JPH1055447A - Object recognizing method and device using the same - Google Patents

Abstract

PROBLEM TO BE SOLVED: To simply and precisely recognize an object. SOLUTION: PC(personal computer) 4 inputs the video of the object from a camera. An extraction part 20 extracts the outline of the photographed object. A calculation part 2 calculates the curvature of the extracted outline. An identification part 24 identifies the object based on the feature of the calculated curvature. The identification part 24 refers to a storage part 26 holding the feature of the curvature on the outline of various objects and classifies the objects.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] The present invention relates to a method for recognizing an object and an apparatus using the method.

[0002]

2. Description of the Related Art Pattern recognition (pattern recognit)
ion) is a technique for determining the meaning of data to be processed.
The data takes various forms, such as one-dimensional audio signals, two-dimensional image data, and three-dimensional three-dimensional data. In such a situation, image pattern recognition substitutes a computer for recognition of the outside world by a human visual system, and active research is being conducted.

[0003] Image pattern recognition has been put to practical use in various industrial settings. As a typical method, a statistical pattern recognition method (statistical pattern recognit
ion) has been used. This approach uses deterministic methods to assign unknown patterns to one or more of a number of classes. For this purpose, first, an N-dimensional vector called a “feature vector” is extracted from the input data, and a matching between the N-dimensional vector and a feature vector of a standard pattern which is a “model” prepared in advance is calculated. As a preprocessing for matching, pattern normalization (pattern
n normalization), that is, adjustment of image enlargement, reduction, translation, rotation, image density, and the like.

[0004]

The problem with the statistical pattern recognition method is that the accuracy of pattern recognition depends on the quality of normalization. However, there are generally many items to be considered in normalization, and it is difficult to make optimal adjustments. For example, when identifying an object such as a part flowing on a production line of a factory, the position where the part is placed will not be constant, and it is necessary to appropriately translate or rotate the image showing the part. However, since the type of component is unknown, it is inherently impossible to determine the optimal amount of movement and rotation. It is hard to imagine a situation in which the part being identified is similar to part A if it is rotated 30 degrees, but more similar to part B if it is rotated 90 degrees. Further, even if parts having the same shape are painted in different colors, the difficulty of normalization increases. Even if the colors of the components are the same, normalization may be required in consideration of the lighting conditions and the amount of light entering from the windows. Additional normalization is required if the type of component is added.

[0005] In recent years, in particular, there has been a demand for flexible manufacturing in factories, and various parts (and parts that vary depending on time zones) are arranged in one line.
You can see the flow Under these circumstances, accurate and real-time identification of objects such as parts is indispensable for improving production efficiency.

[0006]

SUMMARY OF THE INVENTION The present invention has been made in view of the above circumstances, and an object of the present invention is to provide an object recognition technique capable of eliminating normalization or reducing the effort as much as possible. . Another object is to provide a technique that can easily recognize an object correctly with a relatively small calculation load. Still another object is to provide a technique capable of recognizing an object with a relatively inexpensive configuration.

[0007] The present invention utilizes the following rules of thumb. That is, a person can tell the object from the silhouette of the object. This is because the features of the shape of the object appear on the outline. Moreover, a major factor that characterizes the contour is the degree of curvature, that is, curvature. Since object recognition is modeled on the function of the human visual system, a method by which a person grasps the characteristics of an object is considered to be effective even on a computer.

Therefore, in the object recognition method of the present invention, a step of extracting an outline of an object, a step of calculating a curvature of the extracted outline, and an object are identified based on characteristics of the calculated curvature. Steps. In determining the protection range, “curvature” should be interpreted widely, and should include all indications of the degree of curvature of the outline, such as the radius of curvature.

According to the present invention, since only the contour lines need to be extracted, normalization is not required or the number of steps is extremely small. The accuracy of the normalization is not likely to be a problem. For this reason, stable object recognition is realized. Since it is only necessary to pay attention to the outline, a new object can be dealt with relatively easily. Since the outline of the object is closed, the curvature can always be expressed with an arbitrary point on the outline as a start point. This is to return to the original point after making one round, and the present invention which focuses on the curvature is advantageous in that it does not depend on the coordinate system. Since we are only concerned with contours,
The computational load is relatively light.

