JPH09297857A - Picture recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To make model generation work to be efficient by permitting a worker to add an appropriate feature at the time of compiling a model without executing a feature extraction processing again even if the feature which is to be registered to a model cannot be obtained. SOLUTION: Feature data approximated to a straight line and an arc are generated on the outline, projecting recessing parts and the pattern of an object from an input picture and the worker selectively registers feature data as the model features. Furthermore, the arbitrary straight line is added to the model features. When the feature is extracted from the input picture, for example, it is often a case when the feature of the object, which is obtained as the 'straight line' with high probability, can be obtained as the 'arc' in feature extraction executed at the time of model generation work by accident. In this case, the 'arc' is not registered at the time of registering the model feature but the worker adds the appropriate 'straight line' at the time of compiling the model. Thus, the dispersion of feature data extracted from the input picture is absorbed, model manufacture work is made efficient and recognition performance can be improved.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an object recognition apparatus, and is particularly useful when it is used to create a model of an object used for feature matching in the model-based matching method.

[0002]

2. Description of the Related Art When recognizing a three-dimensional object from a two-dimensional plane image, that is, when recognizing a three-dimensional object in a so-called visual recognition device, a two-dimensional projection image is displayed on a monitor to visually confirm the certainty of the recognition. There is a way to do it. In other words, by displaying the two-dimensional image of the target object and the three-dimensional object model image of the target object on the monitor in a superimposed manner, it is possible to easily visually confirm the deviation between the target object and the three-dimensional object model. is there.

In this case, as a monitor of the three-dimensional object model, a three-dimensional object model is used, and the three-dimensional object model is converted into a three-dimensional coordinate into a recognition posture, and further two-dimensionally projected onto camera coordinates to display the monitor. I am doing it.

The production of this three-dimensional object model is
Since the three-dimensional object model expresses the outer shape of the object, it is created from CAD data or created on a monitor using a graphic tool. For this reason, the three-dimensional object model is composed of geometric features such as straight lines, arcs, and ellipses.

On the other hand, so-called computer vision in which visual recognition is performed by a machine using a computer is being researched, and a model-based matching method is one of the methods for performing three-dimensional object recognition. This method stores a model of a three-dimensional object in the computer in advance, compares this model with the target object image, searches for a specific object in the image, and detects the three-dimensional position and orientation of the object. is there.

Even in such a model-based matching method, it is necessary to prepare a three-dimensional object model in advance before performing three-dimensional object recognition, and a model is prepared from CAD data or a third-order model is created by using a graphic tool. We are creating an original object model.

[0007]

As described above, in order to perform three-dimensional object recognition, it is necessary to create a three-dimensional object model. As described above, the model is created using CAD data or using a graphic tool. However, the following problems are occurring.

(1) When a model is created from three-dimensional CAD data, the model is characterized by a straight line, an arc, an ellipse, etc. Therefore, if the object has an arbitrary free-form contour or pattern, it is displayed on the monitor. When a target object and a three-dimensional object model image are superimposed and displayed, an error when approximated to a straight line or a circular arc becomes a display error, and a difference between the target object and the model occurs in detail.

Also, in model-based matching in computer vision, even in the case of feature matching in which an image from a camera is processed to extract the features of the object and the features are matched with the model, the features are the same as those of the CAD model. It does not match the feature image obtained by processing the two-dimensional projection image of the object, and this feature matching makes it difficult to obtain the same feature of the object as the CAD model, and the CAD model cannot be used as it is.

(2) When a three-dimensional object model is created using a graphic tool, depth information of the object is necessary. Therefore, the three-dimensional data of the object must be acquired by means other than the graphic means. I have to.

In order to solve the above-mentioned problems of the prior art, the present inventors need to prepare a three-dimensional object model by approximation of a straight line, an arc, and an ellipse by CAD, or to obtain three-dimensional data separately. A method for producing a three-dimensional object model without using the above has been previously proposed (Japanese Patent Application No. 7-23222).

This is because the operator basically selects one of the feature data that approximates a straight line, a circular arc, etc. by processing one actual image of the contour, unevenness, pattern, etc. of the object and that is necessary for the model data. According to this method, the process up to obtaining the feature data is the same as the recognition process, so there is an advantage that a model suitable for recognition can be created, but the change in the gray value of the input image from the camera and the target As shown in FIG. 13, there are variations in the output of the characteristic data due to changes in the position and orientation of the object, and the following problems may occur.

