JP3437141B2 - Pattern recognition method and apparatus - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a pattern recognition method and apparatus for identifying an object in an image and detecting its position and orientation.

[0002]

2. Description of the Related Art In automated work, especially in robotic work, pattern recognition is essential to know the position and orientation of an object by using an image of the object. Conventionally, as a method of recognizing a pattern in an image, a template matching method in which a template representing a pattern to be detected is applied to the image to calculate a cross-correlation value, and the position where this value is the largest is the position of the object is famous. is there. There is also a template matching method in which the sum of the absolute values of the differences between the two is calculated and the position is detected based on the smallest position.

However, in the template matching method, when there is a possibility that the orientation of the pattern differs from that of the object, the orientation of the pattern is gradually changed until the pattern of the object and the pattern match, and the matching is performed from the edge of the image. Since it is repeated, the amount of calculation becomes huge and there is a drawback that it takes time. Further, when the measurement target is a metal product, the gray value of the target in the image greatly varies depending on the condition of the light beam, and thus the position of the target cannot be detected when the difference between the contour of the template is large. There are cases.

There is also a vector correlation method for detecting the position of an object from the result of shifting and adding images using a vector connecting the edge position of the template and the point to be detected, but when the posture of the object changes Not applicable to.

[0005]

Therefore, the problem to be solved by the present invention is a new pattern recognition method capable of quickly calculating the position and orientation of a pattern, in which the illumination or the reflection state of the surface of an object is changed. Is to provide a pattern recognition method that is not easily affected.

[0006]

In order to solve the above-mentioned problems, the pattern recognition method of the present invention detects a plurality of contour points on the contour of a template image and template normal vectors at the contour points, A shift vector V having a length r of a line segment connecting a point and a predetermined reference point and an angle θ between the line segment and a normal vector is created and stored in advance. Then, the image of the object is acquired, the contour points on the contour of the object pattern in the image are detected, the pattern normal vector at the contour point of the object pattern is calculated, and the attribute is set to the pattern normal vector. Shift vector conversion is performed to calculate the coordinates of the mapping point shifted by the length r of the shift vector in the direction of the angle θ, data is stored in the coordinate correspondence memory corresponding to the calculated mapping point coordinates, and after the shift vector conversion It is characterized in that an address where the number of data is concentrated is detected and the position of the object is obtained based on the coordinates corresponding to the memory address.

In the pattern recognition method of the present invention, the comparison with the template is performed by the shift vector conversion mainly for the simple iterative operation, and thus the high speed operation is possible. Further, since the determination is performed by a statistical means, it is possible to accurately identify the pattern even when the quality of the acquired image is poor, and at the same time, it is possible to determine the position of the object.

Further, by using the angle difference between the pattern normal vector and the template normal vector as the data to be stored in the coordinate correspondence memory, the frequency distribution of the angle difference values stored in the memory address corresponding to the coordinates of the object is obtained. Based on the above, it is possible to obtain the posture of the target object based on, for example, a peak value or an average value. This is because the angle difference has substantially the same value for each mapping point in the pattern that matches the template, and means the angle between the orientations of the orientation pattern of the template.

Note that the pattern method is performed only when the position of the object is obtained based on the memory address where the number of data is concentrated, and then the shift vector conversion is performed again to have the corresponding mapping point near the obtained position. The angle difference between the line vector and the template normal vector may be stored and accumulated, and the posture of the object may be obtained based on the frequency distribution of the accumulated data values. In the method of obtaining the position and orientation all at once, the angle difference is stored in the memory address corresponding to the coordinates of the mapping point, so that the memory structure needs to be three-dimensional and the memory capacity must be large. In the case of performing shift vector conversion by dividing into two as described above, it is sufficient to increment the numerical value stored in the address corresponding to the mapping point coordinate in order to record the mapping point position, and accumulate the angular difference. Sometimes it is sufficient to store only the angle difference data when the mapping point falls within a limited range, so the capacity of the memory for storing the data may be small.

