JP3180985B2 - Object shape description device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an object shape description device for describing the shape of an object.

[0002]

2. Description of the Related Art When an object having various shapes is described by a specific basic model, the object is divided into a plurality of parts so that the object can be described by the model. As a division method, there is a method in which a maximum basic model inscribed in an object is obtained, and a maximum basic model inscribed is sequentially obtained for a portion that cannot be expressed by the first model. As a result, the object is divided into basic models of various sizes. In addition, a criterion for convexity of the shape using the geometric feature is provided, and the object is subjected to convex division according to the criterion,
There is also a method of dividing a component into parts by approximating a basic model to each convex portion. However, these methods have a drawback that an object is divided more than necessary.

[0003] Therefore, as a method of integrating these plurality of divided parts into an appropriate part configuration, an evaluation function is set in advance, and if adjacent parts are integrated, if the evaluation function value becomes better than before integration, A method of repeatedly calculating an operation of integrating and otherwise separating the data is used. In this method, each component is represented by a description having a binary tree structure, and the above-described iterative calculation is a method in which a final divided result is obtained when a predetermined convergence condition is satisfied.

[0004]

The above-described conventional object shape description apparatus has the following drawbacks. (1) It is necessary to set different evaluation functions and convergence conditions depending on the object. (2) In the case where components are integrated using a binary tree structure, if any one portion is incorrectly integrated, all subsequent integrations will be incorrectly (or not appropriate). (3) Further, the use of geometric features is limited to the case where dense data is obtained as shape data.

An object of the present invention is to provide an object shape description device which can use the same evaluation function for different objects and can be easily divided into parts which can be expressed by functions regardless of the shape of the object. Is to provide.

[0006]

The first object shape description device of the present invention, in order to solve the problems] includes a data input unit for inputting shape data, input
The input shape data is divided into n (≧ 2) data sets.
Describe the part division part to be divided and the input data by a function
From the shape description part and the individual data set
k (<n) are selected, the data sets are integrated, and the integration is performed.
Data set and the remaining data set
On the other hand, a function approximation is executed in the shape
Obtain the similarity error, and the approximate error of the approximated function model
And the number of data sets used for the approximation, and
Evaluation with the number of free parameters of the function model as an argument
The value of the value function is calculated, and the data
And a control unit for performing a process for obtaining the cost.

A second object shape description device of the present invention, the shape
A data input unit for inputting data, and input shape data
Part division unit that divides data into n (≧ 2) data sets
And a function model with a variable number of parameters
Shape description part approximated by
Select k (<n) items from the set and merge the data sets
And changing the shape data and the number of parameters.
The error with the described function model and the number of parameters
To determine the value of the evaluation function
And a control unit for obtaining a function model.

[0008]

According to the first object shape description apparatus of the present invention, as a combination of a large number of divided parts, starting from the integration of two different parts, the integration of n parts (this makes the entire data set into one part) ) Are obtained, and function approximation is performed independently for each component of each combination. In each combination, the error of data having a shape that cannot be approximated by the function model has a large error. Therefore, a limited plurality of candidates can be selected from the evaluation function term having the approximation error of each component as an argument. Further, there is another term of the number of data sets and the number of parameters. By finding the combination that minimizes this, even if the approximation error is the same, the number of parts is smaller and the function model is simpler. Model can be selected as the final division result.

According to a second object shape description device of the present invention,
Unlike the apparatus described above , the object is not divided into parts. Instead, the shape data of the object is described by changing the number of parameters using a function model in which the number of parameters is variable. The step of obtaining the value of the evaluation function and the step of obtaining a function model that minimizes the value of the evaluation function are the first steps.
Method is the same. However, since the number of components is always 1, the number of components (the number of all data sets) is not included in the argument of the evaluation function.

[0010]

Next, embodiments of the present invention will be described with reference to the drawings.

