JP2011221939A - Object recognition device, object recognition method and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an image recognition device capable of performing detection processing even when illuminating directions and illumination light to an object for recognition are different while reducing memory capacity.SOLUTION: The image recognition device has: a feature point data storage part in which feature point data of a plurality of first feature points extracted from a registered image and a displacement vector from the first feature points to a reference point of the registered image is stored; a feature point extraction part which extracts a second feature point of the object for recognition of the input image; a candidate coordinate point calculation part which calculates a candidate coordinate point of a reference point candidate from coordinates of the second feature point and the displacement vector for every rotation angle obtained by sequentially rotating the displacement vector at a preset unit angle centering around the second feature point; a candidate coordinate point voting part which performs voting by individually integrating a cosine value and a sine value of the rotation angle of the displacement vector for every candidate coordinate point; and an object for recognition determination part which detects a candidate coordinate point with the largest numeric value obtained by adding a square value of an integrated value, considers the detected candidate coordinate point as coordinates of the reference point, and detects the rotation angle as a rotation angle to the object for recognition with respect to the registered image.

Description

  The present invention relates to an object recognition apparatus, an object recognition method, and a program for detecting a position, posture (rotation angle), and scaling factor of a reference object registered in advance in an input image.

As a detection method for detecting a position of a reference object registered in advance (position coordinates (x, y) in a two-dimensional coordinate system), a rotation angle θ, a scaling factor S, and the like from an input image input from an imaging device or the like, correlation The method, the Fourier descriptor method, the moment method, the Hough transform method, etc. are well known.
In any of the detection methods described above, the feature points of the registered image of the registered reference object are extracted and registered in advance.
Then, using the feature points of the recognition target object extracted from the input image and the feature points of the registered image registered, the position of the image corresponding to the registered image in the input image, and the attitude to the registered image And a scaling factor.

In the detection method using the feature points described above, the Hough transform method is based on a simple majority rule called voting. Therefore, it has robust performance in the recognition target object recognition process against partial occlusion due to contact or overlap of parts of the recognition target objects in the input image, image noise, and the like.
In this Hough transform method, there is a generalized Hough transform method as a method for detecting a recognition target object having an arbitrary shape at high speed (for example, see Non-Patent Document 1).

In the above generalized Hough transform method, in the feature point (pixel unit) extracted from the reference object, the tangent at the feature point or the angle of the normal line, and the position of the reference point serving as the basis of the feature point and the object Prepare the relationship as a reference table. This reference point is an arrangement reference point indicating the arrangement position of the reference object on the two-dimensional coordinates.
That is, as shown in the flowchart of FIG. 10 for generating the reference table of the registered image as a model, the image of the registered image is input (step S1), and the feature point (P) of the reference object as shown in FIG. Is extracted by the second derivative (Laplacian) method or the Moravec method (step S2). In other words, the points on the contour line, which are points having a large difference in value (pixel value) from the surrounding pixels, and the point where the surface state changes greatly are characteristic points.

  Then, binarization processing of feature points in the registered image is performed using a preset threshold value, and binarized data is output (step S3). As a binarization method performed here, a threshold when the inter-class variance becomes maximum when a histogram is divided into two classes with a certain threshold, including a simple binarization method with a preset threshold, is set to 2. Otsu method as a threshold value (“Automatic threshold selection method based on discrimination and least squares criterion” (Otsu), IEICE Transactions, Vol. J63-D, No. 4. pp. 349-356, 1980), or a binarization method for setting a threshold according to the local density for an image having gradation.

  Next, the storage area of the reference table that stores data of each feature point of the registered image is initialized (step S4), and the reference point of the reference object is obtained as an arbitrary position, for example, the barycentric position of the reference object (step S5). ). It is determined whether or not the processing for obtaining the feature point data including the displacement vector from the reference point to the feature point and the unit normal vector at the feature point has been performed for all the binarized data (step S6). ) When the process for acquiring feature point data in all binarized data ends, the process ends. If there is binarized data that has not undergone the process of acquiring feature point data, the process proceeds to step S7.

Unprocessed binarized data is read (step S7), and whether or not the binarized data is a feature point is determined based on the threshold value (step S8). If it is not a feature point, the process proceeds to step S6. If it is a feature point, the process proceeds to step S9.
Then, a tangent or unit normal vector at the feature point (simply referred to as a point in the figure) is obtained, and the process proceeds to step S6 (step S9).
For example, in step S9, the normal vector N1 at the point P1 is obtained, the normal vector at the point P2 is obtained, and the normal vectors are similarly obtained at other points. These normal vectors are standardized as unit normal vectors so that the lengths thereof are the same, and as shown in FIG. 12, x corresponding to the feature point numbers and constituting the unit normal vectors at the feature points A reference table of the component NXn (an integer of 1 ≦ n ≦ M, M is the number of feature points), the y component NYn, and the x component DXn and the y component DYn constituting the displacement vector Dn at the feature point is generated.

The image recognition process according to Non-Patent Document 1 using the reference table is performed by the operation of the flowchart shown in FIG. An input image is input from the imaging device or the like (step S11), and as shown in FIG. 14, the feature point (Q) of the recognition target object in the input image is extracted by the same method as the reference image (step S12). In FIG. 14, the feature point of the registered image is superimposed on the feature point of the recognition target object, the angle of the normal vector at the feature point of the registered image is matched, and the position of the reference point indicated by the displacement vector of the registered image It is a conceptual diagram which shows performing voting.
Then, the binarization processing of the feature points in the input image is performed using a preset threshold value similarly to the processing of the reference image, and binarized data is output (step S13).

Next, the storage area of the voting space (two-dimensional plane showing the coordinates of the reference point for each rotation angle in the voting space conceptual diagram of FIG. 15) for voting on the reference point candidates is initialized (step S14).
It is determined whether or not the processing for obtaining the feature point data including the displacement vector from the reference point to the feature point and the unit normal vector at the feature point has been performed for all the binarized data (step S15). ) When the feature point data acquisition process for all binarized data has been completed, the process ends. When there is binarized data for which feature point data acquisition processing has not been performed, the process proceeds to step S16. .

Unprocessed binarized data is read (step S16), and whether or not the binarized data is a feature point is determined based on the threshold value (step S17). If it is not a feature point, that is, smaller than the threshold value. The process proceeds to step S15, and if it is a feature point, that is, if it is equal to or greater than the threshold value, the process proceeds to step S18.
And the unit normal vector of a feature point is calculated | required and a process is advanced to step 19 (step S18).
It is determined whether or not the reference point voting process corresponding to the selected feature point of the recognition target object has been performed using all the feature points of the registered image (step S19), and the selected feature of the recognition target object is determined. If the voting process using all the feature points of the registered image is completed for the point, the process proceeds to step S15. On the other hand, if there is a feature point of the registered image that is not used for the voting process, the process is performed. Proceed to step S20.

  The feature point of the registered image is superimposed on the feature point of the recognition target object, and the angle difference (rotation angle θ) between the unit normal vector at the feature point of the recognition target object and the unit normal vector at the feature point of the registration image Is obtained (step S20).

Next, the displacement vector of the feature point of the registered image is rotated by the following equation about the feature point by the rotation angle θ (step S21).
A COS value and a SIN value are obtained from the unit normal vector at the feature point of the recognition target object and the unit normal vector at the feature point of the registered image. For example, when it is assumed that the point P1 in FIG. 11 is the point Q2 in FIG. 14, it can be obtained by the following equation (1). COS = Nx1 x nx2 + Ny1 x ny2
SIN = Nx1 × ny2−Ny1 × nx2 (1)
Here, the unit normal vector of the registered image is (Nx1, Ny1), and the unit normal vector of the recognition target object is (nx2, ny2). Here, since the displacement vector is also rotated with respect to the feature point by the rotation angle θ, the components dx and dy of the rotated displacement vector can be obtained by the following equation (2).
dx = COS × DX−SIN × DY
dy = SIN × DX + COS × DY (2)
Here, the displacement vector of the registered image is (DX, DY).

The candidate position of the reference point is obtained by adding the rotated displacement vector to the position of the feature point of the recognition target object in the input image as shown in the following equation (step S22). If the coordinates of the point Q2 are (Qx, Qy) according to the above-described equation, the reference point candidate of the recognition target object is obtained by the following equation (3).
Rx = Qx + dx
Ry = Qy + dy (3)

Then, “1” is added to the position (Rx, Ry) of the reference point candidate with respect to the coordinates of the two-dimensional plane corresponding to the rotation angle θ, that is, voting is performed (step S23). As described above, in the method of Non-Patent Document 1, a displacement vector indicating the angle of each normal line of a registered image and a recognition target object at each feature point (the rotation angle at this time becomes a posture) is shown. Vote for the location of the reference point.
As a result, when the superimposed feature points correspond to the registered image and the recognition target object, when the normal vector angle is adjusted, the displacement vector of the registered image points in the direction of the same point, Since the reference points overlap, it is possible to detect the position where the number of votes is the peak as the position of the reference point of the recognition target object, and to obtain the rotation angle at that time as the posture with respect to the registered image. By this process, the recognition process between the registered image and the recognition target object in the input image is performed very efficiently.