In one aspect of the present invention, an object is identified based on the number of characteristic points of curvature. The number of feature points is naturally irrelevant to the coordinate system, and normalization such as enlargement, reduction, parallel movement, and rotation of the image is unnecessary. The same applies to the case where an object is identified based on the arrangement of characteristic points of curvature.

Another aspect of the present invention includes a step of extracting an outline of an object, a step of calculating a curvature of the extracted outline, and a step of storing the calculated characteristic of the curvature in association with the object. Including. By storing the characteristics of the curvature for various objects, the database construction of the object based on the curvature is realized.

Still another preferred embodiment according to the present invention relates to an object recognizing device. The apparatus includes an input unit for inputting an image of an object, an extraction unit for extracting a contour of the input object, a curvature calculation unit for calculating a curvature of the extracted contour, and a feature of the calculated curvature. And an identification unit for identifying an object based on the identification information. When the input unit is constituted by a camera, one camera may be used. The input unit may be an OCR or a drawing tool. In either case, the cost of the device is relatively small.

[0013]

DESCRIPTION OF THE PREFERRED EMBODIMENTS When an object moves on a screen in a computer moving image, a person can recognize the object regardless of the posture or the moving direction of the object. When a human performs such recognition, a feature of the shape of the object independent of the posture and direction of the object is grasped as an “invariant”. As this invariant,
There are critical points, that is, vertices (peaks) of convex portions and valley bottom points (pits) of concave portions. These points are obtained from a curvature graph of the contour of the object.

FIG. 1 is a configuration diagram of an object recognition device according to an embodiment. As shown in FIG. 1, the apparatus includes a camera 2 for photographing an object, and a PC (personal computer) 4 for capturing an image of the camera 2 and recognizing the object. This apparatus identifies a part 8 flowing on a conveyor 6 in a factory, and a camera 2 is mounted so as to face the conveyor 6. When the part 8 is identified by the PC 4, the PC 4 gives an instruction 12 required for manufacturing to any manufacturing machine 10.

FIG. 2 is a diagram showing the internal structure of the PC 4, and FIG.
Is a flowchart showing a processing procedure according to the first embodiment. As shown in FIG. 2, the PC 4 includes an extracting unit 20 for extracting the contour of the photographed object, a calculating unit 22 for calculating the curvature of the extracted contour, and identifying the object based on the characteristic of the calculated curvature. The identification unit 24 is implemented as a software module. The identification unit 24 refers to a storage unit 26 that stores the features of the curvatures on the contours of various objects.

The extracting section 20 is constituted by a known differential filter such as a Robert filter, a Laplacian filter, a Laplacian-Gaussian filter, and extracts the contour of the object (see FIG. 3). S1).

The calculation unit 22 calculates the curvature at each point on the contour based on the position of the pixel on the contour (S in FIG. 3).
2). The curvature is the reciprocal of the radius of the circle that is most tangent to the contour at that point, but since the contour is now represented as a set of discrete pixels, a plurality of calculation methods can be considered for the curvature.

1. Use of Discrete Pixel Information FIG. 4 is a diagram showing a method of directly obtaining a curvature from discrete pixels. The curvature κ of a certain pixel P0 is determined by L1: an angle θ formed by a straight line determined by the pixel P0 and a pixel P1 n pixels before the pixel L2: a straight line determined by the pixel P0 and a pixel P2 n pixels ahead thereof.

In addition to the curvature, a distance s from the start point of the curvature calculation to each pixel is also defined for later graph display. First, s is set to 0 at the start point, a pixel whose distance from the start point is minimized is searched for, and the distance is set to s = 1. Next, a pixel having the minimum distance from the pixel of s = 1 is searched for, the distance is set to s = 2, and s is determined one after another. As a result, a graph of (s, κ) can be drawn. Of course, there are many other ways to determine the distance s.

2. Use of approximation curve A free curve such as a spline curve approximating discrete pixel positions is derived once. From an ontological point of view: Than 2. Is preferred. Actually, if you want to reduce the computational load: If you want to obtain more accurate results by adopting the method of 2. Should be adopted. The obtained approximate curve can be represented as a parameter by the variable t as follows.