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

This embodiment comprises a data input unit 1 for inputting shape data, a component dividing unit 2 for dividing the input shape data into components, and a microprocessor. It comprises a control unit 3 for performing data processing to be described later according to a program, and a shape description unit 4 for describing the shape of an object. The control unit 3 includes a component integration unit 5 for integrating a plurality of data sets into one data set, a function approximation unit 6 for performing a function approximation to the data set, and a component integration result evaluation for comparing a division result based on evaluation criteria. Means 7 are provided.

FIG. 2 is a flowchart showing the processing of each section of the present embodiment.

First, shape data is input in the data input section 1 (step 10). Next, the shape data is divided into components of a super quadratic function in the component dividing unit 2, and a function inscribed in the input data is obtained by a weighted minimum approximation method (step 11). Next, data existing outside the function is labeled (step 12), and the input shape data is divided into two or more data sets (step 13). The operation of obtaining the inscribed function for the divided data set is similarly repeated (step 14). Steps 10 to 14 described above are the invention described in Japanese Patent Application No. 4-7392. Next, a data set is extracted from all the data sets, and a new data set is generated (step 15). Approximate the newly generated data set by the function least squares method (step 1)
6). The value of the evaluation function F is obtained for all data sets (step 17). A data set having the minimum value among all the values of the evaluation function F is obtained (step 18).

Here, the super quadratic function is defined by the following equation (1).

[0016]

(Equation 1) ε is called a shape parameter, and is a parameter that changes the degree of angularity of the function shape. a ₁ and a ₂ are called scale parameters, and can change the size of the function in the x and y axis directions. That is, by changing the shape parameter and the scale parameter, various shapes (rectangle, ellipse, rhombus, etc.) can be expressed by one function. Generally, a super-quadratic function refers to a three-dimensional function. Here, since the target data is two-dimensional contour data, it is used as a two-dimensional function.

FIG. 3 is a diagram showing an example of dividing input shape data according to the present embodiment, and FIG. 4 is a diagram showing an example of component integration.

First, shape data (FIG. 3A) is input by the data input unit 1 (step 10). The input shape data is divided by the component dividing unit 2 into shapes that can be expressed by a super quadratic function. First, an initial function (having the shape of a unit circle) is placed at the center of gravity G of the data, and the inscribed function shape p0 (FIG. 3) is obtained by the weighted least squares method.
(B)) is obtained (step 11). Here, the weighted least squares method is a least squares method in which data inside a function is heavy and each data is weighted such that the weight decreases as the distance from the function surface increases.

As an initial value, the shape of a unit circle is placed inside the object and the function parameters are changed to obtain a function shape inscribed in the object. Since the hyperquadric function is considered to be a function in which the exponent part is replaced with a real number in the definition expression of the expression representing the ellipse, each data is substituted for S (x, y). Data is outside the function, S = 1
If so, it can be determined that the data is within the function, and if S <1, the data is within the function. The data inside the function means data existing inside the ellipse (sphere). And
Sum of squared errors

[0020]

(Equation 2) And the value of the weight ω is changed depending on whether each data is located outside or inside the function.

Next, by labeling data existing outside the function p0 (step 12), the input shape data is converted into five data sets d0, d1, d2, d
3 and d4 (step 13). Where d0
Is data existing on the function p0 (FIG. 3)
(C)). By repeating the operation of similarly obtaining an inscribed function for each of the divided data (step 14),
The input shape data includes five function models p0, p1,
It will be described by p2, p3, p4 (FIG. 3 (d)).

