correlation is performed according to the following procedure. In this case, the nine spheres are treated as a polyhedron as shown in FIG.

1) Divide the spherical surface by polyhedra. Considering isotropy, each triangle of the regular icosahedron is divided into 4'' (n
=0.1,2. ,,,).

2) For all feature vectors obtained from the input data, determine which region on the spherical division of 1) the end point belongs to. Further, 1) and the operation of the feature vector matching unit 20 are performed in advance for the standard data as well.

3) Let Ai be the number of end points of the feature vector obtained from the input data for each region on the spherical division (where '-1+91.+m, m is the number of divisions), and let B be the standard data. Now Σ A, -B.

Rotate the input data to minimize.

4) Correspond between the data on both sides starting from the closest one.

Through the above procedure, input data can be rotated according to the required accuracy by appropriately selecting n, and correspondence can be established between vectors representing feature amounts.

〔Effect of the invention〕

According to the present invention, since the vector representing the feature quantity of a three-dimensional object is treated as a three-dimensional vector quantity, stable recognition can be expected even when the input data contains errors due to noise in the measurement system, etc. . Furthermore, since the calculated correlation coefficient is a quantity that is not affected by the rotation of the object, it can be recognized regardless of the orientation of the object.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is a diagram illustrating a method of expressing feature quantities of an object, FIG. 3 is a diagram explaining processing in the feature vector matching section, and FIG. A diagram explaining the method of dividing a spherical surface into icosahedrons and then dividing them into 2 and 4″ pieces, Figure 5 is a diagram explaining the conventional method of handling three-dimensional data, and Figure 6 is a diagram explaining the 5-pherical mode. 10: Feature extraction unit 20; Feature vector association unit 30: Standard pattern storage unit 40: Correlation coefficient determination unit 50: Pattern determination unit. Utilizes correlation.

Claims (3)

[Claims] (1) An input part that measures the position information from a three-dimensional object and accumulates the three-dimensional coordinates of the point elements of the object for each X, Y, and Z coordinate, and a A feature extraction unit that calculates direction data represented by a direction, a standard pattern storage unit that stores pre-calculated feature quantities of standard patterns, and a correspondence between feature quantities between input patterns and standard patterns. The present invention is equipped with a feature vector matching section that determines the correspondence, and a correlation coefficient determining section that calculates a judgment measure using spherical correlation based on the correspondence, and uses the judgment measure to determine the type of input data. A three-dimensional shape recognition device with special features. (2) In order to determine the correspondence between point data distributed on two spherical surfaces corresponding to two sets of direction data, the feature vector matching section divides the spherical surface isotropically, and divides the spherical surface isotropically. The number of points existing in each small part is counted for the above two sets of direction data, and based on the distribution, the correspondence between the points is determined so that the sum of the distances between each point is the smallest. The three-dimensional shape identification device according to claim (1), characterized in that: (3) The correlation coefficient determining unit is configured to perform a spherical correlation between the direction data obtained from the input pattern and the standard pattern based on the correspondence between points obtained by the feature vector matching unit. The three-dimensional shape identification device according to claim 1, characterized in that a correlation coefficient is determined.