Since there are variations in the appearance of the feature data, even feature data that is obtained with an appropriate high probability as model data may not be obtained as feature data when selecting model data. It may make the manufacturing work difficult. Even if the feature data of the same part on the target object is feature data that is significantly different from the feature data registered in the model, in this case, the recognition performance deteriorates.

In view of the above problems, the present invention improves the efficiency of model making work by improving the efficiency of model making work and further improving the recognition performance, that is, absorbing the variation of the feature data extracted from the input image. An object of the present invention is to provide an object recognition device that improves performance.

[0015]

The structure of the present invention that achieves the above object has the following features.

1) Recognizing an object by matching the characteristic data obtained by approximating the features such as contours, irregularities, and patterns of the object obtained from the input image with straight lines and arcs, and the characteristic data of the model of the object. In the object recognition device for performing the above, the operator selects this feature data as a model feature based on the feature data obtained by approximating the features such as the contour, unevenness, and pattern of the object obtained from another input image to a straight line or an arc. A model creating means configured to have a function of registering and further allowing a worker to edit the characteristic data of the model such as a straight line and an arc. 2) In 1), the model creating means has a function of allowing an operator to select and register the feature data as a model feature and further add an arbitrary straight line to the model feature. 3) In 1), the model creating means has a function of allowing an operator to selectively register the above-mentioned feature data as a model feature and to add an arbitrary arc to the model feature. 4) In 1), the model creating means has a function of allowing the operator to select and register the above-mentioned feature data as a model feature, and further to move and adjust the position of the registered model feature according to the operator's judgment. . 5) In 1), the model creating means has a function that allows an operator to select and register the above-mentioned feature data as a model feature, and further rotate and adjust the posture of the registered model feature according to the operator's judgment. . 6) In 1), in the model creating means, the operator selectively registers the above-mentioned feature data as a model feature, and after adding an arbitrary feature to the model feature as needed, the feature unsuitable as the model feature is added to the operator. It must have a function that can be erased according to the above judgment.

[0017]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described in detail below with reference to the drawings.

The object recognizing apparatus according to the embodiment of the present invention has a model creating means for allowing a worker to edit the characteristic data of the model created by the conventional model creating function. .

Therefore, the basic processing procedure is, as shown in FIG. 1, after the image input processing of the object similar to the conventional _one (step S ₁ ), the contour, unevenness, pattern, etc. of the object are extracted from this input image. The characteristic data that is approximated to a straight line or an arc is extracted (step S ₂ ), and the operator selectively registers this characteristic data as a model characteristic (step S ₃ ).
After that, the operator appropriately edits the model feature data (straight line, arc) (step S ₄ ).

Various specific contents of such editing can be considered. A more specific embodiment of the present invention will be described focusing on the contents of each edit. Each embodiment is not limited to the two types of characteristic data of “straight line” and “arc” by the processing of the input image (of course, it is not limited to these “straight line” and “arc”, and there is no special limitation on the type of feature data. Not available) is obtained.

<First Embodiment (Measures by Adding a Straight Line)>
In this embodiment, feature data that approximates contours, irregularities, patterns, etc. of an object from an input image to straight lines or arcs is created, the operator selects and registers the feature data as model features, and further, an arbitrary straight line is used as a model feature. To be added to.

When feature extraction is performed from the input image,
The characteristics of the object that can be obtained as a “straight line” with a high probability happen to be “arc” in the feature extraction performed during model creation work.
May be obtained as. In this case, when registering the model feature, the "arc" is not registered, and the operator adds an appropriate "straight line" when the model is edited. This work is shown in FIG. 2 (thick lines in the figure are the edited features (the same applies hereinafter)).

When feature extraction is performed from the input image,
A feature of an object that is obtained as a "single straight line" with a high probability may happen to be obtained as a "set of a plurality of straight lines" in the feature extraction performed during the model creation work. In this case, when registering model features, those "straight lines" are not registered, but the operator adds appropriate "straight lines" when the model is edited. This work is shown in FIG.

When feature extraction is performed from the input image,
The feature of an object obtained as a “set of a plurality of straight lines” with a high probability may happen to be obtained as a “single straight line” in the feature extraction performed during the model making work. In this case, when registering the model feature, the "straight line" is not registered, and the operator adds an appropriate "straight line" when the model is edited. This work is shown in FIG.