[0021] {(f (t), g (t)) | 0 ≦ t ≦ T} f, g: C 2 class mapping that is defined in [0, T], (f (0),
g (0)) is the starting point. The curvature κ (t) at the point (f (t), g (t)) is obtained by Expression 1.

[0022]

(Equation 1) (Equation 1) Note that the length of the curve from (f (0), g (0)) to (f (t), g (t)) is

(Equation 2) (Equation 2) Also in this case, a graph of (s, κ) can be drawn.

FIG. 5 is a diagram showing a contour of a simple object X, and FIG. 6 is a graph plotting (s, κ) obtained by the calculation unit 22 with respect to the contour. In FIG. 5, point A is a starting point, and points B, C,
The curvature is calculated along a path that follows D, E, F, and G and returns to point A. The horizontal axis in FIG. 6 is the distance s on the contour line, and the vertical axis is the curvature of each point. FIG. 6 also shows the position of the predetermined threshold value th.

The identification unit 24 identifies an object using the graph shown in FIG. 5 (S3 in FIG. 3). The identification unit 24
The storage unit 26 stores the characteristic of the curvature of the object X as a key.
The object whose curvature characteristic is close to the curvature characteristic of the object X is searched from a plurality of objects recorded in advance in the object X. In order to materialize the characteristic of curvature, the concept of a feature point is introduced. In the embodiment, taking into consideration the point where the curvature takes an extreme value, more specifically, the selection using the threshold value th, i) the positive maximum point where the curvature κ> th ii) the curvature κ <−th Two types of points are defined as characteristic points, namely, two types of negative minimum points, and iii) a point at which the curvature κ = 0. A code “α” is assigned to a point (peak) satisfying the condition i), a code “β” is assigned to a point (pit) satisfying the condition ii), and a code “γ” is assigned to a point satisfying the condition iii). Therefore, in the case of FIG. 6, points B, C and G are “α”, point E is “β”, and points D and F are “γ”.
And the outline is represented as an array of feature points, ααγβγα. Once you have prepared so far, you can search by specifying the following key.

(Example 1) Using the total number of α and β as a key In the case of FIG. 6, it is “4”. This is because it is considered that α and β more clearly represent the shape of the object than γ.

(Example 2) Using the total number of α, β, and γ as a key In the case of FIG. 6, it is “6”. Although γ exists between α and β, it is unknown whether γ exists between α and α or between β and β. Therefore, the key is used including the number of γ.

(Example 3) Using the Numbers of α and β as Keys In the case of FIG. 6, “3” and “1” constitute a two-variable key. Also in this case, a key of three variables may be created in consideration of γ.

(Example 4) Using a sequence of α, β, and γ as a key In this case, a sequence obtained by cyclic permutation of each other, for example, α
γβγαα and γβγααα must be regarded as identical to the original sequence. This is because the position of the starting point has just changed. Also in this case, γ may be omitted.

(Example 5) The positions of α, β, and γ are added to the position information and used as a key. For example, distances on the circumference of six feature points are s1 to s6, respectively.
Then, these six numbers are also added to the key. In that case, the length of one circumference of the contour line may be normalized to “1” in advance. It should be noted that this normalization is very easy and there is no adjustment difficulty that has been a problem with the prior art.
Of course, in this case as well, γ may be omitted.

(Example 6) Using the magnitude of curvature at a feature point as a key If, for example, the curvatures of six feature points are respectively κ1 to κ6, these six numbers are used as keys instead of α to γ. .
Also in this case, the length of one circumference of the contour line may be normalized. Also in this case, the feature point corresponding to γ may be omitted.

The above is an example of the object identification method. The identification unit 24 searches the storage unit 26 for an object having characteristics similar to the captured object, and classifies or identifies the object. As a result, a necessary instruction 12 is issued to a necessary manufacturing apparatus. Here, six examples have been given, but of course, a plurality of these may be used in combination, or another characteristic point may be employed for the curvature. For example, when a section where the curvature is substantially zero exceeds a predetermined length, both end points of the section can be set as feature points. In that case, a straight line portion of the contour of the object can be extracted. Further, the threshold value th is not essential.