Next, the processing in the control section 4 is performed. First, all data sets d0, d1, d2, d3, d4
, K (2 ≦ k ≦ 5) data sets are extracted, and a new data set is created (step 15). For example, all data sets d0, d1, d2, d3, d4
Is regarded as one part P00 (that is, the data sets d0, d1, d2, d3, and d4 are the data of FIG. 3A itself), the new data set d00 that configures the part P00 is Each data set d0, d
1, d2, d3, d4 sum d0 + d1 + d2 + d3 +
d4 (step 15). This data set d00
The least squares method is applied to to approximate a super quadratic function (step 16). As a result, a part shape P00 as shown in FIG. 4A is obtained. The residual (square error) at this time is r00. When the data sets d0 and d1 are regarded as one component, the new data set d01 is a data set based on the sum of the data sets d0 and d1.
In this case, four data sets d01, d2, d3, and d4 are obtained (step 15). Each data set d01,
As a result of approximating the functions to d2, d3, and d4, as shown in FIG.
1, P2, P3, and P4 (step 16). In this case, each part P01, P2, P3,
The residuals at P4 are r01, r2, r3, r4, respectively.
And When the data sets d0, d1, and d2 are one data set, the data set is d02 (=
d0 + d1 + d2), d3, and d4 are obtained (step 15), and the function approximation result at this time is as shown in P02 of FIG. 4C (step 16). In this case, the residuals of the components P02, P3, and P4 are r02, r3, and r4, respectively. As described above, component descriptions in all combinations are obtained for the five data sets d0, d1, d2, d3, and d4.

When the residuals are obtained for all the combinations, the value of the evaluation function F is obtained for each combination (step 17). Here, the sum of the log likelihood of the residual for evaluating the goodness of approximation and the product of the number of parts and the number of free parameters of the function model is defined as the value of the evaluation function F. That is,

[0024]

F = (sum of log likelihood of residual) + (number of parts)
× (the number of free parameters of the function model) is evaluated.

In the case of this embodiment, the number of free parameters of the hyperquadric function is three for the function itself (the above-mentioned ε, a ₁ , a ₂ ) and the number of parameters for transforming the data set into the function coordinate system. There are three parameters and one distribution parameter for the data set to have data with a specific variance, for a total of seven or more. Here, the parameters for transforming the data set into the function coordinate system include two parameters for the parallel movement in the x and y axis directions and the rotation of the axis of the function in order to move the function in parallel to the center of gravity of the data. Means a total of three rotation parameters for approximating the data shape well.

[0026]

F (Example 1; FIG. 4 (a)) = log (r00) + 1 × 7 F (Example 2; FIG. 4 (b)) = (log (r01) + log (r2) + log (r3) + log (R4)) + 4 × 7 F (Example 3; FIG. 4 (c)) = (log (r02) + log (r3) + log (r4)) + 3 × 7 F (Example k) = Σ (log (ri ) + K × 7 is calculated (step 17).

An evaluation function F having the minimum value (in this case, k) is searched for from all the values of the evaluation function F, and the combination of the data sets at this time is considered as a data set of a part to be divided (step 18). Step 15 above
To 18 are repeated until the total number of components no longer changes (step 19).

Finally, two combinations as shown in FIGS. 4D and 4E are obtained. That is, the component P04 is obtained by integrating the data sets d0, d1, and d3, the component P05 is obtained by integrating the data sets d2, d4, or the component P07 is obtained by integrating the data sets d0, d2, and d4. A part P06 is obtained by integrating d1 and d3 and described (FIG. 4D,
(E), step 20). In this case, each is a shape description composed of two parts.

Next, as a second embodiment of the present invention, an example of application to an object having only sparse data is described. A super quadratic function is used as a function model. A case where the shape data of the target as shown in FIG. 5B is obtained by corner edge detection as the shape data of the target as shown in FIG. 5A will be described.