When feature extraction is performed from the input image,
A part of the object may be obtained as a “straight line” or a “circular arc” with a similar probability, and may be obtained as a “circular arc” by the model making worker. In this case, when the model feature is registered, the “arc” is registered, and the worker adds an appropriate “straight line” to that part. This work is shown in FIG.

When feature extraction is performed from the input image,
A part of the object may be obtained as "one straight line" or "a set of a plurality of straight lines" with similar probability. In this situation, if obtained as a “plurality of straight line sets” during model creation work, register the “plurality of straight line sets” at the time of model feature registration, and in addition to that part To add. This work is shown in FIG.

On the other hand, when it is obtained as "one straight line" during the model making work, the "one straight line" is registered at the time of model feature registration, and the operator can add an appropriate "straight line" to the part. To add. This work is shown in FIG.

<Second Embodiment (Countermeasure by adding a circular arc)>
In this embodiment, feature data in which contours, irregularities, patterns, etc. of an object are approximated to straight lines or arcs from an input image is created, and the feature data is selected and registered by a worker as model features. To be added to.

When feature extraction is performed from the input image,
The characteristics of the object that are obtained as "arcs" with a high probability happen to be "straight lines" in the feature extraction performed during model creation work.
May be obtained as. In this case, when registering the model feature, the "straight line" is not registered, but the operator adds an appropriate "arc" when editing the model. This work is shown in FIG.

When feature extraction is performed from the input image,
A part of the object may be obtained as a “straight line” or a “circular arc” with a similar probability, and may be obtained as a “straight line” during the model making operation. In this case, when the model feature is registered, the "straight line" is registered, and the worker adds an appropriate "arc" to the part. This work is shown in FIG.

<Third Embodiment (Coping with Feature Movement)>
In this embodiment, feature data in which contours, irregularities, patterns, etc. of an object are approximated to straight lines or arcs from an input image is created, the feature data is selected and registered by a worker as model features, and further registered by the worker's judgment. The position of the existing model feature is moved and adjusted.

When feature extraction is performed from the input image,
In some cases, the characteristics of a certain part of the target object may be slightly deviated from the position originally obtained by the influence of the illumination or the surface state of the target object at the time of model production or the position obtained with a high probability. In this case, the feature is registered when the model is registered, and the operator moves to an appropriate position when the model is edited. This work is shown in FIG.

<Fourth Embodiment (Handling by Feature Rotation)>
In the present embodiment, feature data that approximates contours, irregularities, patterns, etc. of an object from an input image to straight lines and arcs is created, and the feature data is selected and registered by a worker as a model feature, and further registered by the worker's judgment. This is for rotating and adjusting the posture of the model feature that is being performed.

When feature extraction is performed from the input image,
In some cases, the characteristics of a certain part of the object are slightly deviated from the attitude originally obtained by the influence of the illumination or the surface state of the object during model production or the attitude obtained with a high probability. In this case, the feature is registered when the model is registered, and the worker rotates to an appropriate posture when the model is edited. This work is shown in FIG.

<Fifth Embodiment (Measures by Eliminating Characteristics)>
In the present embodiment, feature data in which contours, irregularities, patterns, etc. of an object are approximated to straight lines or arcs from an input image is created, and the feature data is selected and registered by a worker as a model feature, and an arbitrary feature is required. Accordingly, after adding to the model feature, the feature that is not appropriate as the model feature is deleted by the operator's judgment.

The features registered in the model are erroneously registered in the model registration, features that are registered as model features and have a low probability of being obtained as features from the input image, or are erroneously added in model editing. Features may be included. In this case, when the model is edited, the features that are not appropriate as the model features are deleted by the operator's judgment. This work is shown in FIG.

[0037]

As described above in detail with the embodiments and examples, the present invention has the following effects.

(1) In the feature extraction at the time of model creation, even if the feature to be registered in the model is not obtained, the operator can add an appropriate feature at the time of editing the model without performing the feature extraction process again. Model production work can be made more efficient.