FIG. 5 shows a simple object. However, if the object is complicated to some extent, a considerable number of feature points can be obtained, which is convenient for retrieval. Especially when there are many feature points, a candidate that completely matches the key may not be found. In that case, the closest candidate may be searched using existing various matching methods.

In the case of FIG. 5, the contour can be represented by a single curve, but the contour can be represented by several smooth curves (C ²
For general expressed as a set of grade), firstly decompose the entire outline into a set of such C ² grade curve. And for each curve, determine the starting point and starting direction,
Calculate s (t) and κ (t). As a result, a set of {(graph of (s, κ), start point, moving direction from start point)} is obtained. If the direction is determined in advance for the entire contour, this set is clearly invariant,
Can be used to identify objects.

The above is the object recognition method and apparatus according to the embodiment. For this embodiment, for example, the following modification technology is also conceivable.

(1) Here, the identification of parts flowing on the factory line was considered. However, the present invention is generally applicable to automation of production. For example, if the animator inputs the outline of an object to a PC using a drawing tool, a past object close to the object can be called from the database based on the outline. In addition, a designer who is involved in the shape of an object in some way can save labor and share the design by taking the same method.

(2) The calculator 22 calculates the curvature graph (s,
κ) was calculated. Conversely, when this graph is given, the original contour can be restored by the following calculation.

[0037]

(Equation 3) These equations are further integrated to obtain a (x, y) two-dimensional curve as follows. x0 and y0 are integration constants.

[0038]

(Equation 4) Therefore, the graph of (s, κ) can be considered as encoded data of the contour of the object. If this graph is quantized and stored according to an appropriate standard, image data can be compressed while maintaining the feature of the shape of the object.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of an object recognition device according to an embodiment.

FIG. 2 is an internal configuration diagram of a PC 4;

FIG. 3 is a flowchart showing a processing procedure by a PC 4.

FIG. 4 is a diagram illustrating a method of directly obtaining a curvature from discrete pixels.

FIG. 5 is a diagram showing an outline of a simple object X;

6 is a diagram in which (s, κ) obtained by a calculation unit is plotted with respect to the outline of the object in FIG. 5;

[Explanation of symbols]

2 cameras, 4 PCs, 6 conveyors, 8 parts, 10
Manufacturing machine, 12 instructions, 20 extraction unit, 22 calculation unit, 24 identification unit, 26 storage unit.

Claims (14)

[Claims] 1. A method comprising: extracting a contour of an object; calculating a curvature of the extracted contour; and identifying an object based on the characteristic of the calculated curvature. Object recognition method. 2. The object recognition method according to claim 1, wherein a plurality of objects recorded in advance are searched for those whose curvature characteristics are close to the calculated curvature characteristics. 3. The method according to claim 2, wherein the object is identified based on a characteristic point of curvature. 4. The method according to claim 3, wherein the feature points include extreme points of curvature. 5. The method according to claim 4, wherein the feature points further include a point at which the curvature becomes zero. 6. The object recognition method according to claim 4, wherein only an extreme point whose absolute value exceeds a predetermined threshold value is determined as a feature point. 7. The object recognition method according to claim 3, wherein the object is identified based on the number of feature points. 8. The method according to claim 3, wherein an object is identified based on an array of feature points. 9. The object recognition method according to claim 8, wherein all the arrays obtained by the cyclic permutation of the elements of the array are regarded as the same array. 10. The object recognition method according to claim 3, wherein an object is identified based on a position of a feature point on a contour line. 11. The object recognition method according to claim 3, wherein the object is identified based on the magnitude of the curvature at the characteristic point. 12. A method comprising: extracting a contour of an object; calculating a curvature of the extracted contour; and storing a characteristic of the calculated curvature in association with the object. Object recognition method. 13. An input unit for inputting an image of an object, an extraction unit for extracting an outline of the input object, a curvature calculation unit for calculating a curvature of the extracted outline, and a feature of the calculated curvature. An object recognizing device, comprising: an identification unit that identifies an object based on the object. 14. The apparatus according to claim 13, further comprising: a storage unit that retains a curvature characteristic of a plurality of objects, wherein the identification unit is one of a plurality of objects recorded in advance. An object recognition device that searches for a feature whose curvature is close to the calculated curvature feature.