Further, after the position of the object is roughly obtained by the first shift vector conversion, further shift vector conversion is performed, and only when there is a mapping point corresponding to the obtained position vicinity, further finely divided coordinate correspondence is performed. It is also possible to accumulate appearance data of mapping points in a memory and obtain a more precise position based on the degree of concentration of the number of data. By combining coarse measurement and precise measurement in this way, it is possible to make the size of the grid corresponding to the coordinates large in the coarse measurement and reduce the memory capacity of the coordinate correspondence memory. Since only the data corresponding to the narrow area in the vicinity of is stored, the data storage memory can be measured with sufficient accuracy even if it is extremely small.

Instead of calculating the contour point from the contour of the grayscale image and calculating the normal vector at the contour point as described above, the distance to the target is recorded at the position corresponding to the pixel of the two-dimensional image memory. It is also possible to obtain the position and orientation of the three-dimensional object by obtaining the normal vector of the object surface from the distance gradient at each point using the range image and performing the shift vector conversion.

Further, in order to solve the above-mentioned problems, a pattern recognition apparatus of the present invention includes an image processing apparatus, an information processing apparatus,
The image processing apparatus includes a first storage device and a second storage device, captures an image of a template, detects a contour point of the template and a normal vector at the contour point, and determines a contour point and a predetermined reference point. A shift vector V having attributes of the length r of the connecting line segment and the angle θ between the line segment and the normal vector
To store the shift vector in advance in the first storage device, the image processing device inputs the image information of the target object at the time of measurement, and the target object in the image is processed in the same manner as the template. The contour point is detected for the object pattern, the pattern normal vector at the contour point is calculated, and the information processing device inputs the pattern normal vector and the shift vector to perform shift vector conversion.

The shift vector conversion is to calculate the coordinates of a mapping point which is deviated by the line segment length r in the direction having the shift vector attribute angle θ with respect to the pattern normal vector. The second storage device is provided with a coordinate correspondence memory having storage cells corresponding to the coordinates, and counts each time the mapping point has the coordinates at the coordinate correspondence memory address corresponding to the coordinates of the mapping point calculated by the information processing device. The frequency is stored by a method such as stepping, and the information processing device reads the recording result after the shift vector conversion to check the coordinates corresponding to the memory with a large frequency and determine the position of the object based on the result. Characterize.

It should be noted that, initially, a coarsely divided coordinate-corresponding memory is used to perform shift vector conversion to roughly determine the position of the object based on the memory in which the number of data is concentrated, and this time, it is further divided. The second coordinate correspondence memory is prepared, the shift vector conversion is performed again, and the appearance data of the mapping point is stored in the second coordinate correspondence memory only when the mapping point comes near the roughly determined position. A more precise position may be obtained based on the degree of concentration of accumulated data.

Further, instead of storing the occurrence frequency in the coordinate correspondence memory, the data of the angular difference between the pattern normal vector and the template normal vector is accumulated, and the data is concentrated from the coordinates of the memory. It is also possible to obtain the position of the object and obtain the posture of the object based on the frequency distribution of the angle difference data values stored in the memory.

[0016]

BEST MODE FOR CARRYING OUT THE INVENTION The present invention will be described below in detail based on embodiments with reference to the drawings.

[0017]

EXAMPLE This example shows a case where the position and orientation of a pattern in an image are detected by using the pattern recognition method of the present invention. FIG. 1 is a block diagram of a pattern recognition apparatus of this embodiment, FIG. 2 is a flow chart showing a pattern recognition method of this embodiment, FIG. 3 is a flow chart of a step of preparing a shift vector as preparation, and FIG. 4 is a shift vector generation. FIG. 5 is an explanatory view showing a method, FIG. 5 is a flowchart of shift vector conversion used in this embodiment, FIGS. 6 to 9 are drawings for sequentially explaining steps of shift vector conversion, and FIG.
Is a drawing for explaining the conditions for determining the position of the pattern, FIG.
1 is a conceptual diagram of a robot apparatus to which this embodiment is applied.