First, by labeling the target data, the shape data is converted into a plurality of data sets d0, d1, d.
2, d3, d4, d5, d6, and d7 (step 12) (FIG. 5C). As a result of approximating the function to each classified data set, descriptions p0, p1, p2, p
3, p4, p5, p6, and p7 are obtained (step 1).
3) (FIG. 5 (d)). Considering these as eight parts,
As described in the embodiment, the data sets are integrated. FIG. 5E shows an example in which the data sets d0 and d1 are integrated as one component. First, data sets are integrated (step 15), and a function is approximated to a new data set to obtain a part p01 (step 1).
6). Similarly, FIG. 5F shows data sets d3 and d3.
The result of integrating d4, d5, and d6 is shown, and the parts p02,
p0, p1, p2, p7 are obtained. FIG. 5 (g) shows the result of integrating the data sets d1 and d7, and the parts p03,
p0, p2, p3, p4, p5, p6 are obtained. FIG.
In (n), as a result of integrating the data sets d0, d1, d2, and d7, a part p04 is obtained, and the data set d3 is obtained.
As a result of integrating d4, d5, and d6, the result of obtaining the part p05 is shown.

By calculating the evaluation function F shown in the first embodiment for each case, the description of FIG. 5 (n) in which the evaluation function F has the minimum value is selected, and shown in FIG. 5 (n). Obtain the division result.

Next, as a third embodiment of the present invention, an example will be described in which the obtained shape data is described by a function model in which the number of parameters is variable. The Fourier descriptor is used as a function model.

In this case, the means for determining the order of the function model corresponds to the means for determining the data set combination to be integrated in the embodiment shown in FIGS. First, FIG.
When the order of the Fourier descriptor is changed for the shape data of the object as shown in FIG.
The description shown by the solid lines of (d) and (e) is obtained.

For example, when the degree of the Fourier descriptor is first, a description as shown in FIG. 6B is obtained. At this time, the error between the descriptor and the shape data is r01. When the degree of the Fourier descriptor is third, a description as shown in FIG. 6C is obtained. At this time, the error between the descriptor and the shape data is r02. When the order of the Fourier descriptor is 10, the description as shown in FIG. 6D is obtained. At this time, an error between the descriptor and the shape data is r10. For example, when the order of the Fourier descriptor is 30, the description as shown in FIG. 6E is obtained. At this time, the error between the descriptor and the shape data is r30.
As described above, as the order of the Fourier descriptor increases, the original data shape can be described in detail (that is, the error can be reduced). I can hardly see it.

Next, the value of the evaluation function F in each description of FIGS. 6B to 6E is obtained. In this case, the number of parts is always one.

[0036]

F (FIG. 6 (b)) = log (r01) + 1 × 1 F (FIG. 6 (c)) = log (r03) + 3 × 1 F (FIG. 6 (d)) = log (r10) +10 × 1 F (FIG. 6E) = log (r30) + 30 × 1 In FIGS. 6B and 6C, the error is large, and as the order of the Fourier descriptor increases, the value of the evaluation function F increases. Gradually decreases. However, as the order of the Fourier descriptor increases, the number of parameters of the second term of the evaluation function gradually becomes larger than the error. From FIG. 6D, the value of the second term becomes larger than the first term (sum of distances).
Becomes larger. As a result, the value of the evaluation function F is minimized in the description of FIG. 6D, and the description of FIG. 6D is obtained as a final result. As described above, the superquadratic function is used as the function in the first and second embodiments, and the Fourier descriptor is used in the third embodiment. If there is, it is not specified. As a part dividing method, here, the part division is performed by obtaining an inscribed function. However, even if a convex part is divided by function contraction from the outside, or even if another part dividing method is used, the part is divided. There is no change in the integration means, and the method of dividing the components into the basic models is not specified. In the first, second, and third embodiments, two-dimensional contour data is used as shape data. However, there is no problem even if a binarized region is used as object shape data. Furthermore, by using a three-dimensional model as a basic model, it can be easily applied to the description of a three-dimensional object (such as a description of an object from a distance image), and the data dimension is not specified. In the third embodiment, an example of a method of determining a function model for one data has been described. However, the method of the third embodiment is applied to each data set used in the first and second embodiments. It is also possible to apply.