(2) In the feature extraction at the time of model production, even if the feature to be registered in the model is slightly deviated from the position originally obtained by the influence of the illumination or the surface state of the object, or the position obtained with a high probability, The operator can streamline the model creation work by moving the feature to an appropriate position when editing the model without performing the feature extraction process again.
Further, the recognition accuracy can be improved.

(3) In the feature extraction at the time of model production, even if the feature to be registered in the model is slightly deviated from the posture originally obtained by the influence of the illumination or the surface state of the object, or the posture obtained with a high probability, The operator can make the model production work more efficiently by rotating the feature to an appropriate posture when editing the model without performing the feature extraction process again.
Further, the recognition accuracy can be improved.

(4) When there are several types of features of a part of an object, for example, the features of the same part are obtained as "straight lines" or "arcs" with similar probability, or "one straight line" or When it is possible to obtain a “set of multiple straight lines” with a similar probability, first register the feature obtained at the time of model feature registration, and then at the time of model editing, the worker identifies the feature of another appearance. The recognition stability performance can be improved by adding it to the part.

(5) In the model production work, the features incorrectly registered in the model registration work and the features incorrectly added in the model edition work are not the model again and the wrong features are judged by the operator during the model edition. By deleting, the model production work can be made more efficient.

(6) For features with a low probability of being obtained in the model feature data, the model creation work can be made more efficient by eliminating inappropriate features in the model editing process without recreating the model again. Performance can be improved.

[Brief description of drawings]

FIG. 1 is a flowchart showing a processing procedure according to an embodiment of the present invention.

FIG. 2 is an explanatory diagram conceptually showing a first example of straight line addition processing according to the first embodiment.

FIG. 3 is an explanatory diagram conceptually showing a second example of straight line addition processing according to the first embodiment.

FIG. 4 is an explanatory diagram conceptually showing a third example of the straight line addition processing according to the first embodiment.

FIG. 5 is an explanatory diagram conceptually showing a fourth example of the straight line addition processing according to the first embodiment.

FIG. 6 is an explanatory diagram conceptually showing a fifth example of straight line addition processing according to the first embodiment.

FIG. 7 is an explanatory diagram conceptually showing a sixth example of the straight line addition processing according to the first embodiment.

FIG. 8 is an explanatory diagram conceptually showing a first example of arc adding processing according to the second embodiment.

FIG. 9 is an explanatory diagram conceptually showing a second example of the arc addition processing according to the second embodiment.

FIG. 10 is an explanatory diagram conceptually showing the feature movement processing according to the third embodiment.

FIG. 11 is an explanatory diagram conceptually showing the feature rotation processing according to the fourth embodiment.

FIG. 12 is an explanatory diagram conceptually showing the arc addition processing according to the fifth embodiment.

FIG. 13 is an explanatory view conceptually showing a conventional problem.

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Makoto Niwakawa 2-17-1 Osaki, Shinagawa-ku, Tokyo Stock company Shameidensha

Claims (6)

[Claims] 1. Recognizing an object by matching the characteristic data obtained by approximating the contour, unevenness, pattern, etc. of the object obtained from the input image with a straight line or an arc and the characteristic data of the model of the object. In the object recognition device that performs the above, the operator selects this feature data as a model feature based on the feature data obtained by approximating the features such as contours, unevenness, and patterns of the object obtained from other input images to straight lines and arcs. An object recognition device characterized by having a model creating means configured to have a function of allowing a worker to edit and to register model feature data such as a straight line and an arc. 2. The model producing means has a function of allowing an operator to select and register the feature data as a model feature and further add an arbitrary straight line to the model feature [claim 1]. The object recognition device described in 1. 3. The model creating means has a function of allowing an operator to selectively register the feature data as a model feature and further add an arbitrary arc to the model feature [claim 1]. The object recognition device described in 1. 4. The model creating means has a function of allowing an operator to select and register the feature data as a model feature and to move and adjust the position of the registered model feature according to the operator's judgment. The object recognition device according to claim 1 characterized. 5. The model creating means has a function of allowing an operator to selectively register the feature data as a model feature, and further to rotate and adjust the posture of the registered model feature according to the operator's judgment. The object recognition device according to claim 1 characterized. 6. The model creating means allows an operator to select and register the feature data as a model feature, add an optional feature to the model feature as necessary, and then select a feature that is not appropriate as a model feature of the operator. It has a function that can be erased by judgment [Claim 1].
The object recognition device described in 1.