In the present embodiment, a pattern recognition method that can be easily calculated by a computer, can judge a target product when the degree of approximation is high in comparison with the shape of the template, and can detect the position and orientation of the target product. Then, as shown in Figure 1,
The pattern recognition device is provided with the image processing device 1, the information processing device 2, the first storage device 3, and the second storage device 4. The pattern recognition procedure of this embodiment is as shown in the flow chart of FIG. 2. First, an image pickup device (not shown) places a template showing the object to be sorted or its shape in a standard posture, shoots it, and generates an image. The image processing apparatus 1 inputs this image information, calculates a shift vector from it, and stores it (S1).

FIG. 3 is a flow chart for explaining a method of creating a shift vector. The standard template TM is photographed by the image pickup device having the same configuration as that used for the measurement of the target product, and the image information as shown in FIG. 4 is obtained (S11). The standard template TM has the shape of the object to be identified,
It was placed in a standard posture. At this time, the information processing apparatus 2 cuts out the area where the template TM exists as a teaching area and designates the reference point O at an appropriate position (S12). The reference point O may be specified inside or outside the template TM.

The image processing apparatus 1 scans the teaching area of the grayscale image on which the template TM is photographed with a contour extraction filter such as a Mexican hat filter, and extracts the contour portion having a density gradient as a zero cross point Pi ( S13). At this time, the direction of the density gradient is also calculated, and this is the direction of the normal to the contour line (S1).
4). In this way, by scanning an appropriate filter in the image, it is possible to mechanically calculate a point on the contour line and a normal vector set up there.

A unit vector on the normal line set at the zero-cross point Pi is called a teaching vector Ti (S15). The information processing device 2 inputs the information of the teaching vector Ti, generates a line segment from the zero cross point Pi to the reference point O, and calculates the length thereof. A vector having the length ri of the line segment and the angle θi formed by the line segment with the teaching vector Ti as an attribute (ri, θi) is called a shift vector Vi (S16). The calculated shift vector Vi (ri, θi) is stored in the first storage device 3
(S17).

For the template TM in the image, the same procedure is repeated until the zero-cross points Pi can be extracted over the entire circumference of the contour at appropriate intervals (S18). The number of contour points Pi to be used on the outer periphery of the template TM may be determined according to the complexity of the shape of the template, but since this method uses a statistical method for determination, a large number of points are used. No need. In this way, the shift vector V1, from the appropriate number of start points Pi determined at appropriate intervals on the outer periphery of the template TM toward the reference point O1,
V2, ..., Vn are obtained.

After the shift vector Vi of the identification object product is prepared in the first storage device 3, the actual object product can be processed. An image of a target article is photographed by an image pickup device (not shown), converted into a grayscale image, and captured in the image processing device 1 (S2). The target article is, for example, a product that is conveyed by a belt conveyor or a semi-finished product that is on a conveyance table and is waiting for processing by a robot. The conditions for shooting the target article are adjusted so as to be the same as those on the template TM so that the images are at the same scale.

The image processing apparatus 1 uses the grayscale image F of the target article.
Then, the contour line is detected to obtain the normal vector Nj at the detection point Qj (S3). The contour can be detected by the same method as when generating the shift vector, and at the same time, the direction of the density gradient at the detected contour point Qj, that is, the normal vector Nj can be calculated. This normal vector is also called a contour line gradient vector because it represents the slope of the contour line.

By scanning the target area in the image with the contour extraction filter, the contour points Qj are detected at appropriate intervals over the entire circumference of the image F of the target article, and the normal vector Nj at that point is obtained to process the information. It is stored in the second storage device 4 via the device 2. After obtaining the contour points Qj over the entire circumference of the target article pattern, the information processing apparatus 2 performs shift vector conversion (S5).