[0037]

As described above, the present invention has the following effects. 1. According to the first aspect of the present invention, as a combination of a large number of divided parts, all combinations from the integration of two different parts to the integration of n parts are obtained, and a function is independently performed for each part of each combination. An approximation is performed, and in each combination, a limited plurality of candidates are selected from terms of the evaluation function with the approximation error of each part as an argument, and further, a combination that minimizes the terms of the number of data sets and the number of parameters is obtained. This has the following effects. (1) Despite the shape of the object, it is possible to easily divide the parts into parts that can be represented by a basic model (function), and to obtain a description with the minimum number of parts. At this time, even when a plurality of division methods are conceivable, division by the unified standard is possible by using the evaluation function. (2) Even if the initially input data is sparse, the data is initially divided into a large number of parts, but ultimately a plurality of data sets that originally constituted one part are 1
Even if the approximation errors are the same and the approximation error is the same, a candidate with a smaller number of components can be obtained as the final division result. (3) When the basic model has a model in which the number of free parameters changes, such as a spherical harmonic function, the higher the degree of the function, the more accurate the description. However, the shape description should be as simple as possible. Is desirable. In the present invention, a model having a smaller number of free parameters is finally selected by the second term of the evaluation function. Therefore, when the approximation errors are the same in a plurality of models, a model with a simpler description can be selected. . 2. According to the second aspect of the invention, the obtained shape data is
2. The method according to claim 1, wherein the number of parameters is described by a variable function model.
In the same manner as in the invention described above, by obtaining a function model that minimizes the value of the evaluation function, the object shape can be accurately described with a simpler function model.

[Brief description of the drawings]

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

FIG. 2 is a flowchart showing a process of the embodiment of FIG. 1;

FIG. 3 is a diagram showing an example of division of input shape data.

FIG. 4 is a diagram illustrating an example of component integration.

FIG. 5 is a diagram illustrating an example of component integration in a case where obtained shape data is only an object having sparse data in the second embodiment of the present invention.

FIG. 6 is a diagram showing a result of describing obtained shape data with a function model having a variable number of parameters in a third embodiment of the present invention.

[Explanation of symbols]

Reference Signs List 1 data input unit 2 parts division unit 3 control unit 4 shape description unit 5 parts integration means 6 function approximation means 7 parts integration result evaluation means 10 to 20 steps G center of gravity p0 to p4 function model P00, P01, P02, P03, P04, P05, P
06, P07 Function approximation result (parts) d0 to d7 Data set d01 to d07 Integrated data set p0 to p7, p01, p02, p03, p04, p05
Function approximation results for parts

────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-5-73673 (JP, A) JP-A-5-197785 (JP, A) JP-A-4-246786 (JP, A) 96871 (JP, A) JP-A-5-94524 (JP, A) Horikoshi et al., "Basic element division of three-dimensional object by hyperquadric", IEICE Transactions D- ▲ II ▼, Vol. J76-D-IIII, No. 1, pp. 30-39, January 1993 (58) Fields investigated (Int. Cl. ⁷ , DB name) G06F 17/50 G06T 1/00 JICST file (JOIS)

Claims (2)

(57) [Claims] 1. A data input unit for inputting shape data, and converting the input shape data into n (≧ 2) data sets
A part dividing unit to be divided, a shape description unit that describes the input data as a function, and k (<n) pieces are selected from the divided individual data sets.
Select the data set, merge the data set,
And the shape for the remaining dataset
Execute function approximation in the description part and obtain approximation error
The approximation error of the approximated function model and the approximation error
The number of data sets used for
The value of the evaluation function with the number of free parameters
Find the data set that minimizes the value of the evaluation function
An object shape description device having a control unit for performing processing. 2. A data input unit for inputting shape data, and converting the input shape data into n (≧ 2) data sets.
The part division part to be divided and the data set are close by a function model with a variable number of parameters.
A similar shape description section and k (<n) pieces are selected from the divided individual data sets.
Select the data set, integrate the data set,
Error from the function model described by changing the number of meters
And the value of the evaluation function with the number of parameters as arguments
Control unit that finds a function model that minimizes the value of the evaluation function
An object shape description device having: