JP3372419B2 - Object recognition method - 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 recognition method for measuring the position and shape of an object based on an image of the object.

[0002]

2. Description of the Related Art Conventionally, various object recognition techniques for measuring the position and shape of an object based on an image of the object are known. In this kind of object recognition technology, it becomes a polygon (rectangle, square, regular hexagon, etc.) that has circles, ellipses, cruciforms, and even number of sides in the image, and the opposite sides have the same length. A method for obtaining the center position of such a recognition target (a positioning mark on a printed circuit board is illustrated) is proposed in Japanese Patent Laid-Open No. 5-107030.

In the technique described in the above publication, the recognition target and the background are separated by binarizing a grayscale image obtained by imaging the recognition target. After that, as shown in FIG. 25, the intersections between the horizontal scanning line Lh and the vertical scanning line Lv and the boundary of the recognition target Ob are obtained by scanning in the horizontal direction and the vertical direction, respectively. Since two intersections are obtained for each horizontal scanning line Lh and each vertical scanning line Lv, a midpoint R between the two points on each horizontal scanning line Lh and each vertical scanning line Lv is obtained. When the histogram of the midpoint R thus obtained is obtained in the horizontal direction and the vertical direction, the point (mode) at which the frequency becomes maximum can be regarded as the position of the midpoint in the horizontal direction and the vertical direction. Of. That is,
Even if the edge (boundary line) of the recognition target Ob has some deformation such as chipping, if most of the edges have a normal shape, the center position of the recognition target Ob can be obtained as described above. .

[0004]

However, in the above method, the density value of the pixel is binarized by an appropriate threshold value when the recognition object Ob and the background are separated. When the brightness changes gently between the target Ob and the background (for example, when the change of the density value is illustrated in FIG. 26), the threshold values Ta and Tb to be set are set.
Accordingly, the position of the edge of the recognition target Ob changes. In the illustrated example, when the threshold value is set to a large value, the recognition target O is detected in the binary image after binarization as shown in FIG.
If b is small and the threshold value is set to a small value,
As shown in 7 (b), the recognition target Ob becomes large in the binarized image after binarization (appropriate edge positions are indicated by broken lines). Therefore, in the method of binarizing the density value with the threshold value, the edge position changes depending on the setting method of the threshold value depending on the recognition target Ob, and the center position obtained by the above method and the actual position The deviation from the center position may increase.

Further, as shown in FIG. 28 (a), the recognition target O
When the center line of b does not coincide with the horizontal direction or the vertical direction of the image, there is no point where the frequency becomes maximum in the histogram of the middle points as shown in FIG. The center position cannot be detected.
Further, in the method described in the above publication, since the midpoints of both ends of the intersections of the horizontal scanning line Lh and the vertical scanning line Lv with the edge of the recognition target Ob are used,
As shown in FIG. 29A, one scanning line (horizontal scanning line Lh or vertical scanning line Lv) has a plurality of recognition targets O.
When passing through b ₁ and Ob ₂ , the intersections are obtained from the recognition objects Ob ₁ and Ob ₂ located at both ends of the scanning line. That is, according to the method described in the above publication, when a plurality of recognition targets Ob ₁ and Ob ₂ exist, FIG.
Recognized Ob _1, power is maximized at the intermediate position of the Ob ₂ as would not be able to obtain the center position of each recognition target Ob _1, Ob _2. Further, with the method described in the above publication, the center position of the recognition target Ob can be obtained, but the shape of the recognition target Ob cannot be recognized.

The present invention has been made in view of the above circumstances, and an object thereof is to detect the edge between the recognition target and the background with good reproducibility when obtaining the midpoint of the recognition target in the image, and to recognize the recognition target. The object of the present invention is to provide an object recognition method capable of accurately obtaining the position of a representative point of an individual recognition target regardless of the direction and the number of targets and further recognizing the shape of the recognition target.

[0007]

According to a first aspect of the present invention, a grayscale image obtained by picking up an image of a polygonal object in which the shape of a boundary line of interest is point symmetry or at least two pairs of opposite sides are parallel to each other is formed. The point where the differential value of each pixel changes is detected as an edge to be recognized, the edge direction that can define the direction of the edge is determined for each point on the edge, and two points on the edge where the edge direction is equal are paired. As a feature, a histogram relating to the coordinate value of the midpoint between the pair of points is created, and the coordinate value at which the maximum frequency is obtained in the histogram is adopted as the coordinate of the representative point to be recognized.

In this method, the grayscale image including the object is not binarized by the threshold value of the density, but the grayscale image is differentiated and the change point of the differential value is detected as the edge of the recognition target. It is possible to extract an edge with good reproducibility even when a point including a point that changes relatively slowly is included. In addition, since a histogram regarding the coordinate value of the midpoint between a pair of points with the same edge direction on the edge is created and the coordinate value that gives the maximum frequency is adopted as the coordinate of the representative point to be recognized,
Since the representative point of the recognition target is statistically calculated based on a large number of points, the position of the representative point of the recognition target can be accurately calculated. Here, since the edge direction is a direction that can define the direction of the edge, it usually means a tangential direction or a normal direction.

According to a second aspect of the present invention, in the first aspect of the present invention, the point pairs in which the edge directions are equal on the edges,
It is characterized in that the limiting condition is set such that only the point pair in which the midpoint between the point pairs is located near the representative point to be recognized is selected to obtain the midpoint. In this method, it is possible to eliminate a point pair unnecessary for obtaining the representative point of the recognition target out of a large number of point pairs by setting a limiting condition, so that the representative point of the recognition target can be accurately measured with a relatively small amount of processing. You can ask well.

According to a third aspect of the present invention, in the first or second aspect of the invention, an edge of a line of points formed by a pair of points having the coordinate value at which the maximum frequency is obtained in the histogram as the coordinate of the middle point is arranged. The feature is that the graphic feature of the recognition target is extracted based on the feature amount. In this method, the point sequence consisting only of the point pairs used to determine the representative point of the recognition target is used as the points on the edge, so it is necessary to extract the point sequence which is considered to well reflect the shape of the recognition target. become,
By using such a sequence of points, it becomes possible to accurately recognize the graphical feature of the recognition target.

According to a fourth aspect of the present invention, in the second aspect of the invention, the limiting condition is set so that only the point pair having a difference in concentration gradient within the predetermined range is selected from the point pairs to obtain the midpoint. It is characterized by having done. The invention of claim 5 relates to claim 2
In the invention of (1), the middle point is obtained by selecting only the point pair having the direction of the concentration gradient with respect to one point of the point pair on the center line and having the other point within the predetermined angle range from the center line. The above-mentioned limiting condition is set.

According to a sixth aspect of the invention, in the second aspect of the invention, the limiting condition is set so that only the point pair in which the distance between the point pairs is within a predetermined range is selected to obtain the midpoint. Characterize. The invention according to claims 4 to 6 is a specific embodiment of the limiting condition, in which the position of the representative point of the recognition target is obtained based on the difference in concentration gradient between the point pairs, the positional relationship between the point pairs, and the like. Since the position of the representative point to be recognized is obtained by excluding the point pairs which are apparently unnecessary, the processing is simplified and the accuracy of the obtained position of the representative point is increased.

The invention of claim 7 is characterized in that, in the invention of claims 1 to 6, the grayscale image is a color image, and a midpoint is obtained for a recognition target existing in a region of a specific color range. To do. By thus limiting the color range, when it is necessary to individually obtain the position of the representative point for a plurality of objects of different colors, it is possible to identify each object individually by color.

The invention according to claim 8 is characterized in that, in the invention according to any one of claims 1 to 6, the histogram is created for the remaining edges from which some of the edges have been deleted. In this method, a histogram is created for the remaining edges after deleting some of the edges, so if there are multiple objects, create the histogram after deleting the edges of unnecessary objects in advance.
Meaningless information is less likely to occur, and the position of the representative point to be recognized can be accurately obtained.

According to a ninth aspect of the present invention, in the third aspect of the invention, the shape of the recognition target is obtained by obtaining the distribution of the density gradient for each point of the point sequence and comparing the density gradient distribution pattern with a known distribution pattern. It is characterized by recognizing. According to this method, since the distribution pattern of the density gradient for each point of the point sequence is obtained, the shape of the recognition target can be recognized by the pattern matching of the distribution pattern of the density gradient. Moreover, since the point sequence for obtaining the distribution pattern of the density gradient is the point sequence used when obtaining the midpoint of the recognition target, it is considered that the shape of the recognition target is well reflected.
By determining the shape of the recognition target from such a sequence of points, the shape of the recognition target can be accurately grasped.

According to a tenth aspect of the invention, in the invention of the third aspect, the distribution of the concentration gradient for each point of the point sequence is obtained,
The feature is that the orientation of the recognition target is detected based on the distribution pattern of the density gradient. In this method, the orientation of the recognition target can be obtained based on the density gradient distribution pattern by obtaining the density gradient distribution for each point in the point sequence. Moreover, since the point sequence for obtaining the distribution pattern of the density gradient is the point sequence used for obtaining the coordinates of the representative point of the recognition target, it is considered that the figure characteristic of the recognition target is well reflected. By obtaining the orientation of the recognition target from the point sequence, the orientation of the recognition target can be accurately grasped.

The invention of claim 11 is characterized in that, in the invention of claim 3, the number of the point sequences is obtained, and the size of the recognition target is detected based on the number of the point sequences. With this method, the size of the recognition target can be obtained based on the number of point sequences. Moreover, since the point sequence for obtaining the number of points is the point sequence used when obtaining the midpoint of the recognition target, it is considered that the figure characteristic of the recognition target is well reflected, and the number of such point sequences is calculated. As a result, the size of the recognition target can be obtained with good reproducibility.

According to a twelfth aspect of the present invention, in the first aspect of the present invention, the first step of deleting the edge to be recognized having the coordinate value of the representative point whose coordinate value gives the maximum frequency in the histogram, and after the first step It is characterized in that the coordinates of the representative points of the plurality of recognition targets are individually obtained by repeating the second step of recreating the histogram for the remaining edges.

According to this method, when there are a plurality of recognition targets of different sizes, the positions of the representative points are calculated in order from the largest recognition target, and the edges of the recognition target for which the positions of the representative points are calculated are excluded. Since the histogram is created in this manner, the position of the representative point of each recognition target can be individually obtained. According to a thirteenth aspect of the present invention, in the first aspect of the invention, the coordinates of the representative point of each recognition target for the object having a plurality of recognition targets are obtained, and the position of the object and the position of the object are determined based on the positional relationship of the obtained coordinates of the representative points. It is characterized by recognizing the direction.

According to this method, since the positional relationship between the representative points of the respective recognition targets is obtained from the object having a plurality of recognition targets, the position and orientation of the object can be recognized. According to a fourteenth aspect of the invention, in the first aspect of the invention, the coordinates of the representative point of the recognition target are obtained for a plurality of grayscale images obtained by imaging the object from a plurality of different positions, and the representative of the obtained recognition targets is obtained. After associating the coordinates of the points,
The feature is that the distance to the object is obtained by applying the triangulation method.

According to this method, when the distance to the recognition target is calculated using the stereo image, the position of the representative point of the recognition target calculated by the method of claim 1 is used to associate a plurality of images, which is high. Correspondence can be performed with accuracy. As a result, the distance to the recognition target can be accurately obtained. According to a fifteenth aspect of the present invention, in the third aspect of the present invention, a plurality of grayscale images obtained by imaging an object from a plurality of different positions are provided.
The coordinates of the representative point of the recognition target are obtained respectively, and the graphic feature of the recognition target is extracted based on the feature amount of the edge formed by the point sequence formed by the point pairs for which the coordinates of the representative point are obtained, and the same in each image. After associating the regions having the graphic features of, the distance to the object is obtained by applying the triangulation method.

According to this method, when the distance to the recognition object is obtained using the stereo image, the graphic feature of the recognition object obtained by the method of claim 3 is used to associate a plurality of images, so that the accuracy is high. Can be matched with. As a result, the distance to the recognition target can be accurately obtained. The invention of claim 16 relates to claim 1
In the invention of claim 1, for a plurality of grayscale images obtained by imaging an object having a plurality of recognition targets from a plurality of different positions, the coordinates of the representative point of each recognition target are calculated from the respective images, and are calculated in each image. After associating the coordinates of the representative points with each other, the triangulation method is applied to obtain the positions of the coordinates of the representative points of each recognition target in the three-dimensional space, and the three-dimensional space is calculated based on the positional relationship of the coordinates of the representative points. It is characterized by recognizing the position and orientation of an object inside.

According to this method, when three-dimensional information is obtained using a stereo image, the positions of a plurality of positions of the object are obtained, and the position and orientation of the object are recognized based on the positional relationship between them. Claim 1 for associating images of
Since the position of the representative point of the recognition target obtained by the method is used, the three-dimensional information of the object can be obtained with high accuracy and reproducibility.

[0024]

DETAILED DESCRIPTION OF THE INVENTION

(Embodiment 1) FIG. 1 is a block diagram showing a schematic configuration of an apparatus according to this embodiment. This device is basically realized by using a computer device. The recognition target described below has a point-symmetrical shape, and the case where the representative point is the center will be described. However, the recognition target is not necessarily limited to this, and as a recognition target, at least two pairs of facing sides are parallel to each other. It also includes polygons. In this case, the position of the representative point is not the center, but a position that can be defined as the representative point can be obtained. Although the tangential direction is basically used as the edge direction below, the normal direction may be used as the edge direction when either the tangential direction or the normal direction may be used.

The TV camera 1 picks up an image of a spatial region including a target object, and the light source (not shown) for illuminating the object and the TV camera 1 are the amount of light from the object incident on the TV camera 1. The positional relationship is set so that there is a sufficient difference between the light amount from the background and the amount of light from the background. In addition, the TV camera 1 that can obtain an output according to the amount of incident light is used.
The output of the V camera 1 is converted into a digital signal by the A / D converter 2 and then stored in the frame memory 3. That is, a grayscale image is stored in the frame memory 3.
The grayscale image stored in the frame memory 3 is the differential processing unit 4
Is spatially differentiated at.

In the differential processing section 4, the edge E of the recognition object Ob is obtained as described above, and the direction in which the density gradient is the largest for each point (pixel) on the edge E is the normal direction of that point. (Which means a direction substantially orthogonal to the edge E). Here, the recognition target O
b is an elliptical shape, and the edge E obtained by the differential processing unit 4 has a shape as shown in FIG. The normal direction is indicated by the arrow n in the figure. Various methods are conceivable for obtaining the absolute value | G | of the differential value and also obtaining | θ | in the direction in which the concentration gradient becomes maximum (almost coincident with the normal direction: hereinafter referred to as differential direction value). However, the Sobel method is used here. That is, the following equation is calculated using a well-known operator as a Sobel operator. Now, the pixel (point) of interest has coordinates (x, y) (where x is the horizontal coordinate value and y is the vertical coordinate value,
Here, it is assumed that the right direction is the positive direction in the horizontal direction and the upward direction is the positive direction in the vertical direction. If the density is I _{x, y} , the first-order differential values G _x , G _{y in} each direction. Is expressed by the following equation. G _x = (I _{x + 1, y−1} + 2I _{x + 1, y} + I _{x + 1, y + 1} ) − (I
_{x-1, y-1} + 2I _{x-1, y} + I _{x-1, y + 1} ) G _y = (I _{x-1, y-1} + 2I _{x, y-1} + I _{x + 1, y-1} ) − (I
_{x-1, y + 1} + 2I _{x, y + 1} + I _{x + 1, y + 1} ) Using these first-order differential values G _x and G _y , the absolute value | G | of the differential value and the differential direction value | θ | can be expressed by the following equation. | G | = (G _x ² + G _y ² ) ^1/2 | θ | = tan ⁻¹ (G _y / G _x ) When the differential value of the differential value | G | It is possible to slice the differential value by the threshold value and extract the ridge line having the maximum absolute value | G | within the area equal to or more than the threshold value as the edge E. The edge E obtained in this way may be discontinuous, but since the coordinates of the center of the recognition target Ob are obtained using a statistical method in the method described later, even if the edge E is discontinuous, there is a particular problem. Does not occur.

The normal direction n is quantized with reference to the horizontal or vertical direction in the image. For example, when the angle is set counterclockwise with the right direction in the horizontal direction as a reference line and sections are set at equal angular intervals, and the normal direction n obtained as described above is included in each section, the section is represented. The angle to be assigned is assigned. Alternatively, an integer value may be assigned to the quantized angle, known as the direction code. When the normal direction n is determined in this way, the direction orthogonal to the normal direction n is defined as the edge direction (meaning the direction in which the edge E can be defined). Where the edge direction is 0
It shall be determined within the range of ~ 180 °. That is, when the normal direction n differs by 180 °, the edge directions are the same. The edge direction is indicated by a straight line t in the figure.

The point pair extraction unit 5 classifies and extracts points for each edge direction t based on the edge direction t obtained by the differential processing unit 4. Also, two points having the same edge direction t are combined to form a pair. By the way, when the recognition target Ob is an ellipse or a circle as shown in FIG. 2A, the midpoint C of the line segment connecting the pair of points is expected to pass near the center of the recognition target Ob. In other words, a space having the same coordinate space as the coordinates on the image and a coordinate value is given to each coordinate value as shown in FIG. 2B is set, and the points between pairs of points for the recognition target Ob of an ellipse or a circle are set. Find the coordinates of the midpoint C of
When a histogram is created by repeating the operation of adding 1 to the frequency of the coordinate value in the space, the coordinate of the midpoint C between the pair of points is the center (xc, y) of the recognition target Ob.
It becomes the largest near c).

When the recognition object Ob is a point-symmetrical polygon, for example, a parallelogram as shown in FIG. 3A, there are many points having the same edge direction t. If the midpoint C of a pair of all points having the same edge direction t is obtained, the existence probability of the midpoint C is expected to be the highest near the center (xc, yc) of the recognition target Ob. That is, if the frequency of the coordinates of the middle point C between the pair of points is calculated, the coordinates of the middle point C between the pair of points are calculated as shown in FIG. 3B.
A histogram is obtained which is the largest in the vicinity of c).

Although the above discussion assumes that the edge E of the recognition object Ob does not have a chip or a bulge, the above description is given even if the edge E has some chip or bulge. The existence probability of the midpoint C of the pair of points obtained as described above becomes highest near the center of the recognition target Ob, and
The position of the center of can be calculated. However, when the center position of the recognition target Ob is obtained, obtaining the middle point C for all combinations of paired points results in many useless combinations, resulting in poor processing efficiency.
Therefore, the condition verification unit 6 limits the occurrence of useless combinations by extracting pairs of points that satisfy the following specific conditions. In short, the condition verification unit 6 selects only the pair of points having a high probability that the midpoint between the points extracted by the point pair extraction unit 5 substantially coincides with the center of the recognition target Ob. .

First Method for Limiting Useless Combinations
Is a method that uses the difference in the normal direction n of each point extracted as a pair
Is the law. That is, the recognition target Ob is a circle, an ellipse, or point symmetry.
If the polygon is a simple polygon, the middle point of the pair of points is the recognition target Ob.
When coincident with the center, the normal direction n of both points is 180 °
Because of the difference, the normal direction n is about 180 °
You can find the center of the recognition target Ob by selecting only
It is thought to be. Therefore, it is extracted by the point pair extraction unit 5.
The difference in the normal direction n between the paired points is calculated, and the difference is specified
For pairs of points that are within (for example, 180 ° ± 5 °)
Adopt only. For example, as shown in FIG.
Then, a pair of points on the same side (P in the figure₁, P ₂)
Of the points on the side (point P in the figure₁, P₃) And the edge direction
Although t is equal, the above-mentioned restriction condition regarding the normal direction n
If used, point P₁, P₂Is removed and the point P₁, P₃
The midpoint is sought only for the pair. That is, this
The processing amount is almost halved only by using the limiting conditions of
It

The second method of limiting the useless combinations is to limit the area where the other exists by the normal direction n of one of the points extracted as a pair, and extract only the pairs where the other point exists in that area. Is the way to do it. Here, at each point, the recognition target Ob has a higher density than the background. That is, an area of an appropriate angle range is set with one point extracted as a pair of points in the point pair extraction unit 5 as the center and the normal direction as the center line.
Now, as shown in FIG. 5, two elliptical recognition targets Ob
Assuming that ₁ and Ob ₂ exist apart from each other in the image, there are four points having the same edge direction t in the image, so that there are six types of point pairs. For example, the point P ₄ shown in FIG.
If the edge directions t of P ₅ and P ₆ are equal, the point P
The pair of P ₄ and P ₅ can be used to determine the center of the recognition object Ob ₁ , but the middle point of the pair of points P ₄ and P _{6 and} the pair of points P ₅ and P ₆ is the recognition object Ob _1. , Ob ₂ centers are also irrelevant. Then, for example, paying attention to the point P ₄ ,
Appropriate angle range centered line normal direction n of the point P _{4 4} around a (for example, ± 5 °) by setting the area D ₁ of the point P ₅ is included in this region D ₁ Therefore, the midpoint of the point pair can be obtained without considering other points having the same edge direction t (such as P ₆ ).

A third method for limiting useless combinations is to use the distance between points extracted as a pair.
That is, as shown in FIG. 6, assuming that two elliptical recognition targets Ob ₁ and Ob ₂ exist apart from each other in the image, four points having the same tangential direction t exist. , If the distance range between the points extracted as a pair is limited, the points P ₇ , P ₈ and P ₉ shown in the figure are
It is possible to select only the points P ₇ and P _{8 so} that the pair of points P ₇ and P _{9 and} the pair of points P ₈ and P ₉ are not the targets of the pair for which the midpoint is sought.

The fourth method of limiting the useless combinations is a method of ignoring a specific component of the edge E of the recognition object Ob and obtaining the center position using only a part of the edge E. For example, as shown in FIG. 7, when the recognition target Ob has an elliptical shape and both side edges are straight lines, when paired points are extracted from the straight line portion Ea, a pair of paired points is combined. Will be very many. On the other hand, in this recognition object Ob, since the center position can be obtained using only the curved portion Eb, the straight line portion Ea can be ignored in obtaining the center of the recognition object Ob. Therefore, only the pair of points existing in the curved line portion Eb is selected from the pair of points. If this method is adopted, when the center position of the recognition target Ob is obtained, rapid processing can be performed without using useless pairs of points.

As described above, the straight line portion Ea and the local line portion Eb
The following method can be used to separate and. In other words, from the start point set appropriately on the edge E to the edge E
To determine the relationship between the directions before and after the point of interest. For example, when the direction of the points before and after the tracking is within a predetermined angle (for example, 10 degrees), it is considered as a straight line portion Ea, and when it exceeds the predetermined angle, it is determined as a curved portion Eb.

The first to fourth methods can be used independently, but if these are used in an appropriate combination, a pair of unnecessary points can be more reliably obtained when the center position of the recognition object Ob is obtained. Can be removed. For example, in the position where the region D ₁ is set to the second method may include the point where the recognition object Ob ₂ other in areas D ₁ set in one of the recognition target Ob ₁ is tangential at the same However, there is a possibility that such a point can be eliminated by adopting the third method together.

For the pair of points selected by the condition verification section 6, the midpoint detection section 7 obtains the coordinates of the midpoint of both points.
The midpoint detection unit 7 calculates the coordinate values of the midpoint as shown in FIG.
The histogram as described with reference to (b) and FIG. 3 (b) is created. Here, since the limiting condition in the condition verification unit 6 is applied, as compared with the histograms shown in FIG. 2B and FIG. 3B, other conditions for the frequency of coordinates near the center of the recognition target Ob are compared. A frequency distribution in which the frequency of the coordinate values is significantly small can be obtained. That is, the recognition target O
Unnecessary information can be greatly reduced in obtaining the center coordinates of b, and the center coordinates can be accurately obtained with a small amount of processing.

In the histogram as described above, the frequency near the coordinate corresponding to the coordinate of the midpoint of the recognition target Ob is maximized, so that the central coordinate detection unit 8 has the coordinate value (xc, yc) that maximizes the frequency in the histogram. Is used as the coordinates of the center position of the recognition target Ob. Here, since the coordinate value of the frequency counting space is quantized, the frequency of one coordinate value is maximum in most cases, but the frequency that becomes the same or almost the same around the coordinate value with the maximum frequency is Occurrence of coordinate values. In such a case, the average coordinate value of the coordinate values whose frequency is within the predetermined range may be used as the coordinate of the center position.

Assuming that the recognition target Ob is an ellipse as shown in FIG. 2A, it is possible to obtain coordinates (xc, yc) which substantially coincide with the center C of the ellipse, and FIG. If the recognition target Ob is a parallelogram as shown in (a), it is possible to obtain coordinates (xc, yc) that substantially match the intersection D of the diagonal lines. The above processing is summarized as shown in FIG. That is, the object is imaged, the grayscale image is stored in the frame memory 3 (S1), the edge E is extracted by the differential processing in the differential processing unit 4, and the normal direction n (density gradient) of each point on the edge E is extracted. ) Is obtained (S2). Next, edge E
, The point pairs having the same tangential direction are obtained (S3), and it is determined whether or not each point pair satisfies the restriction condition (S4). For all point pairs satisfying the restriction conditions, the coordinates of the midpoint between the point pairs are obtained, and a histogram of these coordinate values is created (S5). In this way, if there are no other point pairs (S
6) The coordinate value of the maximum frequency is calculated from the created histogram, and this is set as the center position of the recognition target Ob (S7).

As described above, since the edge E of the recognition object Ob is extracted and the edge direction is obtained by differentiating the grayscale image, the change point of the density gradient is determined even for the object whose density change is relatively gentle. By extracting, it becomes possible to extract the boundary line of the object with good reproducibility. In other words, instead of binarizing the density with the threshold value as in the conventional case, the differential value is binarized with the threshold value, so that the boundary line of the object can be accurately extracted. Further, since only the pairs of points satisfying the above conditions are extracted, it is possible to reduce the pairs of unnecessary points when obtaining the recognition object Ob, and the plurality of recognition objects O
Even if b ₁ and Ob ₂ exist, the center position can be obtained separately.

In this embodiment, the gradient direction of the density change is expediently adopted as the normal direction for each point of the edge E because the process of obtaining is simple, but the actual normal direction n is Since it is geometrically determined based on the shape of the edge, processing for geometrically determining the normal direction or the tangential direction may be performed. (Embodiment 2) This embodiment is an example applicable when there are a plurality of objects in the field of view of the TV camera 1 and the colors of the objects are different, and the specific color extraction unit as shown in FIG. The structure is almost the same as that of the first embodiment except that the ninth embodiment is provided. A TV camera 1 capable of picking up a color image is used, and the frame memory 3 has an R (red) G color.
The grayscale image is stored in the form of a combination of basic colors such as (green) and B (blue). The specific color extraction unit 9 extracts a desired color range (hue, saturation,
A region having at least one range of lightness) is extracted. In other words, even when you image multiple objects,
The recognition target Ob can be individually extracted according to the color of the object.

The subsequent processing is the same as that of the first embodiment. The point pair extraction unit 5 obtains a pair of points having the same tangential direction, and the condition verification unit 6 selects only the pair of points satisfying the restriction condition. The midpoint detection unit 7 finds the midpoint of a pair of points and gives the frequency to the frequency counting space. After obtaining the frequency in this way,
The central coordinates of the recognition target Ob are calculated by the central coordinates detection unit 8. The point pair extraction unit 5, the condition verification unit 6, the midpoint detection unit 7, and the center coordinate detection unit 8 have the same functions as those in the first embodiment, and therefore their explanations are omitted.

Here, a monochrome TV is filtered through a filter.
The specific color range can also be extracted by imaging the spatial area with the camera 1. In this case, the specific color extraction unit 9 may not be provided. (Third Embodiment) Although the first embodiment has described the technique for obtaining the center position of the recognition target Ob, the present embodiment will describe a technique for obtaining not only the center position of the recognition target Ob but also the shape thereof. A schematic configuration of the apparatus of this embodiment is shown in FIG.
Is shown as a block diagram.

In FIG. 10, the configuration denoted by the same reference numerals as in FIG. 1 has the same function, but in the present embodiment, the center position of the recognition object Ob is obtained without providing the condition verification unit 6. Further, in order to obtain the shape of the recognition target Ob, the edge extraction unit 10 that extracts the edge E used when obtaining the center position, and the recognition target Ob based on the edge E obtained by the edge extraction unit 10
The shape recognition unit 11 for recognizing the shape of is added. That is, the edge extraction unit 10 obtains the edge E from the arrangement of the point sequences at both ends of the line segment having the midpoint used when obtaining the center position of the recognition target Ob. Further, when the center position of the recognition target Ob is obtained by averaging a plurality of coordinate values, the points at both ends of the line segment having the midpoint used when the center position of the recognition target Ob is obtained for each coordinate value. The edge E is obtained from the row of columns.

That is, similarly to the process of obtaining the center position of the recognition target Ob, the process of obtaining the midpoint of the point pair on the edge E is performed, and the position of this midpoint is the center of the recognition target Ob already obtained. When within a predetermined distance from
The point pair is used as the edge E used when obtaining the center position of the recognition target Ob. By the way, the shape recognition unit 11 obtains the feature amount of the edge E by the method described below, and recognizes the shape of the recognition target Ob based on the feature amount.

In the first method of recognizing the shape of the recognition object Ob, the distribution in the direction in which the density gradient of each point included in the point sequence used when the edge extraction unit 10 extracts the edge E is maximized is used. . That is, the distribution of the values used in the normal direction n in the first embodiment is obtained. For example, when the recognition target Ob is elliptical as shown in FIG. 11A, the frequency distribution in the normal direction n is as shown in FIG. 11B. Although the normal direction n is shown as a continuous value in the figure, it is actually a discrete value obtained by dividing 0 to 360 ° by 18 or 36. Therefore, the frequency distribution in the normal direction n has a shape close to a sine waveform. Further, the recognition target Ob is shown in FIG.
If it is a parallelogram as shown in (a), all the points on one side have the same normal direction n.
There are four angles with high frequencies as in (b),
The ratio of each angle is proportional to the length of each side.
Therefore, it is possible to identify whether it is a circle (the frequency distribution is linear), an ellipse (the frequency distribution is sinusoidal), or a polygon (the frequency peak appears discretely) by the frequency distribution pattern in the normal direction n.

The shape recognition unit 11 may detect the orientation of the recognition target Ob together. In order to detect the orientation of the recognition target Ob, the frequencies in the normal direction n in the range of 0 ° to 360 ° are sequentially obtained, and the angle at which the frequency becomes the first maximum value is obtained. For example, if the recognition target Ob is an ellipse, the maximum value of the frequency is 0 ° to
It appears twice in the range of 360 °, and when the frequencies are sequentially obtained starting from 0 ° as described above, the first maximum value appears in the range of 0 ° to 180 °. Since this angle is substantially orthogonal to the tangential direction parallel to the major axis of the ellipse, the direction of the major axis of the ellipse can be obtained. If the recognition target Ob is a parallelogram, the maximum value of the frequency appears four times in the range of 0 ° to 360 °, and 0 ° to 9 °.
The first maximum appears in the range of 0 °. This angle approximately corresponds to the direction of the short side or long side of the parallelogram (including rectangle). Therefore, the maximum value of these maximum values is obtained, and if there are two maximum values, the one with the smaller angle is defined as the direction of the parallelogram. In order to obtain the maximum value of the frequency distribution in the normal direction n, first, one-dimensional smoothing is performed along the angle axis to remove noise components, and then a one-dimensional differential filter is applied to obtain a differential curve, and then a differential curve is obtained. , The point at which the sign of the differential value changes from positive to negative along the angle axis is determined, and when the frequency at this point is equal to or greater than the predetermined threshold value, that point is adopted as the maximum value of the frequency. .

Further, in the shape recognition section 11, the recognition target Ob
May also be detected together. The number of points on the edge E that contributed when obtaining the center position of the recognition target Ob, that is, the frequency of the peak value in the histogram is
Since it is considered that the recognition target Ob is substantially proportional to the circumference of the recognition target Ob, the size of the recognition target Ob can be obtained by the frequency in the histogram for the similar recognition target Ob.
For example, when there are two objects of different sizes in the field of view of the TV camera 1, there are two large and small recognition targets Ob ₁ and Ob ₂ as shown in FIG. 13A. Will be done. Therefore, after the recognition targets Ob ₁ and Ob ₂ are individually recognized by the method described in the first embodiment, the midpoint between the two points having the same tangential direction on the edge E is taken and the histogram is set to the middle. Taking the appearance frequency of the coordinates of the points,
As shown in FIG. 13B, a frequency distribution having two peaks will be obtained. Here, if the recognition target Ob _1, Ob ₂ is a similar, as described above, since the magnitude of power reflects the area of each recognition target Ob _1, Ob _2, recognized Ob ₁ based on the frequency value , Ob ₂ can be known. This method is applicable not only when there are two objects but also when there are three or more objects.

(Embodiment 4) In this embodiment, as shown in FIG. 14, the condition verifying unit 6 is omitted and an edge erasing unit 12 is added to the configuration of Embodiment 1 shown in FIG. Therefore, as shown in FIG. 15A, a plurality of recognition targets Ob ₁ , Ob ₂ , and Ob ₃ exist in the image, and some recognition targets Ob ₁ , Ob ₂ , and Ob ₃ overlap. In this case, it is possible to individually extract the center positions of the recognition targets Ob ₁ , Ob ₂ , and Ob ₃ .

That is, the point pair extraction unit 5 obtains the midpoint of the pair of points selected by the same method as in the first embodiment,
When the frequency of the coordinates of the midpoint is calculated by the midpoint detecting unit 7 for all pairs of points, a plurality of peaks appear in the histogram as shown in FIG. Fig. 15 shows the coordinates of the peak.
Because they also contain extraneous peaks for recognition Ob _1, Ob _2, the center position C ₁ of the Ob _3, C _2, C ₃ as shown in (c), the recognition object Ob _1, Ob _2, Ob ₃ In order to obtain the center positions C ₁ , C ₂ and C ₃ of P, the irrelevant peaks must be removed by some method. Since the irrelevant peak has a relatively low frequency, it is easy to extract the peak with a high frequency as the center of the recognition target Ob ₁ , but the peak with a low frequency is the center of the recognition targets Ob ₂ and Ob ₃ . Hard to identify.

Therefore, in the present embodiment, the edge erasing unit 1
2 is used to remove the edge E including the point that contributed to obtaining the center coordinate of the recognition target Ob ₁ having the coordinate with the maximum frequency as the center. After that, the remaining edge E
On the other hand, the point pair extraction unit 5 extracts the pair of points again, obtains the coordinates of the middle point of the pair of points, and obtains the histogram.
In the state in which the edge E of the recognition target Ob ₁ having the coordinate with the maximum frequency as the center is deleted, the state becomes as shown in FIG. 15D, and therefore the histogram has only about two peaks as shown in FIG. 15E. It will not appear. In this state, the edge E of the recognition object Ob ₂ having the higher frequency is erased, and the frequency of the middle point is obtained again for the remaining edge E to determine the coordinates of the middle point. By repeating such processing until the edge E disappears, the centers C ₁ , C ₂ , C ₃ of all the recognition objects Ob ₁ , Ob ₂ , Ob ₃ can be individually obtained. Other configurations and operations are similar to those of the first embodiment.

(Fifth Embodiment) In this embodiment, the center of the recognition object Ob is obtained from the pair of points satisfying a predetermined condition as in the first embodiment. The shape, orientation, and size can be obtained. Therefore, as shown in FIG. 16, the edge extraction unit 10 and the shape recognition unit 11 are added to the configuration of the first embodiment shown in FIG. 1 as in the third embodiment.

According to the present embodiment, wasteful processing can be reduced by limiting the pair of points used for obtaining the center position when obtaining the center position of the recognition object Ob, and the unnecessary processing in the histogram is unnecessary. It is possible to suppress the occurrence of peak values. Further, not only the position of the center of the recognition target Ob but also the shape, orientation, size, etc. can be obtained. In FIG. 16, the portions denoted by the same reference numerals as those in FIGS. 1 and 10 function similarly.

(Sixth Embodiment) In this embodiment, the recognition target Ob
The shape and direction of the object can be recognized, and the method is applicable to the case where the object has a relatively complicated shape. That is, as shown in FIG. 18A, the case where there are portions Q ₁ , Q ₂ , Q ₃ where the edge E is circular for one recognition target Ob is targeted. An object from which such an image can be obtained is, for example, an object having three circular openings.

In this embodiment, as shown in FIG.
The configuration has a configuration in which the shape recognition unit 11 is added to the configuration of the first embodiment shown in FIG. 1, and the operation of the shape recognition unit 11 is characterized. That is, the condition verification unit 6 applies the respective methods described in the first embodiment to apply the parts Q ₁ ,
Select a pair of points to find the center of Q ₂ and Q ₃ ,
Each part Q in the midpoint detection unit 7 and the central coordinate detection unit 8
The centers C ₁₁ , C ₁₂ and C ₁₃ of ₁ , Q ₂ and Q ₃ are obtained. The shape recognition unit 11 identifies whether or not the recognition target Ob is registered in advance by collating the obtained positional relationship between the centers C ₁₁ , C ₁₂ , and C _{13 with} the positional relationship registered in advance. To do. Further, in the case of a registered object, the orientation of the recognition target Ob can be known with the position of the registered object as the reference position. As this type of technique, a well-known pattern matching technique can be applied. In FIG. 17, the parts denoted by the same reference numerals as those shown in FIG. 1 have the same function.

(Embodiment 7) In this embodiment, the techniques of Embodiments 1 to 6 are applied to align a stereo image and improve the accuracy of three-dimensional measurement based on the stereo image. It is what That is, although various techniques are known for three-dimensional measurement, parallax of both TV cameras 1a and 1b can be obtained by imaging the same spatial region using two TV cameras 1a and 1b as shown in FIG. There is a technique for obtaining three-dimensional information of a spatial region by using.

A three-dimensional measurement using this type of stereo image will be briefly described. To obtain a stereo image, as shown in FIG. 20A, one object N is imaged by two TV cameras 1a and 1b, and the obtained gray-scale image is subjected to parallel stereo conversion. Here, the parallel stereo conversion means that the two TV cameras 1a and 1b are arranged such that the optical axis and the horizontal direction in the image are parallel to each other and the focal lengths are matched with each other. 1
This is to transform each image so that the vertical coordinates of the specific points located at the same position in the spatial region in the image obtained by b are matched. The image subjected to the parallel stereo conversion is as shown in FIG. 20 (b). Therefore, when the horizontal positions x1 and x2 of the specific points in the image are obtained, the lenses used in the TV cameras 1a and 1b are calculated. Using the focal length F and the distance B between the centers of the lenses of both TV cameras 1a and 1b, the distance Z to the object N can be calculated by the following formula by the principle of the triangulation method. Z = B · F / (x1−x2) Then, as a specific point of the object N, each TV camera 1a, 1
Center C of recognition targets Oba and Obb in the image obtained in b
If a and Cb are used, the techniques described in the first to sixth embodiments can be applied to obtain the centers Ca and Cb, and the centers Ca and Cb can be obtained accurately. , As a result center Ca, Cb
The three-dimensional position (distance from the TV cameras 1a and 1b to the object N) can be accurately obtained.

That is, the images picked up by the TV cameras 1a and 1b are images having the A / D converter 2 described in the first to sixth embodiments at the input stage and the central coordinate detection unit 8 at the output stage. The positions of the centers Ca and Cb of the recognition targets Oba and Obb can be obtained by processing by the processing devices Aa and Ab. The processing after obtaining the positions of the centers Ca and Cb is conventionally known, and the three-dimensional position measuring unit 15 obtains the distance Z to the object N by applying the above equation.

(Embodiment 8) In Embodiment 7, FIG.
When there are a plurality of objects N ₁ and N ₂ as shown in 2 (a), a plurality of recognition objects Oba ₁ , Oba ₂ and Obb ₁ are included in the images obtained by the TV cameras 1a and 1b. , O
bb ₂ will be present. In such a case, since a plurality of center positions are obtained for each image, it is necessary to identify which of the objects N ₁ and N ₂ is the center position obtained. Therefore, in the present embodiment, the same processing as that of the third embodiment is performed to recognize the recognition targets Oba ₁ , Oba ₂ , O.
By recognizing the shapes of bb ₁ and Obb ₂ and determining the center for each shape, it is possible to recognize a plurality of objects N ₁ and N ₂ separately.

In the example shown in FIG. 22, recognition objects Oba ₁ , O
Since there are an ellipse and a rectangle in ba ₂ , Obb ₁ , and Obb ₂ , each recognition target Oba ₁ ,
The shapes of Oba ₂ , Obb ₁ , and Obb ₂ can be recognized. Therefore, as shown in FIG. 21, the image processing apparatus A
Each of the recognition objects Oba ₁ , Oba ₂ , Ob obtained by a and Ab
The center position of each of b ₁ and Obb _{2 is} assigned to the association processing unit 16
Object to be recognized in Oba ₁ , Oba ₂ , Obb ₁ , Ob
Correspond to each shape of b ₂ . That is, since the images obtained by both TV cameras 1a and 1b are aligned in the vertical direction by horizontal stereo conversion, the association processing unit 16 performs the same alignment on one scanning line Ls passing through both images. Shape recognition objects Oba ₁ , Oba ₂ , Obb ₁ , Obb ₂
The object N ₁ and N ₂ can be individually recognized by this correspondence, and the distances to the objects N ₁ and N ₂ can be individually calculated by the three-dimensional position measuring unit 15. Can be done. Other configurations and operations are similar to those of the seventh embodiment.

(Ninth Embodiment) In the eighth embodiment, an example in which a plurality of objects exist is shown. In the present embodiment, an object N having a relatively complicated shape as shown in FIG. An example is shown in which not only the three-dimensional position but also the direction of the object N is detected. Now, it is assumed that the object N has a hexagonal shape and three holes M ₁ , M ₂ , and M _{3 that} are circularly opened are formed. When the object N is imaged by the two TV cameras 1a and 1b, parallel stereo conversion is performed, and an edge is obtained by differential processing, FIG.
An image as shown in (b) is obtained. Image processing device A
In a and Ab, holes M ₁ and O b are provided for recognition objects Oba and Obb in the images obtained by the TV cameras 1 a and 1 b, respectively.
Portions Qa ₁ , Qa ₂ , Qa ₃ , Q corresponding to M ₂ , M ₃
b _1, we obtain the center position of Qb _2, Qb _3. The same processing as that of the sixth embodiment can be applied to this processing.

As described above, each part Qa ₁ , Qa ₂ ,
When the center positions of Qa ₃ , Qb ₁ , Qb ₂ and Qb ₃ are obtained, as shown in FIG. 23, by using the method of the seventh embodiment in the three-dimensional position measuring unit 15, the holes M ₁ ,
The three-dimensional position of M ₂ and M ₃ can be determined. In this way, the center ₃ of the holes M ₁ , M ₂ and M 3
When the three-dimensional position is determined, the three-dimensional posture measuring unit 17 detects the three-dimensional orientation of the object N using the positions of these centers. Other configurations and operations are the same as the seventh embodiment.
Is the same as.

[0063]

According to the first aspect of the present invention, the differentiation of each pixel of a grayscale image obtained by capturing an image of a polygonal object in which the shape of the focused boundary line is point-symmetrical or at least two sets of facing sides are parallel to each other. The point where the value changes is detected as an edge to be recognized, the edge direction that can define the direction of the edge with respect to each point on the edge is obtained, and two points having the same edge direction on the edge are used as a point pair and the point pair A histogram for the coordinate value of the middle point is created, and the coordinate value at which the maximum frequency is obtained in the histogram is adopted as the coordinate of the representative point of the recognition target. Differentiating the grayscale image and detecting the change point of the differential value as the edge of the recognition target instead of converting There is an advantage that it is possible to extract the di. In addition, since a histogram for the coordinate value of the midpoint between pairs of points with the same edge direction on the edge is created and the coordinate value that gives the maximum frequency is adopted as the coordinate of the representative point to be recognized, it is based on a large number of points. As a result, the representative point of the recognition target is statistically obtained, and the position of the representative point of the recognition target can be accurately obtained.

According to the second aspect of the present invention, for the point pairs having the same edge direction on the edge, only the point pair in which the midpoint between the point pairs is located near the representative point to be recognized is selected to obtain the midpoint. Since a limit condition is set for each of the points, it is possible to remove a point pair that is unnecessary for obtaining a representative point of the recognition target among a large number of point pairs. This has an advantage that the representative point of the object can be accurately obtained.

According to the third aspect of the present invention, the graphical feature to be recognized is based on the feature amount of the edge in which the point sequence formed by the point pair formed by the point pair having the coordinate value at which the maximum frequency is obtained as the coordinate of the middle point is arranged. The point sequence consisting of only the point pairs used to determine the representative point of the recognition target is used as the points on the edge. By extracting such a sequence of points, it is possible to accurately recognize the graphical feature of the recognition target.

According to a fourth aspect of the present invention, the limiting condition is set such that only a point pair having a difference in concentration gradient within a predetermined range is selected from among the point pairs to obtain a midpoint. According to the fifth aspect of the present invention, the middle point is determined by selecting only the point pair having the direction of the concentration gradient with respect to one of the above point pairs as the center line and having the other point within a predetermined angle range from the center line. In the invention of claim 6, the limiting condition is set such that only the point pair in which the distance between the point pairs is within a predetermined range is selected to obtain the midpoint. However, when determining the position of the representative point of the recognition target based on the difference in the density gradient of the point pair, the positional relationship of the point pair, etc. Since the position is calculated, the processing is simple and the accuracy of the calculated representative point position is high. There is an advantage in that.

According to the seventh aspect of the present invention, since the grayscale image is a color image, and the midpoint is obtained for the recognition target existing in the area of the specific color range, the color range is limited.
For example, when it is necessary to obtain the position of the representative point individually for a plurality of objects of different colors, there is an advantage that each object can be individually identified by color. According to the invention of claim 8, the histogram is created for the remaining edges obtained by deleting a part of the edges. Therefore, if there are a plurality of objects, the edges of unnecessary objects are deleted in advance and then the histogram is created. The creation has an advantage that meaningless information is less likely to occur and the position of the representative point of the recognition target can be accurately obtained.

According to a ninth aspect of the invention, the shape of the recognition target is recognized by obtaining the distribution of the density gradient for each point of the point sequence and comparing the distribution pattern of the density gradient with a known distribution pattern. Since the density gradient distribution pattern for each point in the row is obtained, the shape of the recognition target can be recognized by pattern matching of the density gradient distribution pattern. Moreover, since the point sequence for obtaining the distribution pattern of the concentration gradient is the point sequence used when obtaining the midpoint of the recognition target, it is considered that the shape of the recognition target is well reflected. There is an advantage that the shape of the recognition target can be accurately grasped by obtaining the shape of the target.

According to the tenth aspect of the present invention, the distribution of the density gradient for each point of the point sequence is obtained, and the orientation of the recognition target is detected based on the distribution pattern of the density gradient. By obtaining the distribution of, the orientation of the recognition target can be obtained based on the distribution pattern of the concentration gradient. Moreover, since the point sequence for obtaining the distribution pattern of the density gradient is the point sequence used when obtaining the midpoint of the recognition target, it is considered that it reflects the graphical characteristics of the recognition target well. The direction of the recognition target can be accurately grasped by obtaining the direction of the recognition target from the above.

According to the eleventh aspect of the present invention, the number of the point sequences is calculated, and the size of the recognition target is detected based on the number of the point sequences. The size of the recognition target is calculated based on the number of the point sequences. be able to. Moreover, since the point sequence for obtaining the number of points is the point sequence used when obtaining the midpoint of the recognition target, it is considered that the figure characteristic of the recognition target is well reflected, and the number of such point sequences is calculated. This has the advantage that the size of the recognition target can be obtained with good reproducibility.

According to a twelfth aspect of the present invention, the first step of deleting the edge of the recognition object having the coordinate value of the representative point whose coordinate value gives the maximum frequency in the histogram, and the edge remaining after the first step By repeating the second step of creating again, the coordinates of the representative points of the plurality of recognition targets are individually obtained. When there are a plurality of recognition targets of different sizes, the recognition targets are sorted in descending order. Since the histogram is created by obtaining the positions of the points and excluding the edges of the recognition target for which the positions of the representative points have been obtained, there is an advantage that the positions of the representative points of the respective recognition targets can be obtained individually.

According to the thirteenth aspect of the present invention, the coordinates of the representative point of each recognition target are obtained for the object having a plurality of recognition targets, and the position and orientation of the object are recognized based on the positional relationship of the coordinates of the obtained representative points. Therefore, since the positional relationship of the representative points of each recognition target is obtained from the object having a plurality of recognition targets, there is an advantage that the position and orientation of the object can be recognized.

According to a fourteenth aspect of the present invention, the coordinates of the representative point of the recognition target are respectively obtained for a plurality of grayscale images obtained by imaging the object from a plurality of different positions, and the coordinates of the obtained representative point of the recognition target are calculated. After matching, the distance to the object is obtained by applying the triangulation method.
When the distance to the recognition target is calculated using the stereo image, the position of the representative point of the recognition target calculated by the method of claim 1 is used to correlate a plurality of images, so that the correlation can be performed with high accuracy. . As a result, there is an advantage that the distance to the recognition target can be accurately obtained.

According to a fifteenth aspect of the present invention, the coordinates of the representative point of the recognition target are obtained for a plurality of grayscale images obtained by imaging the object from a plurality of different positions, and the coordinate of the representative point is obtained by the point pair. By extracting the graphic feature of the recognition target based on the feature amount of the edge where the formed point sequence is aligned, and associating the regions having the same graphic feature in each image, the triangulation method is applied. Since the distance to the object is calculated, and when the distance to the recognition target is calculated using the stereo image, the graphic feature of the recognition target calculated by the method of claim 3 is used for associating a plurality of images. There is an advantage that the correspondence can be made with. As a result, the distance to the recognition target can be accurately obtained.

According to a sixteenth aspect of the present invention, for a plurality of grayscale images obtained by imaging an object having a plurality of recognition targets from a plurality of different positions, the coordinates of the representative point of each recognition target are obtained from the respective images. , After the coordinates of the representative points obtained in each image are associated with each other, the position of the coordinates of the representative points of each recognition target in the three-dimensional space is obtained by applying the triangulation method, and the positional relationship of the coordinates of the representative points is obtained. The position and the orientation of the object in the three-dimensional space are recognized based on, and when the three-dimensional information is obtained using the stereo image, the positions of the plurality of positions of the object are obtained, and the object is determined based on the positional relationship between them. The position of the representative point of the recognition target obtained by the method of claim 1 is used for associating a plurality of images with each other, and therefore the three-dimensional information of the object is highly accurate and reproducible. Get It has the advantage that the door can be.

[Brief description of drawings]

FIG. 1 is a block diagram showing a first embodiment.

FIG. 2 is a diagram illustrating the principle of the first embodiment.

FIG. 3 is a diagram illustrating the principle of the first embodiment.

FIG. 4 is a diagram showing an example of a limiting condition according to the first embodiment.

FIG. 5 is a diagram showing an example of a limiting condition according to the first embodiment.

FIG. 6 is a diagram showing an example of restriction conditions according to the first embodiment.

FIG. 7 is a diagram showing an example of a limiting condition according to the first embodiment.

FIG. 8 is an operation explanatory diagram of the first embodiment.

FIG. 9 is a block diagram showing a second embodiment.

FIG. 10 is a block diagram showing a third embodiment.

FIG. 11 is a diagram illustrating the principle of the third embodiment.

FIG. 12 is a diagram illustrating the principle of the third embodiment.

FIG. 13 is a diagram illustrating the principle of the third embodiment.

FIG. 14 is a block diagram showing a fourth embodiment.

FIG. 15 is an operation explanatory diagram of the fourth embodiment.

FIG. 16 is a block diagram showing a fifth embodiment.

FIG. 17 is a block diagram showing a sixth embodiment.

FIG. 18 is a diagram illustrating the principle of the sixth embodiment.

FIG. 19 is a block diagram showing a seventh embodiment.

FIG. 20 is a diagram illustrating the principle of the seventh embodiment.

FIG. 21 is a block diagram showing an eighth embodiment.

FIG. 22 is a diagram illustrating the principle of the eighth embodiment.

FIG. 23 is a block diagram showing Embodiment 9.

FIG. 24 is a diagram illustrating the principle of the eighth embodiment.

FIG. 25 is an operation explanatory diagram of a conventional example.

FIG. 26 is a diagram illustrating a problem of the conventional example.

FIG. 27 is a diagram illustrating a problem of the conventional example.

FIG. 28 is a diagram illustrating a problem of the conventional example.

FIG. 29 is a diagram illustrating a problem of the conventional example.

[Explanation of symbols]

1 TV camera 2 A / D converter 3 frame memory 4 Differentiation processing unit 5 point pair extraction unit 6 Condition verification section 7 Midpoint detector 8 Center coordinate detector

─────────────────────────────────────────────────── ─── Continuation of the front page (56) Reference JP-A-5-107030 (JP, A) (58) Fields investigated (Int.Cl. ⁷ , DB name) G01B 11/00-11/30 G06T 1 / 00 G06T 7/00 G06T 7/60 G06T 9/20

Claims (16)

(57) [Claims] 1. A recognition target is a change point of a differential value for each pixel of a grayscale image obtained by capturing an image of a polygonal object in which the shape of a boundary line of interest is point symmetry or at least two pairs of opposite sides are parallel to each other. The edge direction that can define the direction of the edge for each point on the edge, and regards the coordinate value of the midpoint between the point pairs as two point pairs having the same edge direction on the edge. An object recognition method characterized in that a histogram is created, and the coordinate value at which the maximum frequency is obtained in the histogram is adopted as the coordinate of the representative point of the recognition target. 2. Regarding the point pairs having the same edge direction on the edge, a limiting condition is set so that only the point pair in which the midpoint between the point pairs is located near the representative point of the recognition object is selected to obtain the midpoint. The object recognition method according to claim 1, wherein the object recognition method is performed. 3. A graphic feature to be recognized is extracted based on a feature amount of an edge in which a sequence of points formed by a pair of points having a coordinate value at which the maximum frequency is obtained in a histogram as a coordinate of a middle point is arranged. The object recognition method according to claim 1 or 2, which is characterized. 4. The object according to claim 2, wherein the limiting condition is set so as to obtain a midpoint by selecting only a point pair having a difference in concentration gradient within a predetermined range from the point pair. Recognition method. 5. A midpoint is obtained by selecting only a point pair having a direction of a density gradient with respect to one of the points of the point pair as a center line and having the other point within a predetermined angle range from the center line. 3. The object recognition method according to claim 2, wherein the limiting condition is set to. 6. The object recognition method according to claim 2, wherein the limiting condition is set so that only the point pair in which the distance between the point pairs is within a predetermined range is selected to obtain the midpoint. 7. The object recognition method according to claim 1, wherein the grayscale image is a color image, and a midpoint is obtained for a recognition target existing in a region of a specific color range. 8. The object recognition method according to claim 1, wherein the histogram is created for the remaining edges that are obtained by deleting a part of the edges. 9. The shape of a recognition target is recognized by obtaining a distribution of a density gradient for each point of the point sequence and comparing the distribution pattern of the density gradient with a known distribution pattern. Object recognition method. 10. The object recognition method according to claim 3, wherein the distribution of the density gradient is calculated for each point of the point sequence, and the orientation of the recognition target is detected based on the distribution pattern of the density gradient. 11. The object recognition method according to claim 3, wherein the number of the point sequences is obtained, and the size of the recognition target is detected based on the number of the point sequences. 12. A first step of deleting an edge to be recognized having a coordinate value of a representative point having a coordinate value for obtaining the maximum frequency in the histogram, and a second step of recreating the histogram with respect to the edges remaining after the first step. The object recognition method according to claim 1, wherein the coordinates of a plurality of representative points of the recognition target are individually obtained by repeating the process. 13. An object having a plurality of recognition targets, the coordinates of a representative point of each recognition target are obtained, and the position and orientation of the object are recognized based on the positional relationship of the coordinates of the obtained representative points. The object recognition method according to claim 1. 14. With respect to a plurality of grayscale images obtained by imaging an object from a plurality of different positions, the coordinates of the representative point of the recognition target are obtained, and the coordinates of the obtained representative point of the recognition target are associated with each other, The object recognition method according to claim 1, wherein the distance to the object is obtained by applying a triangulation method. 15. A sequence of points formed by a pair of points for which the coordinates of a representative point of a recognition target are obtained for a plurality of grayscale images obtained by imaging an object from a plurality of different positions, and the coordinates of the representative point are obtained. After extracting the graphic features of the recognition target based on the feature amount of the edges that are lined up and associating the regions having the same graphic features in each image, the distance to the object is determined by applying the triangulation method. The object recognition method according to claim 3, wherein the object recognition method is performed. 16. Regarding a plurality of grayscale images obtained by imaging an object having a plurality of recognition targets from a plurality of different positions, the coordinates of the representative point of each recognition target are obtained from each image, and within each image. After associating the obtained coordinates of the representative points with each other, the triangulation method is applied to obtain the positions of the coordinates of the representative points of the respective recognition targets in the three-dimensional space, and three-dimensionally based on the positional relationship of the coordinates of the representative points. The object recognition method according to claim 1, wherein the position and orientation of the object in the space are recognized.