The shift vector conversion is a procedure for obtaining the mapping point Mji by applying the shift vector Vi to each point Qj as shown in the flowchart of FIG. FIG. 6 is an explanatory diagram showing how the mapping point Mj1 is obtained by applying the first shift vector V1 to the image including the first pattern F1 and the second pattern F2.

As shown in FIG. 6, each contour point Q1, Q2,
The shift vector V1 obtained around the template TM is drawn starting from Qm (S21). At this time, the direction of the shift vector V1 is the normal vector N1,
An attribute angle θ1 which is the attribute of each of N2, ..., Nm is formed. Then, each point Q1, Q2, ...
The mapping point Mj1 moved by the shift vector V1 is obtained for each Qm. At this time, each point Q1, Q2, ...
.., Nm normal vectors N1, N2, ..., N
An angle difference Φj1 between m and the teaching vector T1 of the shift vector V1 is calculated (S22).

An area having an address corresponding to the coordinate, that is, a coordinate corresponding memory is prepared in the second storage device 4, and the address has a three-dimensional structure including a plurality of memory elements capable of storing a plurality of data. ing. Mapping point Mj1
And the angle difference Φj1 is obtained, the angle difference Φj1 is recorded in the address corresponding to the coordinates of the mapping point Mj1 (S2).
3, S6).

As shown in FIG. 7, the second shift vector V
The same operation is performed for 2. Further, FIG. 8 is a conceptual diagram when processing is performed for the third shift vector V3,
FIG. 9 is a drawing showing the processing situation for the final shift vector Vn. In this way, the above operation is performed for all the shift vectors V1, V2, ..., Vn (S24, S7).

As described above, all shift vectors Vi are applied to all the extracted contour points Qj to complete the shift vector conversion, and the mapping point Mji and the angle difference Φji.
Is recorded in the coordinate correspondence memory of the second storage device 4. Figure 1
0 is a drawing in which the result of the shift vector conversion is superimposed on all the shift vectors. As shown in FIG. 10, the first pattern F of the patterns to be measured is
Since 1 has the same shape as the template TM, the mapping reference point Ot in which the mapping points Mji overlap by the number n of the shift vector Vi is created. This mapping reference point Ot is the template T
A position corresponding to the reference point O of M, which is the mapping reference point Ot
The position of the pattern F1 can be represented by the coordinates of.
It should be noted that the mapping points for the second pattern F2 different from the template TM are dispersed, and no concentrated mapping reference points appear.

As described above, when the same pattern as the template TM exists, an address where the number of data is concentrated appears in the coordinate corresponding memory after the shift vector conversion is completed. The coordinates corresponding to this address become the position of the pattern (S8). However, the contour points Qj are extracted on the pattern contour at appropriate intervals, and are not always at positions corresponding to the contour points Pi on the template. For this reason, the coordinates where the mapping points are gathered have a statistical spread. Therefore, it is preferable to set the cell of the coordinate correspondence memory to be slightly larger, or to collect some adjacent addresses to perform the position determination. The information processing device 2 reads out the associated memory area, performs information processing, and determines the position of the pattern.

The memory corresponding to the mapping reference point Ot also contains, as noise, the angle difference Φji data calculated for mapping points that do not match the template.
However, since the angle difference Φji has substantially the same value over the entire circumference of the contour for the mapping point Mji that matches the template, the frequency difference of the stored angle difference Φji is taken to obtain the angle difference for the target pattern. Φji can be extracted. The pattern F1 targeted by this angle difference
Represents the posture (S9). In this way, the photographed pattern is collated with the template, the target pattern is recognized, and its position and orientation can be obtained and output (S).
10).

According to the pattern recognition method of the present embodiment, the contour point and the inclination of the contour line at that point are obtained by simple image processing, and the shift vector conversion is mechanically performed by a very simple vector operation, and the memory is stored. Since the position information is obtained by the majority decision of the number of data stored in and the posture is obtained from the frequency distribution of the data contents, the pattern recognition can be performed at a very high speed by a simple processing device. Further, since the arithmetic processing is statistically performed, even if the illumination or the reflection state of the surface of the object is slightly changed, it is not affected, and the position and orientation can be reliably identified and the position and orientation can be detected.

The number or intervals of contour points used on the outer periphery of the template and the pattern may be determined according to the complexity of the shape of the target article or pattern, but since a statistical method is used, It is not necessary to use that many points. Further, it is convenient in terms of arithmetic processing to use a number such as 16 or 32 according to the number of arithmetic bits of the computer central processing unit.

In the above embodiment, the number of mapping points and the angle difference data are stored at the same time. However, only the mapping points are calculated by the first shift vector conversion and stored in the coordinate correspondence memory to store the mapping points. Pattern matching is performed based on the degree of concentration, then shift vector conversion is performed again, this time specifying an appropriate area surrounding the mapping reference point where the mapping points are concentrated, and storing the angle difference data only in this area. The posture may be detected from the frequency distribution. According to such a two-pass method, it is not necessary to prepare a memory having a three-dimensional structure for storing the angle difference data at every coordinate corresponding address, and the first mapping point data is stored in the coordinate corresponding memory as the number of data. Since the subsequent angle difference data may be selected and stored for the limited addresses, the memory capacity can be saved.

In the above embodiment, the shift vector is sequentially applied to the contour points on the entire circumference of the pattern in the shift vector conversion. However, the order is reversed and the contour points of the pattern are shifted one by one. You may make it operate all the vectors. Needless to say, the results obtained are the same, and if the degree of data integration is being monitored, a sufficient amount of data can be accumulated in the same coordinate correspondence memory to know the coordinates of the target even during shift vector conversion. You may be able to In such a case, when it is found that the accuracy is sufficiently high, it is possible to switch to precision measurement with a limited address so as to obtain precise position information.

As another mode of this embodiment, there is a three-dimensional position / orientation detection method using a range image. The distance image is a two-dimensional image in which the distance information to the object surface is stored at each point corresponding to the pixel in the two-dimensional memory space. The range image is obtained by, for example, scanning an object with a laser beam and imaging the reflected light on the CCD with a lens.
It can be generated by a method of converting the distance into a distance and storing it in an image memory according to the principle of triangulation using the light spot position information of the CD. In such a range image, the normal vector of the object surface can be obtained by obtaining the distance gradient near the target point.

Regarding the reference solid corresponding to the template,
Generate normal vectors on the surface at appropriate intervals. This normal vector is a three-dimensional vector called a teach vector. Next, a shift vector from the foot of the teach vector to the reference point is generated, and the length between the shift vector and the angle between the teach vector and the direction of the teach vector are recorded as attributes.

After the preparation for measurement is completed in this way,
A range image is acquired for the target object. The distance image information is fetched by the image processing device, and the normal vector of the object surface is calculated from the distance gradient at appropriate intervals. Shift vector conversion is performed for each normal vector. In this aspect, since the shift vector is a three-dimensional vector, the mapping points are distributed in the three-dimensional space. Therefore, the coordinate corresponding memory for storing the coordinates of the mapping point is selected to have a three-dimensional structure. The coordinate correspondence memory stores the solid angle between the teaching vector and the normal vector of the target object.

If there is a coordinate where the mapping points are concentrated, the figure matches the template. Also, when the distribution of the solid angles stored in the memory corresponding to the coordinates where the mapping points are concentrated is investigated and the angle with the highest frequency is found, this represents the pose of the target object. When a portion different from the template is photographed due to the posture of the three-dimensional figure, the concentration of mapping points is smaller than the number of shift vectors, and therefore the identity may be determined by an appropriate number of concentrations.

The pattern recognition method of the present invention can be applied to a robot apparatus as shown in FIG. The semi-finished product 12 transported by the transport conveyor 11 is photographed by the imaging device 13 and the image information is sent to the image processing device 14. The image processing device 14 detects the contour of the photographed pattern by filtering the grayscale image, and calculates the normal vector on the contour. The calculation result is stored in the storage device 16 via the central processing unit 15.

The central processing unit 15 executes shift vector conversion using the teaching vector and the shift vector regarding the template stored in the storage unit 16 in advance,
The obtained angle difference data is sequentially stored in the coordinate correspondence memory in the storage device 16. When the shift vector conversion is completed, the address where the data in the coordinate correspondence memory is concentrated is found and the frequency distribution of the angular difference stored at that address is examined. The detected address is the position of the pattern,
When the obtained information is sent to the robot controller 17 with the most frequently used angle difference value as the posture of the pattern, the robot controller 17 operates the robot hand 18 to produce the semi-finished product 12
And then adjust the posture relationship accurately and transport it to the target position.

[0043]

As described above, according to the pattern recognition method of the present invention, a required object is accurately and quickly identified from a photographed product image, and its position and orientation are calculated, and Information can be transmitted to a robot control device or the like to automatically operate a robot hand or the like to grasp an item.

[Brief description of drawings]

FIG. 1 is a block diagram of an embodiment of a pattern recognition device of the present invention.

FIG. 2 is a flowchart illustrating an embodiment of the pattern recognition method of the present invention.

FIG. 3 is a flowchart of a process of creating a shift vector in this embodiment.

FIG. 4 is a diagram illustrating a method of generating a shift vector according to the present exemplary embodiment.

FIG. 5 is a flowchart of shift vector conversion used in this embodiment.

FIG. 6 is a diagram illustrating a first step of shift vector conversion according to the present exemplary embodiment.

FIG. 7 is a diagram illustrating a second step of shift vector conversion according to the present exemplary embodiment.

FIG. 8 is a diagram illustrating a third step of shift vector conversion according to the present exemplary embodiment.

FIG. 9 is a diagram illustrating a fourth step of shift vector conversion according to the present exemplary embodiment.

FIG. 10 is a diagram illustrating conditions for determining the position of a pattern in this embodiment.

FIG. 11 is a conceptual diagram of a robot apparatus to which this embodiment is applied.

[Explanation of symbols]

14 Image processing device 15 Central processing unit 16 storage 17 Robot controller

─────────────────────────────────────────────────── ─── Continued Front Page (58) Fields surveyed (Int.Cl. ⁷ , DB name) G06T ⁷ /00-7/60 G06T 1/00 JISST file (JOIS)

Claims (11)

(57) [Claims] 1. A normal vector at a plurality of points on an object surface is detected from an image of a template, a length r of a line segment connecting the surface point and a predetermined reference point, and an angle between the line segment and the normal vector. A shift vector V having θ as an attribute is created and stored in advance, an image of the object is acquired, a pattern normal vector at a surface point of the object pattern in the image is calculated, and the pattern normal vector is calculated. The attribute angle θ
Shift vector conversion for calculating the coordinates of the mapping point deviated by the line segment length r in the direction having, and storing the data in the coordinate correspondence memory corresponding to the mapping point coordinates, the number of data after the shift vector conversion. A pattern recognition method, characterized in that the position of an object is obtained based on the coordinates corresponding to the memory in which is concentrated. 2. The image of the template is a distance image, and the normal vector of the object surface is a normal vector calculated for a point on the object surface in the distance image. Pattern recognition method. 3. A ring detected from a grayscale image of a template
By defining a plurality of contour points on the contour line,
A normal vector set on the contour point is calculated, and a shift vector V having the length r of a line segment connecting the contour point and a predetermined reference point and the angle θ between the line segment and the normal vector as attributes is created. Then, the grayscale image of the object is acquired, the pattern normal vector at the contour point of the object pattern in the image is calculated, and the pattern normal vector and the attribute angle θ
Shift vector conversion for calculating the coordinates of the mapping point deviated by the line segment length r in the direction having, and storing the data in the coordinate correspondence memory corresponding to the mapping point coordinates, the number of data after the shift vector conversion. A pattern recognition method, characterized in that the position of an object is obtained based on the coordinates corresponding to the memory in which is concentrated. 4. A plurality of contour points on the contour and normal vectors at the contour points are detected from an image of a template, and a length r of a line segment connecting the contour point and a predetermined reference point and the line segment are included. A shift vector V having the angle θ of the normal vector as an attribute is created and stored in advance, an image of the target object is acquired, a contour point on the contour of the target object pattern in the image is detected, and the target object is detected. A pattern normal vector at the contour point of the object pattern is calculated, shift vector conversion is performed to calculate the coordinates of the mapping point deviated by the line segment length r in the direction having the pattern normal vector and the attribute angle θ, A pattern characterized by accumulating data in a coordinate correspondence memory corresponding to the mapping point coordinates and obtaining the position of the object based on the coordinates corresponding to the memory in which the number of data is concentrated after the shift vector conversion.識方 method. 5. The pattern when the position of an object is determined based on a memory in which the number of data is concentrated and further shift vector conversion is performed to have a mapping point corresponding to the vicinity of the determined position. 5. The angle difference between a normal vector and the normal vector of the template is stored and accumulated, and the posture of the object is obtained based on the frequency distribution of the accumulated data values. The pattern recognition method according to any one of 1. 6. When the position of the object is determined based on the memory in which the number of data is concentrated, further shift vector conversion is performed, and when there is a mapping point corresponding to the vicinity of the determined position, it is more detailed. 6. The appearance data of the mapping point is stored in the divided coordinate correspondence memory, and a more precise position is obtained based on the degree of concentration of the stored data number. Pattern recognition method. 7. The object to be stored in the memory is an angular difference between the pattern normal vector and the template normal vector, and the object is based on coordinates corresponding to the memory in which the number of data is concentrated. 5. The position of the object is obtained, and the posture of the object is obtained based on the frequency distribution of the data values accumulated in the memory in which the number of data is concentrated. Pattern recognition method. 8. An image processing device, an information processing device, a first storage device, and a second storage device, wherein the image processing device extracts a plurality of contour points on the contour from a template image and normal vectors at the contour points. And a shift vector V having the length r of a line segment connecting the contour point and a predetermined reference point and the angle θ between the line segment and the normal vector as attributes, and the first storage device stores the shift vector V. The shift vector is stored in advance, the image processing apparatus inputs the image information of the target object, detects a contour point on the contour of the target object pattern in the image, and detects the contour point of the target object pattern. A pattern normal vector is calculated, and the information processing device inputs the pattern normal vector from the image processing device and the shift vector from the first storage device to obtain the shift vector of the shift vector with respect to the normal vector. Shift vector conversion for calculating the coordinates of the mapping point deviated by the line segment length r in the direction having the sex angle θ, and the second storage device is provided with a coordinate correspondence memory having storage cells corresponding to the coordinates. The coordinate corresponding memory corresponding to the calculated mapping point coordinates stores the frequency at which the mapping point has the coordinates, and the information processing device, based on the coordinates corresponding to the memory having the large frequency after the shift vector conversion, the object. A pattern recognition device characterized by determining the position of the pattern recognition device. 9. The image processed by the image processing device is a range image, the contour points are points on the surface of an object, and the normal vector is obtained by a range gradient derived from the range image. 9. The pattern recognition device according to claim 8, which is characterized in that. 10. When the position of the object is determined based on the memory in which the number of data is concentrated, further shift vector conversion is performed, and when the mapping point corresponding to the vicinity of the determined position is provided, it is further detailed. 10. The pattern according to claim 8 or 9, wherein the appearance data of the mapping point is accumulated in the divided second coordinate correspondence memory, and a more precise position is obtained based on the concentration of the accumulated data number. Recognition device. 11. A memory in which the data of the angular difference between the pattern normal vector and the template normal vector is accumulated instead of storing the frequency in the coordinate correspondence memory, and the data number is concentrated. 11. The pattern recognition apparatus according to claim 8, wherein the posture of the object is obtained based on the frequency distribution of the angle difference data values stored in.