However, considering the dimension of this voting space, even if the scaling factor S is not taken into consideration, as shown in FIG. 15, the data structure is vertical position (vertical coordinate), horizontal position (horizontal coordinate), and rotation angle. 3D. In addition, when the size change due to the scaling factor is also taken into consideration in the object detection, one dimension is added and becomes four dimensions.
For this reason, when applied to actual object recognition, a very large amount of memory is required to construct three-dimensional and four-dimensional voting spaces. In the following description, evaluation is performed on a three-dimensional voting space excluding the dimension of the scaling factor.

In the case of this three-dimensional voting space, the number of bits “NX × NY × NA × NV” is required in the method of Non-Patent Document 1. Here, when the position of the recognition target object in the input image in the two-dimensional coordinates and the detection result of the rotation angle with respect to the registered image are shown, the number of divisions of the detection position of the x-axis (horizontal axis) of the two-dimensional coordinates (corresponding to the resolution) The number of coordinate points NX, the number of coordinate points NY that is the number of divisions of the detection position of the y-axis (vertical axis), the angle division NA that is the number of divisions of the detection angle of the rotation angle, and the number of bits NV of the memory that accumulates the number of votes (Normally 2 bytes or 4 bytes).
That is, in Non-Patent Document 1, a large amount of memory capacity is required because each vote corresponds to each position and angle of a recognition target object candidate.
In order to reduce this large memory capacity, there is a method of simply integrating the COS value and the SIN value corresponding to the rotation angle θ of the object detection candidate for each position of the object detection candidate (see, for example, Patent Document 1).

Image recognition processing using Patent Document 1 is performed by the operation of the flowchart shown in FIG.
An input image is input from an imaging device or the like (step S31), and as shown in FIG. 14, feature points of the recognition target object in the input image are extracted by the same method as the reference image (step S32).
Then, the binarization processing of the feature points in the input image is performed using a preset threshold value similarly to the processing of the reference image, and binarized data is output (step S33).

Next, the storage area of the voting space for voting on the reference point candidates (each two-dimensional plane for accumulating COS values and SIN values described later) is initialized (step S34).
It is determined whether or not the process of integrating the COS value and SIN value at each feature point with respect to the position of the reference point candidate is performed for all binarized data (step S35). When the COS value and SIN value integration process for the binarized data is completed, the process ends. When there is binarized data that has not been subjected to the COS value and SIN value integration process, the process proceeds to step S36. Proceed to

Unprocessed binarized data is read (step S36), and whether or not the binarized data is a feature point is determined based on the threshold value (step S37). If it is not a feature point, the process proceeds to step S35. If it is a feature point, the process proceeds to step S38.
And the unit normal vector of a feature point is calculated | required and a process is advanced to step 39 (step S18).
Using all the feature points of the registered image, it is determined whether or not the COS value and the SIN value have been added to the reference point corresponding to the selected feature point of the recognition target object (step S39). When the integration process using all the feature points of the registered image is completed for the selected feature point of the object, the process proceeds to step S35, while there are feature points of the registered image that are not used for the integration process. If it exists, the process proceeds to step S40.

Based on the unit normal vector (nx, ny) of the feature point of the recognition target object in the input image and the unit normal vector (Nx, Ny) of the feature point of the registered image, A COS value and a SIN value are obtained (step S40).
COS = Nx × nx + Ny × ny
SIN = Nx × ny−Ny × nx (4)

The displacement vector (dx, dy) rotated by the rotation angle θ from the COS value and SIN value obtained by the equation (4) and the displacement vector (DX, DY) of the feature point of the registered image that has been superimposed is expressed as follows: Obtained by equation (5).
dx = COS × DX−SIN × DY
dy = SIN × DX + COS × DY (5)

Then, from the feature point (Qx, Qy) of the recognition target object in the input image and the rotated displacement vector (dx, dy), the position of the reference point position candidate is obtained by the following equation (6). Rx = Qx + dx
Ry = Qy + dy (6)
With respect to the obtained candidate position (Rx, Ry) of the reference point position, COS (θ) and SIN (θ) are respectively represented as voting space of COS (θ) and SIN as shown in FIG. It accumulates in the voting space of (θ). As a result, the memory capacity as the voting space is obtained by “NX × NY × 2 × (number of bits for integration)”. Since the number of bits to be integrated is both 2 to 4 bits, the angle resolution number NA is set to 360 and the resolution of 1 degree is considered. It can be realized with a memory capacity of 180. Here, 2 indicates two two-dimensional coordinates for integrating the COS value and the SIN value, respectively.

By this voting process, when the candidate point of the reference point of the recognition target object is a true reference point having a rotation angle θ with respect to the registered image, from all of the feature points of the registered image with respect to the candidate point. It can be considered that the majority of the votes of the rotation angle θ are. For this reason, each of the integrated values of the COS value and the SIN value at the point that is the true reference point is determined by the vote of the rotation angle θ of the recognition target object, and the COS corresponding to the rotation angle θ. Value, an integer multiple of the SIN value, and the number of votes. As a result, a peak can be obtained in one or both of the COS integrated value and the SIN integrated value.
Therefore, the peak position of the COS integrated value and the SIN integrated value is the reference point of the recognition target object in the input image, and the ratio between the integrated value of the COS value integrated to the reference point and the integrated value of the SIN value, that is, tan ^{− The} rotation angle θ can be obtained from ¹ (integrated value of SIN value / integrated value of COS value).

On the other hand, if the reference point candidate point is not a true reference point, the vote at the feature point of the recognition target object may be a false vote or may be caused by noise, and the voted rotation angle θ is uniformly random. It is assumed to have a distribution.
As a result, the integration of the COS value and the SIN value at a point that is not the true reference point is an integral multiple of the COS function and the SIN function, each of which is simply integrated from 0 degrees to 360 degrees with a uniform weight. It is expected that the integrated value will approach zero.
As a method for detecting a recognition target object having an arbitrary shape, a correlation method or convolution is used (for example, see Non-Patent Document 2).

JP 2007-128374 A

D.H. Ballard, “Generalizing The Hough Transform To Detect Arbitrary Shapes”, Pattern Recongition, Vol. 13, no. 2, pp. 111-122, 1981 “New Image Analysis Handbook”, 2004, University of Tokyo Press, Part 2, Chapter 4, “Matching”

However, in the object recognition methods shown in Patent Document 1 and Non-Patent Document 1, the displacement vector is rotated based on the rotation angle of the recognition target object in the input image with respect to the registered image image, and the reference point candidate points are obtained. In order to calculate, the feature point extracted from the registered image and the vector information of the normal (or tangent) at the feature point are required.
For this reason, a process for obtaining normal (or tangent) vector information for each feature point of the registered image is required.
In addition, systematic errors may occur in the normal vectors and tangent values obtained from calculations due to differences in the illumination direction and illumination light for the recognition target object when the registered image and the input image are captured. It is done.

In addition, due to conditions such as different imaging distances when the registered image and the input image are captured, an error also occurs in scaling corresponding to the difference in the apparent size (size) of the recognition target object. Can be considered.
That is, as already described, the estimation of the magnitude of the vector and the normal vector at an arbitrary feature point can be obtained by the change in luminance near that point.
Therefore, the normal vector and the like are greatly affected by changes in the illumination conditions.
FIG. 18 is a conceptual diagram showing that an error occurs because the resolution of pixels that capture an input image is equal when the sizes of the input images are different. As shown in FIG. 18, the resolution of the pixels of the image sensor is constant with respect to the change in the size of the image to be extracted. If the input image is reduced as shown in FIG. 18B or the input image is enlarged as shown in FIG. 18C with respect to the registered image shown in FIG. The normal vector at each feature point varies depending on the change in the luminance value of the neighboring image, depending on the size of the object on the image sensor.
If there is a change in the apparent size of the input image (ie, scaling) due to the error as described above, the normal vector at the feature point of the registered image and the normal vector at the feature point of the input image An error also occurs in the result of voting using reference point candidate coordinates.

In the object recognition method of Patent Document 1, the voting operation at the feature point is performed isotropically except for the rotation angle θ at the correct position, and the voting results are expected to be averaged and approach “0”. ing.
However, due to an error in the normal vector between the registered image and the recognition target object of the input image, if a large number of votes are accumulated at incorrect coordinates, the recognition target object detection process will not be performed correctly.
Further, the object recognition method disclosed in Patent Document 1 includes a plurality of recognition target objects in an input image, and when these recognition target objects overlap or are arranged close to each other, integration by a large number of votes at incorrect coordinates. The recognition target object detection process cannot be performed correctly. At this time, since the voting process itself is normally performed, it cannot be directly detected whether or not the detection result is correct.

The case where patent document 1 cannot perform a recognition process is demonstrated using FIG. FIG. 19 is a conceptual diagram illustrating the configuration and arrangement of a recognition target object.
Model 1 in FIG. 19A and model 2 in FIG. 19B are registered images having the reference point at the center and the reference point at the same position. FIG. 19C shows a case where the models 1 of the same registered image overlap with each other with the rotation angles θ set to 0 degrees and 180 degrees. In this case, in Patent Document 1, since the integration of 0 degrees and 180 degrees overlaps at the same coordinate, both the COS value and the SIN value are averaged to “0”, and accurate detection cannot be performed.
In FIG. 19D, model 1 is arranged at 0 degrees and model 2 is arranged at 180 degrees. Also in this case, since the reference points are the same, the state is the same as in FIG. 19C, and it is highly possible that neither model 1 nor model 2 can be detected.

On the other hand, since the object recognition method of Non-Patent Document 2 uses a correlation method and does not require normal vectors or tangents at feature points, problems in Patent Document 1 and Non-Patent Document 1 do not occur.
However, in the case of the simple correlation method, the method for reducing the calculation amount described in Non-Patent Document 1 and the method for reducing the memory capacity described in Patent Document 1 are not used.
In other words, in the correlation method, a large amount of memory capacity is required for voting on the two-dimensional candidate coordinates of the reference point provided for every rotation angle θ of the registered image, and calculation of correlation values is performed. It is necessary to perform a large amount of arithmetic processing in

  The present invention has been made in view of such circumstances, and its purpose is to reduce the amount of calculation and memory capacity required for image recognition, and even if the illumination direction and illumination light for the recognition target object are different, An object of the present invention is to provide an image recognition apparatus, an image recognition method, and a program capable of performing this recognition target object detection processing.

The object recognition apparatus of the present invention stores feature point data including a plurality of first feature points extracted from a registered image and a displacement vector from the first feature point to a reference point of the registered image. A point data storage unit, a feature point extraction unit that extracts a second feature point from the input image, and for each of the second feature points, all the first feature points with the second feature point as a rotation center Each of the displacement vectors is sequentially rotated at a preset unit angle, and for each rotated angle, candidate coordinate points that are candidates for reference points are determined from the coordinates of the second feature point and the displacement vector. A candidate coordinate point calculation unit for calculating, a candidate coordinate point voting unit for individually integrating voting values based on a cosine value and a sine value corresponding to the rotation angle of the displacement vector for each candidate coordinate point, a cosine value and a sine The value obtained by adding the square value of the integrated value with the value is A candidate coordinate point that is large is detected from among the candidate coordinate points, and the detected candidate coordinate point is used as a reference point coordinate, and a rotation angle corresponding to the coordinate point is set as the rotation angle of the input image with respect to the registered image. And a recognition target object determination unit that detects the rotation angle.
In the object recognition method of the present invention, the registered image data generation unit includes a plurality of first feature points extracted from the registered image, and a displacement vector from the first feature point to a reference point of the registered image. A feature point data storing process in which point data is stored, a feature point extracting unit extracting a second feature point from the input image, and a candidate coordinate point calculating unit for each second feature point Further, with the second feature point as the rotation center, each of the displacement vectors of all the first feature points is sequentially rotated by a preset unit angle, and for each rotated rotation angle, the second feature point is A candidate coordinate point calculation process for calculating a candidate coordinate point that is a candidate for a reference point from the coordinates of a feature point and the displacement vector, and a candidate coordinate point voting unit corresponding to the rotation angle of the displacement vector for each candidate coordinate point Voting value based on cosine value and sine value Candidate coordinate point voting process for individual accumulation, and the target object determination unit detects the candidate coordinate point having the maximum numerical value obtained by adding the square value of the accumulated value of the cosine value and the sine value from the candidate coordinate points. And a recognition target object determining step of detecting the detected candidate coordinate point as a reference point coordinate and detecting a rotation angle corresponding to the coordinate point as a rotation angle of the input image with respect to the registered image. And
The program of the present invention is a feature point data comprising a plurality of first feature points extracted from a registered image by a registered image data generation unit and a displacement vector from the first feature point to a reference point of the registered image. For each of the second feature points, a feature point extraction process for extracting a second feature point from the input image, and a candidate coordinate point calculation unit for each of the second feature points. With the second feature point as the center of rotation, each of the displacement vectors of all the first feature points is sequentially rotated at a preset unit angle, and the second feature point at each rotated rotation angle. A candidate coordinate point calculation process for calculating a candidate coordinate point that is a candidate for a reference point from the coordinates and the displacement vector, and a candidate coordinate point voting unit corresponding to the rotation angle of the displacement vector for each candidate coordinate point Value and sine value The candidate coordinate point voting process to be separately accumulated, and the target object determination unit detects the candidate coordinate point having the maximum value obtained by adding the square value of the accumulated value of the cosine value and the sine value from the candidate coordinate points. A program for causing a computer to execute recognition target object determination processing in which a detected candidate coordinate point is used as a reference point coordinate and a rotation angle corresponding to the coordinate point is detected as a rotation angle of the input image with respect to the registered image. It is.
With the configuration described above, according to the object recognition device of the present invention, the posture of the object is detected using the angle of the displacement vector. Object detection can be performed.
Further, according to the object recognition apparatus of the present invention, since the cosine value and the sine value at each angle rotated from the angle of the displacement vector of the registered image are integrated, the rotation angle is obtained from the cosine value and the sine value. Therefore, the angle of the attitude displacement of the input image with respect to the registered image can be easily obtained as the rotation angle.

The object recognition apparatus of the present invention uses the x-direction component and the y-direction component of the unit vector of the displacement vector in the two-dimensional coordinate system as the cosine value and sine value of the displacement vector in the registered image, and the vote value And calculating a cosine value and a sine value corresponding to the rotation angle of the displacement vector.
According to the object recognition apparatus of the present invention, the voting value for detecting the posture of the object is calculated using the angle of the displacement vector. And robust object detection.

In the object recognition device of the present invention, the candidate coordinate point voting unit sets the angle of the displacement vector in the registered image as a first angle, and sets the rotation angle for rotating the displacement vector by a unit angle as a second angle. In this case, the cosine value and the sine value when the displacement vector is rotated are calculated and stored in advance from the first angle and the second angle that increases for each unit angle.
With this configuration, according to the object recognition device of the present invention, the vote value is sequentially integrated to the candidate coordinate points on the circle whose radius is the length of the displacement vector, and the displacement vector angle stored in advance is Since the cosine value and sine value corresponding to the unit rotation angle are used as voting values, it is possible to shorten the time for calculating the voting values to be integrated with respect to the candidate coordinate points, and to speed up the processing.

In the object recognition apparatus of the present invention, the candidate coordinate point calculation unit expands and contracts the length of the displacement vector within a preset magnification range, and the candidate coordinate points in the range of the longest length and the shortest length. Is a coordinate point range, coordinate points corresponding to the length of the displacement vector in the registered image in the coordinate point range are candidate point coordinates represented by the coordinate point range, and all of the candidate point coordinate values in the coordinate point range The voting value of is integrated.
With this configuration, according to the object recognition device of the present invention, even when the input image is scaled with respect to the registered image, it is possible to easily and robustly detect an object as long as the magnification range is expanded and contracted in advance. Since it is not necessary to prepare a memory corresponding to scaling, the voting memory can be reduced.

In the object recognition device of the present invention, the candidate coordinate point voting unit
A random number is generated, the absolute value of each cosine value and sine value is compared with the random number, the voting value is defined as 0 when the random number is larger, and the cosine value is positive when the random number is smaller. In this case, the voting value corresponding to the cosine value is defined as 1, the voting value corresponding to the sine value is defined as 1 when the sine value is positive, and the cosine value corresponds to the case when the cosine value is negative. The voting value to be defined is defined as -1, and when the sine value is negative, the voting value corresponding to the sine value is defined as -1.
With this configuration, according to the object recognition device of the present invention, when distributed processing is performed using a plurality of personal computers, the amount of memory for holding intermediate vote results while the vote values are being accumulated is reduced. Therefore, it is possible to reduce the amount of communication when counting the voting results between the personal computers.

It is a block diagram which shows the structural example of the object identification device by 1st Embodiment of this invention. It is a figure explaining the displacement vector which shows the relative position of the reference point R from the 1st feature point extracted from a registration image. The figure which shows the structure of the 1st feature point table memorize | stored in the feature point data storage part 2 which shows the x direction unit component in the unit vector of the displacement vector of a 1st feature point, the y direction unit component, and a displacement vector. It is. It is a figure explaining how to obtain the candidate coordinate point which is the candidate of the reference point of the input image by superimposing the first feature point on the second feature point of the input image. It is a flowchart explaining the example of an operation | movement of the recognition process of the input image and registration image of the recognition target object in 1st Embodiment. It is a figure which shows the result of having calculated the vote value by superimposing the 1st feature point of a registration image on each coordinate point of 2nd feature point P1 and P2 of an input image. It is a figure explaining the process which performs the object recognition in consideration of the scaling with respect to the registration image of an input image. It is a flowchart explaining the example of an operation | movement of the recognition process of the input image and registration image of the recognition target object in 2nd Embodiment. It is a flowchart explaining the example of an operation | movement of the recognition process of the input image and registration image of a recognition target object in 3rd Embodiment. It is a flowchart explaining operation | movement of extraction of the 1st feature point of a registration image, and extraction of the feature point data in this 1st feature point. It is a conceptual diagram explaining extraction of feature point data. It is a conceptual diagram which shows the table structure in which feature point data are memorize | stored. 10 is a flowchart showing an operation of image recognition processing using generalized Hough transform of Non-Patent Document 1. It is a conceptual diagram explaining the voting process of the reference point in FIG. It is a conceptual diagram explaining the memory capacity required in the case of the image recognition in the generalized Hough conversion of a nonpatent literature 1. 10 is a flowchart for explaining the operation of image recognition processing in Patent Document 1. It is a conceptual diagram explaining the integration | accumulation process of the COS value and SIN value in patent document 1. FIG. It is a figure which shows that an error generate | occur | produces when the magnitude | size of an input image differs because the resolution of the pixel which images an input image is equal. It is a conceptual diagram explaining the structure and arrangement | positioning of a recognition target object.

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a schematic block diagram showing a configuration example of an object recognition apparatus according to an embodiment of the present invention.
In this figure, the object recognition apparatus according to the present embodiment has a registered image data generation unit 1, a feature point data storage unit 2, and an object recognition unit 3.
The registered image data generation unit 1 extracts feature points extracted from the registered image as an array of first feature points, sets a reference point (for example, the center of gravity of the registered image) O of the registered image, and starts from the first feature points. The displacement vector to the reference point is obtained, and the magnitude of the displacement vector is obtained.
FIG. 2 is a diagram for explaining a displacement vector indicating a relative position of the reference point O from the first feature point extracted from the registered image in the present embodiment. The reference point is an arrangement reference point indicating the arrangement position of the registered image on the two-dimensional coordinates. The reference point O may be an arbitrary point on the two-dimensional coordinate where the converted image is arranged. For example, the center of gravity of the converted image may be used as the reference point.
In FIG. 2, for example, a displacement vector R1 from the feature point P1 to the reference point O and a displacement vector R2 from the feature point P2 to the reference point O are shown. The angle θ of the displacement vector is set as an angle formed from the positive direction of the shaft in the direction opposite to the direction in which the timepiece of the watch rotates, taking the displacement vectors R1 and R2 of FIG. 2 as an example.

Returning to FIG. 1, the registered image data generation unit 1 obtains a unit vector of the displacement vector, and obtains an x-direction unit component of this unit vector (a component in the direction parallel to the x-axis of the displacement vector in the two-dimensional coordinate system) and A y-direction unit component (a component in a direction parallel to the y-axis of the displacement vector in the two-dimensional coordinate system) is obtained.
The registered image data generation unit 1 associates the obtained displacement vector and the x-direction unit component and the y-direction unit component in the unit vector of the displacement vector with the feature point number that is identification information of the first feature point, Each feature point is stored in the feature point data storage unit 2.
FIG. 3 shows, for each first feature point, the feature point data storage unit 2 that shows the x-direction unit component, the y-direction unit component, and the displacement vector in the unit vector of the displacement vector of the first feature point. It is a figure which shows the structure of a 1st feature point table. As shown in FIG. 3, the x-direction unit component and the y-direction unit component in the unit vector of the displacement vector and the displacement vector for each feature point are stored as a first feature point table corresponding to the feature point number. ing.

The object recognition unit 3 obtains the second feature points extracted from the recognition target objects in the input image as an array.
Further, the object recognition unit 3 superimposes the first feature point of the converted image on the second feature point, obtains a candidate coordinate point of the reference point of the recognition target object from the displacement vector from the first feature point, The correlation between the recognition target object and the registered image is determined based on the voting result for the candidate coordinate points.
Hereinafter, the configuration of the registered image data generation unit 1 and the object recognition unit 3 described above will be described in detail.

The registered image data generation unit 1 includes a registered image acquisition unit 11, a feature point extraction unit 12, a feature point binarization unit 13, and a displacement vector extraction unit 14. The operations of the registered image acquisition unit 11, the feature point extraction unit 12, and the feature point binarization unit 13 in the registered image data generation unit 1 are the same as the operations in the flowchart of FIG.
The registered image acquisition unit 11 inputs image data of a registered image to be used later for recognizing the recognition target object from the imaging device or the CAD system.
The feature point extraction unit 12 extracts the first feature point (pixel unit) of the registered image using the second-order differential (Laplacian) method or the Morabek method described in the feature extraction in Non-Patent Document 1. .

The feature point binarization unit 13 performs binarization based on the gradation between the pixel of the first feature point of the registered image and the pixels other than the first feature point. For example, the feature point binarization unit 13 sets the gradation level of the pixel of the first feature point to “1” and the gradation level of the pixels other than the first feature point to “0” based on a threshold value of the gradation level set in advance. And Here, as the preset threshold value, a threshold value for performing the binarization processing described in the generalized Hough transform method is used.
The displacement vector extraction unit 14 obtains a displacement vector from the first feature point to the reference point for each first feature point. Further, the displacement vector extraction unit 14 uses the obtained displacement vector and the x-direction unit component and the y-direction unit component in the unit vector of the displacement vector as first feature point data, and the first feature corresponding to the feature point number. It is written and stored in the feature point data storage unit 2 as a point table.

Next, the object recognition unit 3 includes an input image acquisition unit 31, a feature point extraction unit 32, a feature point binarization unit 33, a candidate coordinate point calculation unit 34, a candidate coordinate point voting unit 35, and a recognition target object determination unit 36. Have. The input image acquisition unit 31 inputs image data of an input image including a recognition target object from an imaging device such as a camera.
The feature point extraction unit 32 extracts the coordinate value of the second feature point (pixel unit) of the recognition target object in the input image using the same method as the feature point extraction unit 12.

  The feature point binarization unit 33 performs binarization based on the gradation of pixels of the second feature point of the recognition target object and pixels other than the second feature point. For example, the feature point binarization unit 33 sets the gradation levels of the pixels of the second feature point to “1” and the gradation levels of the pixels other than the second feature point to “0” based on a threshold value of the gradation level set in advance. And Here, a numerical value similar to that of the feature point binarization unit 13 is used as the preset threshold value.

The candidate coordinate point calculation unit 34 superimposes the coordinate value of the second feature point and the coordinate value of the first feature point, and calculates the displacement vector of the registered image from the angle θ around the second feature point. The x coordinate value X _pos of the candidate coordinate point, which is the candidate coordinate of the reference point indicated by the displacement vector R when rotated by the rotation angle φ, is obtained by the following equation (7).
X _pos = R cos (φ) + X (7)
In this equation (7), R is the length of the displacement vector, the angle φ is the angle of the rotation angle that rotates the displacement vector of the registered image from the angle θ with the rotation center as the second feature point, and X is This is the x coordinate value of the second feature point.
Further, the candidate coordinate point calculation unit 34 superimposes the coordinate value of the second feature point and the coordinate value of the first feature point, and calculates the displacement vector of the registered image with the second feature point as the center. The y-coordinate value Y _pos of the candidate coordinate point, which is the candidate coordinate of the reference point indicated by the displacement vector R when rotated by the rotation angle φ from the angle θ, is obtained by the following equation (8).
Y _pos = Rsin (φ) + Y (8)
In this equation (8), R is the length of the displacement vector, the angle φ is the angle of the rotation angle that is rotated from the position of the displacement vector angle θ of the registered image with the rotation center as the second feature point, and Y Is the y coordinate value of the second feature point.

FIG. 4 is a diagram illustrating obtaining candidate coordinate points from the above-described equations (7) and (8) by superimposing the first feature points on the second feature points of the input image. In this figure,
When the second feature point P1 is the first feature point P1, the candidate coordinate point calculation unit 34 rotates the candidate coordinate points arranged on the circle C having the radius R (displacement vector R) by unit angle. For each rotation angle φ at the time of calculation, the calculation is performed using the equations (7) and (8).
For example, when the second feature point P1 is the first feature point 1, candidate calculation points arranged on the circle C11 are obtained. On the other hand, when the second feature point P1 is the first feature point 2, candidate calculation points arranged on the circle C12 are obtained. When the second feature point P1 of the input image is the first feature point 1 of the registered image, the angle θ of the displacement vector is θ1, and the second feature point P1 of the input image is the first feature point of the registered image. When it is point 2, the angle θ of the displacement vector is θ.
Similarly, when the second feature point P2 is the first feature point 1, candidate calculation points arranged on the circle C21 are obtained. On the other hand, when the second feature point P2 is the first feature point 2, candidate calculation points arranged on the circle C22 are obtained.
For example, when the unit angle for changing the rotation angle φ is 1 degree, candidate coordinate points with an angle φ from 0 degree to 359 degrees are calculated on the above-described circle.

Further, when the angle θ of the displacement vector is obtained and the rotation angle φ of the converted image is rotated for each unit angle using the angle θ as a reference value, for example, when the unit angle is 1 degree, cos from 0 to 359 is obtained. By setting the value in the storage unit inside the candidate coordinate point calculation unit 34 in advance, the candidate coordinate point calculation unit 34 can easily calculate the X coordinate value X _pos at each angle φ from the equation (7). Become. Similarly, when the unit angle is set to 1 degree, a sine value with φ ranging from 0 to 359 is set in advance in the storage unit inside the candidate coordinate point calculation unit 34, so that the Y coordinate value Y at each angle φ is set. _pos can be easily calculated by the equation (8).

Returning to FIG. 1, the candidate coordinate point voting unit 35 calculates the voting value to vote for the coordinate position of the candidate coordinate point on the cos plane and the sin plane of FIG. 17 by the following formulas (9) and (10).
Cos = cos (θ−φ) = Ux · cos (φ) + Uy · sin (φ) (9)
Sin = sin (θ−φ) = Uy · cos (φ) −Ux · sin (φ) (10)
Here, as described in the first feature point table in FIG. 3, Ux = sin (θ) by the x-direction unit component and the y-direction unit component in the unit vector of the displacement vector of each first feature point. ) And Uy = cos (θ). The vote value is obtained as numerical data in a floating-point or fixed-point format.
In the present embodiment, the candidate coordinate point voting unit 35 obtains each of the cosine value Cos that is the voting value calculated by the equation (9) and the sine value Sin that is the voting value calculated by the equation (10). The candidate coordinate points on the Cos plane and Sin plane are added to the candidate coordinate points calculated by the calculation unit 34. That is, the candidate coordinate point voting unit 35 accumulates voting values in a portion corresponding to a coordinate point having the same coordinate value as that of the candidate coordinate point in the storage table corresponding to each coordinate point on the Cos plane and the Sin plane.

The recognition target object determination unit 36 sequentially selects coordinate points that are candidate coordinate points of the same coordinate value on the Cos plane and the Sin plane described above, and the square value of the vote value accumulated at the candidate coordinate points of the Cos plane. And the square value of the voting value accumulated at the candidate coordinate points on the Sin plane, and the respective square values are added.
Further, the recognition target object determination unit 36 detects candidate coordinate points exceeding a preset peak value as reference points of the input image.
In addition, the recognition target object determination unit 36 uses the cosine value Cos and the sine value Sin obtained by integrating the peak positions to obtain the rotation angle φ of the input image with respect to the registered image according to the following expression (11).
φ = tan ⁻¹ (Sin / Cos) (11)
Through the processing described above, the coordinates of the reference point of the input image and the rotation angle of the input image with respect to the registered image can be obtained.

Further, when it is necessary for the detail detection unit 37 connected to the object recognition unit 3 to know the position of the recognition target object in detail, the reference point and the rotation angle θ extracted by the recognition target object determination unit 36 are A process for extracting in detail the position where the recognition target object is arranged is performed.
Therefore, if there is no need to extract the detailed arrangement position of the recognition target object, it is not necessary to provide the detail detection unit 37.

Next, the recognition process between the input image of the recognition target object and the registered image in the present embodiment will be described with reference to FIG. FIG. 5 is a flowchart for explaining an operation example of the recognition process.
The input image acquisition unit 31 inputs an input image from an imaging device (not shown), and outputs it to the feature point extraction unit 32 and the candidate coordinate point calculation unit 34 as image data expressed by pixels, for example (step S51).
When the input image is input, the feature point extraction unit 32 extracts the second feature point of the recognition target object in the input image and outputs the extracted input image to the feature point binarization unit 33 (step S52). .
Next, the feature point binarization unit 33 binarizes the gradation of the pixel of the second feature point and the other pixels for each pixel of the input image to be input, and stores it internally. (Step S53). The binarization performed here is the same as the processing described in the flowchart of FIG. 13 of the conventional example.

Then, the candidate coordinate point voting unit 35 initializes a storage table corresponding to each coordinate point in each of the Cos plane and the Sin plane, that is, sets the integrated value in the coordinate value of each coordinate point to “0” (step S54). .
Next, the candidate coordinate point calculation unit 34 determines whether or not the voting process as the evaluation value at the second feature point has been performed on the binarized data of all the pixels of the input image (step S55). ), When the voting at the second feature point is performed on the binarized data of all the pixels of the input image, while the recognition process is finished, on the other hand, when there is a pixel that has not been voted at the feature point, The process proceeds to step S56.
Here, the detection of the pixel for which the evaluation value is not calculated at the feature point is performed by, for example, fixing the y-axis axis coordinates and sequentially processing the pixels corresponding to the x-axis axis coordinates. When the processing on the x-axis axis coordinates corresponding to the axis coordinates is completed, the y-axis axis coordinates are changed by one, and the pixels corresponding to the x-axis axis coordinates are sequentially processed. Thus, the end point can be detected by forming a loop of the combination of the x-axis coordinate and the y-axis coordinate according to the number of pixels on the x-axis and the y-axis.

Next, the candidate coordinate point calculation unit 34 reads the unprocessed pixel binarization data from the feature point binarization unit 33 (step S56), and the read binarization data exceeds a preset threshold value. Then, it is determined whether or not this pixel is the second feature point (step S57).
At this time, if the read pixel is the second feature point, the candidate coordinate point calculation unit 34 advances the process to step S58, and if the pixel is not the second feature point, advances the process to step S55.
When the pixel is the second feature point, the candidate coordinate point calculation unit 34 calculates the candidate coordinate points of all the first feature points of the registered image in the coordinates of the second feature points of the input image, and this It is determined whether or not the voting value accumulation process at the calculated candidate coordinate points has been performed (step S58). If there is no first feature point that has not been processed, the process proceeds to step S55. If the first feature point that has not been processed remains, the displacement vector corresponding to the remaining coordinate point is read from the feature point data storage unit 2 and the rotation angle φ is initialized to “0”. Advances to step S59.

  Next, the candidate coordinate point calculation unit 34 rotates the displacement vector by a unit angle (for example, 1 degree) by 0 to 360 degrees (for processing, when the unit angle is 1 degree, from 0 degree to 359 degrees) It is determined whether all the candidate coordinate points on the circle whose radius is the length of the displacement vector have been generated due to the change in the rotation angle φ to the rotation angle φ in increments of 1 degree (step S59). When the displacement vector at the first feature point is rotated for all rotation angles φ from 0 to 360 degrees per unit angle, the process proceeds to step S58, while the rotation angle φ is 0 to 360. If the displacement vector is not completely rotated, the rotation angle of the displacement vector is rotated by the unit angle, and the process proceeds to step S60.

The candidate coordinate point voting unit 35 calculates a cosine value Cos and a sine value Sin, which are voting values, from the unit vector of the displacement vector at the first feature point of the registered image, which is model data, and the rotation angle φ. ) And (10) are calculated (step S60).
FIG. 6 is a diagram illustrating a result of calculating the vote value by superimposing the first feature point of the registered image on the coordinate points of the second feature points P1 and P2 of the input image. FIG. 6A shows the calculation result of the cosine value Cos that is the vote value for the cos plane, and FIG. 6B shows the calculation result of the sine value Sin that is the vote value for the Sin plane.
For example, in the voting process for the candidate coordinate points on the Cos plane in FIG. 6A, the candidate coordinate point voting unit 35 determines that the displacement vector in the direction of the point A when the second feature point P1 is the first feature point 1. Since the unit vector of is oriented so that φ is 0 degree at the point A, both the x-direction unit component Ux and cos (φ) in the unit vector of the displacement vector are “1”. Calculated as “1”. On the other hand, when φ is 90 degrees at the point B, the candidate coordinate point voting unit 35 has the y-direction unit component Uy in the unit vector of the displacement vector as 0 and sin (φ) as 1. Calculate “0”.

Further, for example, in the voting process for the candidate coordinate point on the Cos plane in FIG. 6A, the candidate coordinate point voting unit 35 determines the direction of the point A ′ when the second feature point P1 is the first feature point 2. Since the unit vector of the displacement vector is oriented at the point A ′, φ is 0 degree at the point A ′, and therefore the x-direction unit component Ux and cos (φ) in the unit vector are “1”, and the vote value is “1”. calculate. On the other hand, when φ is 90 degrees at the point B ′, the candidate coordinate point voting unit 35 has “0” as the voting value because the y-direction unit component Uy in the unit vector is 0 and sin (φ) is 1. Is calculated.
In the case of the voting process for the candidate coordinate points on the Sin plane in FIG. 6B, the candidate coordinate point voting unit 35 (10) is similar to the voting process for the candidate coordinate points on the cos plane in FIG. ) Calculate the vote value from the formula

Next, the candidate coordinate point calculation unit 34 calculates the cosine value Cos and the sine value Sin obtained by the candidate coordinate point calculation unit 34 from the displacement vector at the first feature point of the registered image, which is model data, and the rotation angle φ. The coordinate values of the candidate coordinate points on the Cos surface and the Sin surface for voting are calculated from each of the equations (7) and (8) (step S60).
For example, when the second feature point P1 of the input image is the feature point 1, the candidate coordinate point calculation unit 34 uses the displacement vector R1 at the first feature point 1 and the rotation angle φ to As a candidate coordinate point in each of the Cos plane and the Sin plane, when the rotation angle φ is 0 degree, the coordinate point of the point A is set as the displacement vector angle (θ1-0 = θ) as a candidate coordinate point When the rotation angle φ is 90 degrees, the coordinate point of point B is obtained as a candidate coordinate point as the angle (θ1-90) of the displacement vector.
Further, when the second feature point P1 of the input image is the feature point 2, the candidate coordinate point calculation unit 34 uses the displacement vector R2 at the first feature point 2 and the rotation angle φ to As a candidate coordinate point on each of the Cos plane and the Sin plane, when the rotation angle φ is 0 degree, the coordinate point of the point A ′ is set as the candidate coordinate point (θ2−0 = θ) as a candidate coordinate point. When the rotation angle φ is 90 degrees, the coordinate point of the point B ′ is obtained as a candidate coordinate point as the angle (θ2-90) of the displacement vector.

For example, when the second feature point P2 of the input image is the feature point 2, the candidate coordinate point calculation unit 34 uses the displacement vector R2 at the first feature point 2 and the rotation angle φ to As a candidate coordinate point in each of the Cos plane and the Sin plane, when the rotation angle φ is 0 degree, the coordinate point of the point A is set as the displacement vector angle (θ2-0 = θ) as a candidate coordinate point. Asking.
Further, when the second feature point P2 of the input image is the feature point 1, the candidate coordinate point calculation unit 34 uses the displacement vector R1 at the first feature point 1 and the rotation angle φ to Using the equation (8), as a candidate coordinate point on each of the Cos plane and the Sin plane, when the rotation angle φ is 90 degrees, the coordinate point of the angle (θ1-90) point D of the displacement vector is obtained as a candidate coordinate point.

The candidate coordinate point voting unit 35 adds the cosine value, which is the vote value, to the coordinate value of the storage table corresponding to the candidate coordinate point on the Cos surface obtained by the candidate coordinate point calculation unit 34.
Further, the candidate coordinate point voting unit 35 integrates the sine value as the vote value with respect to the coordinate value of the storage table corresponding to the candidate coordinate point of the Sin surface obtained by the candidate coordinate point calculating unit 34 (step S61). .

According to the object recognition apparatus of the first embodiment described above, since the posture of the object is detected using the angle of the displacement vector, it is more robust against changes in illumination or the like than in the conventional case using a normal vector or the like. Object detection can be performed.
Further, according to the object recognition device of the present invention, even when the input image is rotated with respect to the posture of the registered image, the cosine value and the sine value at the rotated angle from the angle of the displacement vector of the registered image are obtained. Each integration is performed by rotating the displacement vector around the second feature point, calculating candidate coordinate points for each rotation angle, and rotating the candidate coordinate points for the calculated candidate coordinate points. Since the cosine value and sine value for each angle are obtained and integrated with the corresponding candidate coordinate points, the rotation angle can be easily obtained by tangent, and the orientation displacement angle of the input image with respect to the registered image can be rotated. It can be easily obtained as an angle.

<Second Embodiment>
Next, an object recognition apparatus according to a second embodiment of the present invention will be described. The object identification device according to the second embodiment has the same configuration as that of the first embodiment, and performs the same object recognition process. Hereinafter, only processing points different from the first embodiment will be described.
The object recognition apparatus according to the second embodiment is different from the first embodiment in that recognition is performed in consideration of scaling of an input image with respect to a registered image. Thereby, the object recognition apparatus of the second embodiment enables detection even when the input image is enlarged or reduced with respect to the registered image.

FIG. 7 is a diagram for explaining processing for performing object recognition in consideration of scaling of a registered image of an input image.
FIG. 7 (a) shows that the candidate coordinate points change in a displacement vector length R1 described in the first feature table in a reduced region of 90% or more and less than 100%, and an enlarged region in the range of 100% to 110%. Let Then, every time the rotation angle φ is changed by the unit angle from the angle θ of the displacement vector stored in the first feature point table in advance, the candidate coordinate point calculation unit 34 reduces and enlarges the reduced region and the enlarged region in the direction of the displacement vector. The coordinate points obtained in step 1 are set as candidate coordinate points. Here, in FIG. 7A, two annular regions having a predetermined width centered on the second feature point P1 are shown, but the inner annular region is the length R of the displacement vector. Is a set of candidate coordinate points in the reduced area with 90% to 100%, and the outer annular area is a set of candidate coordinate points in the enlarged area with the displacement vector length R from 100% to 110%.

Further, the candidate coordinate point voting unit 35 obtains the cosine value Cos and the sine value Sin, which are voting values calculated for each unit angle change of the rotation angle φ, in the reduced region and the enlarged region obtained by the candidate coordinate point calculation unit 34. Integration processing is performed on all candidate coordinate points.
FIG. 7B shows the position of the candidate coordinate point in the first feature table as in FIG. 7A when the input image is reduced to 95% of the registered image and input. It is a figure which shows the result which the candidate coordinate point calculating part 34 computed in the reduction | restoration area | region where the length R1 of the displacement vector described is 90 to 100%, and the expansion area | region of the range of 100% to 110%. Here, in FIG. 7B, two annular regions having a predetermined width centered on the second feature point P1 are shown, but the inner annular region is the length R of the displacement vector. Is a set of candidate coordinate points in the reduced area with 90% to 100%, and the outer annular area is a set of candidate coordinate points in the enlarged area with the displacement vector length R from 100% to 110%. In FIG. 7B, the length R of the displacement vector is the same as that of the displacement vector described in the first feature table, and the enlargement area and the reduction area are calculated in the same manner as in FIG. On the other hand, since the input image is reduced to 95%, an area where the enlarged area and the reduced area overlap is generated.

Next, a recognition process between the input image of the recognition target object and the registered image in the second embodiment will be described with reference to FIG. FIG. 8 is a flowchart for explaining an operation example of the recognition process.
The input image acquisition unit 31 inputs an input image from an imaging device (not shown), and outputs the input image to the feature point extraction unit 32 and the candidate coordinate point calculation unit 34 as image data expressed by pixels, for example.
When the input image is input, the feature point extraction unit 32 extracts the second feature point of the recognition target object in the input image, and outputs the extracted input image to the feature point binarization unit 33.
Next, the feature point binarization unit 33 binarizes the gradation of the pixel of the second feature point and the other pixels for each pixel of the input image to be input, and stores it internally. (Step S71). The binarization performed here is the same as the processing described in the flowchart of FIG. 13 of the conventional example.

Then, the candidate coordinate point voting unit 35 initializes the storage table corresponding to each coordinate point in each of the Cos plane and the Sin plane, that is, sets the integrated value in the coordinate value of each coordinate point to “0” (step S72). .
Next, the candidate coordinate point calculation unit 34 determines whether or not voting processing as an evaluation value (voting value) at the second feature point has been performed on the binarized data of all the pixels of the input image. If the second feature point is voted on the binarized data of all the pixels of the input image (step S73), the recognition process is terminated, while the pixels that have not been voted on the feature point If it exists, the process proceeds to step S74.
Here, detection of pixels for which evaluation values at feature points are not calculated is the same as in the first embodiment, and thus description thereof is omitted.

Next, the candidate coordinate point calculation unit 34 reads the unprocessed pixel binarization data from the feature point binarization unit 33 (step S74), and the read binarization data exceeds a preset threshold value. Then, it is determined whether or not this pixel is the second feature point (step S75).
At this time, if the read pixel is the second feature point, the candidate coordinate point calculation unit 34 advances the process to step S76. If the pixel is not the second feature point, the candidate coordinate point calculation unit 34 advances the process to step S73.
When the pixel is the second feature point, the candidate coordinate point calculation unit 34 calculates and votes the candidate coordinate points of all the first feature points of the registered image in the coordinates of the second feature point of the input image. It is determined whether or not a value integration process has been performed (step S76). If there is no first feature point that has not been processed, the process proceeds to step S73, and the process has not been performed. If one feature point remains, the displacement vector corresponding to the remaining coordinate point is read from the feature point data storage unit 2, the rotation angle φ is initialized to “0”, and the process proceeds to step S77.

  Next, the candidate coordinate point calculation unit 34 rotates the displacement vector by a unit angle (for example, 1 degree) and changes the rotation angle φ from 0 to 360 degrees. It is determined whether or not all candidate coordinate points on the circle to be generated have been generated (step S77), and the displacement at the first feature point for all the rotation angles φ from 0 to 360 degrees per unit angle. If the vector is rotated, the process proceeds to step S76. On the other hand, if the rotation angle φ has not changed from 0 to 360 and the displacement vector has not completely rotated, the rotation angle of the displacement vector is the unit angle. Rotate and proceed to step S78.

The candidate coordinate point voting unit 35 calculates a cosine value Cos and a sine value Sin, which are voting values, from the unit vector of the displacement vector at the first feature point of the registered image, which is model data, and the rotation angle φ (9). ) And (10) (step S78).
The calculation of the voting value in step S78 is the same as in the first embodiment, and is the same as the calculation of the voting value by the candidate coordinate point calculation unit 34 described with reference to FIGS. 6 (a) and 6 (b). Processing is performed.

Next, when calculating the candidate coordinate value, the candidate coordinate point calculation unit 34 determines whether or not the length R of the displacement vector is measured within a predetermined scaling range (step S79).
At this time, the candidate coordinate point calculation unit 34 determines whether or not a later-described coefficient α in the scaling range exceeds the coefficient range, and multiplies the coefficients in the range by the length R of the displacement vector to obtain a candidate. When the point coordinates are calculated, the coefficient α is returned to the initial value and the process proceeds to step S77. On the other hand, when the candidate point coordinates are not calculated by multiplying all the coefficients of the range by the length R of the displacement vector, the coefficient α Is increased by a preset numerical value, and the process proceeds to step S80.

For example, when the input image is the size of the registered image, the candidate coordinate point calculation unit 34 has a scaling range of +/− 10%, that is, the length of the displacement vector from 0.9R to 1.1R, and the rotation angle. A coordinate point obtained from φ and the coordinate value of the second feature point using the equations (7) and (8) is set as a candidate coordinate point. At this time, if the range of the coefficient α is set as 0.9 ≦ α ≦ 1.1, the initial value of the coefficient α is set to 0.9, and the coefficient α depends on whether the coefficient α exceeds 1.1. It is determined whether or not the range has been exceeded.
At this time, when the coefficient α to be multiplied by the displacement vector length R is changed in increments of 0.05 as the scaling range coefficient α (0.9 ≦ α ≦ 1.1), the rotation angle is used as a candidate coordinate point. In the direction of the displacement vector at φ, 20 candidate coordinate points are generated in increments of 0.05R in the range from 0.9R to 1.1R.

  Next, the candidate coordinate point calculation unit 34 calculates the cosine value Cos and the sine value Sin obtained by the candidate coordinate point calculation unit 34 from the displacement vector at the first feature point of the registered image, which is model data, and the rotation angle φ. The coordinate values of the cos plane and the sin plane for voting are calculated from each of the expressions (7) and (8) (step S80). At this time, the candidate coordinate point calculation unit 34 calculates the coordinate value of the candidate point coordinates using α · R obtained by multiplying the length R of the displacement vector by the coefficient α in step S79.

The candidate coordinate point voting unit 35 adds the cosine value, which is the vote value, to the coordinate value of the storage table corresponding to the candidate coordinate point on the Cos surface obtained by the candidate coordinate point calculation unit 34.
Further, the candidate coordinate point voting unit 35 integrates the sine value, which is the vote value, with respect to the coordinate values of the storage table corresponding to all the candidate coordinate points of the Sin plane obtained by the candidate coordinate point calculation unit 34 (step S81).
At this time, in all of the Cos plane and the Sin plane, the candidate coordinate point voting unit 35 sequentially increases the coefficient α in increments set in advance within the range of the coefficient α obtained by the candidate coordinate point calculation unit 34. Are integrated with respect to all candidate coordinate points obtained using the equations (7) and (9).

  With the configuration described above, according to the object recognition device of the present embodiment, in addition to the effects of the first embodiment described above, a preset magnification can be used even when the input image is scaled with respect to the registered image. If the range is expanded and contracted, robust object detection can be easily performed, and since it is not necessary to prepare a memory corresponding to scaling, the memory for voting can be reduced.

<Third Embodiment>
Next, an object recognition apparatus according to a third embodiment of the present invention will be described. The object identification device according to the third embodiment has the same configuration as that of the first embodiment, and performs the same object recognition process. Hereinafter, only processing points different from the first embodiment will be described.
The difference between the third embodiment and the first embodiment is that the candidate coordinate point voting unit 35 uses the unit vector of the displacement vector and the rotation angle φ to obtain from the equations (9) and (10). Instead of voting each of the cosine value Cos and the sine value Sin on the Cos plane and the Sin plane, the results obtained from the following formulas (11) and (12) are used as voting values.
if (Rnd ()> Abs (Cos)) then C = 0 elsC = Sign (Cos) (11)
if (Rnd ()> Abs (Sin)) then S = 0 elsS = Sign (Sin) (12)

Rnd () in the equations (11) and (12) is a function that generates a random number from 0 to 1. Abs (X) is a function for calculating the absolute value of X. Sign (Y) is a function for obtaining the sign value of Y, and returns 1 if Y is a positive numerical value, 0 if Y is 0, and -1 if Y is negative.
Therefore, the candidate coordinate point voting unit 35 uses the expression (11), and when the random number generated by the function Rnd () is larger than the cosine value Cos, the candidate coordinate point voting unit 35 outputs the voting value C as 0 and the function Rnd () is generated. When the random number is smaller than the cosine value Cos, the sign value of the cosine value Cos obtained by the function Sign (Cos) is output as the vote value C.
Similarly, the candidate coordinate point voting unit 35 uses the expression (12), and when the random number generated by the function Rnd () is larger than the sine value Sin, the candidate coordinate point voting unit 35 outputs the vote value S as 0 and the function Rnd () is generated. When the random number is smaller than the sine value Sin, the sign value of the sine value Sin obtained by the function Sign (Sin) is output as the vote value S.
At this time, a large number of cosine or sine value votes having the same angle are concentrated on the candidate point coordinates serving as the reference point of the input image, and even if a stochastic vote is performed, the final integrated value is the first. In addition, the cosine value Cos and the sine value Sin obtained in the second embodiment are asymptotic. On the other hand, at the candidate coordinate points other than the reference point of the input image, in principle, the cosine value and sine value of random angles are integrated, and asymptotically approaches “0”.

That is, instead of the floating point or fixed point in the first embodiment, the vote value is a numerical value obtained from the probability. As the voting value, +1, 0, and −1 types obtained from the sign values of the cosine value Cos and the sine value Sin are used. As described above, whether or not the vote values are accumulated is determined depending on whether or not it is larger than the random number.
The voting by the types of +1, 0, and -1 makes it possible to reduce the intermediate memory value of voting. When voting processing is performed in a distributed manner by a plurality of personal computers (hereinafter referred to as PCs), each PC performs processing. The amount of data to be reduced can be reduced. Therefore, the intermediate value of voting in each PC can be reduced, and when performing aggregation in any PC, the data capacity to be transmitted / received is reduced, so that the object recognition process can be speeded up.

Next, a recognition process between the input image of the recognition target object and the registered image in the third embodiment will be described with reference to FIG. FIG. 9 is a flowchart for explaining an operation example of the recognition process.
In FIG. 9, steps S91 to S98 are the same as steps S71 to S78 in the flowchart of FIG. 8 according to the second embodiment, and a description thereof will be omitted.
After obtaining the cosine value Cos and the sine value Sin in step S98, the candidate coordinate point voting unit 35 obtains the vote value C from the calculated cosine value Cos using the equation (11), and similarly calculates the calculated sine value Sin. To (12) to obtain a vote value S (step S99).

Next, the candidate coordinate point calculation unit 34 calculates the cosine value Cos and the sine value Sin obtained by the candidate coordinate point calculation unit 34 from the displacement vector at the first feature point of the registered image, which is model data, and the rotation angle φ. The coordinate values of the candidate coordinate points on the Cos plane and the Sin plane for voting are calculated from each of the expressions (7) and (8) (step S100).
The candidate coordinate point voting unit 35 adds the cosine value, which is the vote value, to the coordinate value of the storage table corresponding to the candidate coordinate point on the Cos surface obtained by the candidate coordinate point calculation unit 34.
Further, the candidate coordinate point voting unit 35 integrates the sine value, which is the vote value, with respect to the coordinate values of the storage table corresponding to all the candidate coordinate points of the Sin plane obtained by the candidate coordinate point calculation unit 34 (step S101).

  With the configuration described above, according to the object recognition device of the present embodiment, in addition to the effects of the first embodiment, when distributed processing is performed using a plurality of personal computers, an intermediate period during which vote values are accumulated It is possible to reduce the amount of memory for holding the voting results and reduce the amount of communication when counting the voting results between the personal computers.

<Fourth Embodiment>
Next, an object recognition apparatus according to a fourth embodiment of the present invention will be described. The object identification device according to the fourth embodiment has the same configuration as that of the second embodiment, and performs the same object recognition process. Only processing points different from those of the second embodiment will be described below.
In the flowchart of FIG. 8, the fourth embodiment does not vote the cosine value Cos and the sine value Sin obtained in step S78 as they are, but, similarly to the third embodiment, the vote value C and the vote value S. Are obtained by the equations (11) and (12), and the candidate coordinate points are voted.

Therefore, the candidate coordinate point calculation unit 34 considers scaling of the input image with respect to the registered image and sequentially multiplies the displacement vector R by a coefficient α increased by a step size within a preset range. A plurality of candidate coordinate points are obtained in the direction of the displacement vector of the angle obtained by adding the rotation angle φ to the angle θ.
Then, the candidate coordinate point voting unit 35 adds the cosine value that is the vote value to the coordinate value of the storage table corresponding to the candidate coordinate point of the cos surface obtained by the candidate coordinate point calculating unit 34.
Further, the candidate coordinate point voting unit 35 accumulates the sine value as the vote value with respect to the coordinate values of the storage table corresponding to all the candidate coordinate points on the sine plane obtained by the candidate coordinate point calculating unit 34.
At this time, in all of the Cos plane and the Sin plane, the candidate coordinate point voting unit 35 sequentially increases the coefficient α in increments set in advance within the range of the coefficient α obtained by the candidate coordinate point calculation unit 34. Are integrated with respect to all candidate coordinate points obtained using the equations (7) and (9).

  Further, the program for realizing the function of the object recognition apparatus in FIG. 1 is recorded on a computer-readable recording medium, and the program recorded on the recording medium is read into a computer system and executed to execute object recognition. Processing may be performed. Here, the “computer system” includes an OS and hardware such as peripheral devices.

Further, the “computer system” includes a homepage providing environment (or display environment) if a WWW system is used.
The “computer-readable recording medium” refers to a storage device such as a flexible medium, a magneto-optical disk, a portable medium such as a ROM and a CD-ROM, and a hard disk incorporated in a computer system. Furthermore, the “computer-readable recording medium” dynamically holds a program for a short time like a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. In this case, a volatile memory in a computer system serving as a server or a client in that case, and a program that holds a program for a certain period of time are also included. The program may be a program for realizing a part of the functions described above, and may be a program capable of realizing the functions described above in combination with a program already recorded in a computer system.

  The embodiment of the present invention has been described in detail with reference to the drawings. However, the specific configuration is not limited to this embodiment, and includes design and the like within a scope not departing from the gist of the present invention.

  DESCRIPTION OF SYMBOLS 1 ... Registered image data generation part 2 ... Feature point data storage part 3 ... Object recognition part 11 ... Registered image acquisition part 12 ... Feature point extraction part 13 ... Feature point binarization part 31 ... Input image acquisition part 32 ... Feature point extraction Unit 33 ... Feature point binarization unit 34 ... Candidate coordinate point calculation unit 35 ... Candidate coordinate point voting unit 36 ... Recognition target object determination unit

Claims (7)

A feature point data storage unit for storing feature point data including a plurality of first feature points extracted from a registered image and a displacement vector from the first feature point to a reference point of the registered image;
A feature point extraction unit for extracting a second feature point from the input image;
For each of the second feature points, each of the displacement vectors of all the first feature points is sequentially rotated by a preset unit angle with the second feature point as a rotation center, and the rotation angle is rotated. A candidate coordinate point calculation unit that calculates a candidate coordinate point that is a candidate for a reference point from the coordinates of the second feature point and the displacement vector;
A candidate coordinate point voting unit that individually accumulates voting values based on cosine values and sine values corresponding to the rotation angles of the displacement vectors, for each candidate coordinate point;
The candidate coordinate point having the maximum numerical value obtained by adding the square value of the integrated value of the cosine value and the sine value is detected from the candidate coordinate points, and the detected candidate coordinate point is used as the coordinate of the reference point. An object recognition apparatus comprising: a recognition target object determination unit that detects a rotation angle corresponding to a coordinate point as a rotation angle of the input image with respect to the registered image.   The rotation angle of the displacement vector used as a voting value using the x-direction component and the y-direction component of the unit vector of the displacement vector in the two-dimensional coordinate system as the cosine value and sine value of the displacement vector in the registered image The object recognition apparatus according to claim 1, wherein a cosine value and a sine value corresponding to are calculated. The candidate coordinate point voting unit,
When the angle of the displacement vector in the registered image is the first angle and the rotation angle for rotating the displacement vector by unit angle is the second angle,
3. A cosine value and a sine value when the displacement vector is rotated from a first angle and a second angle that increases for each unit angle are calculated and stored in advance. The object recognition apparatus described in 1. The candidate coordinate point calculation unit
The length of the displacement vector is expanded and contracted within a preset magnification range, and the candidate coordinate point in the longest length and shortest length range is taken as a coordinate point range, and the displacement in the registered image in the coordinate point range The coordinate point corresponding to the length of the vector is set as a candidate point coordinate represented by the coordinate point range, and all voting values for the candidate point coordinate values in the coordinate point range are integrated. Item 4. The object recognition device according to any one of Items 3 to 3. The candidate coordinate point voting section
A random number is generated, the absolute value of each cosine value and sine value is compared with the random number, the voting value is defined as 0 when the random number is larger, and the cosine value is positive when the random number is smaller. In this case, the voting value corresponding to the cosine value is defined as 1, the voting value corresponding to the sine value is defined as 1 when the sine value is positive, and the cosine value corresponds to the case when the cosine value is negative. The voting value to be defined is defined as -1, and when the sine value is negative, the voting value corresponding to the sine value is defined as -1. Object recognition device. Feature points in which feature point data including a plurality of first feature points extracted from the registered image by the registered image data generation unit and displacement vectors from the first feature points to the reference points of the registered images are stored. Data storage process;
A feature point extraction process in which a feature point extraction unit extracts a second feature point from the input image;
The candidate coordinate point calculation unit sequentially rotates each of the displacement vectors of all the first feature points by a preset unit angle with the second feature point as the rotation center for each of the second feature points. A candidate coordinate point calculation process for calculating a candidate coordinate point that is a candidate for a reference point from the coordinates of the second feature point and the displacement vector at each rotated rotation angle;
A candidate coordinate point voting process in which the candidate coordinate point voting unit individually accumulates voting values based on a cosine value and a sine value corresponding to the rotation angle of the displacement vector for each candidate coordinate point;
The target object determination unit detects the candidate coordinate point having the maximum numerical value obtained by adding the square value of the integrated value of the cosine value and the sine value from the candidate coordinate points, and uses the detected candidate coordinate point as a reference And a recognition target object determination step of detecting a rotation angle corresponding to the coordinate point as a rotation angle of the input image with respect to the registered image. Feature points in which feature point data including a plurality of first feature points extracted from the registered image by the registered image data generation unit and displacement vectors from the first feature points to the reference points of the registered images are stored. Data storage processing;
A feature point extraction unit that extracts a second feature point from the input image;
The candidate coordinate point calculation unit sequentially rotates each of the displacement vectors of all the first feature points by a preset unit angle with the second feature point as the rotation center for each of the second feature points. A candidate coordinate point calculation process for calculating a candidate coordinate point that is a candidate for a reference point from the coordinates of the second feature point and the displacement vector at each rotated rotation angle;
Candidate coordinate point voting process in which the candidate coordinate point voting unit individually accumulates voting values based on the cosine value and sine value corresponding to the rotation angle of the displacement vector for each candidate coordinate point;
The target object determination unit detects the candidate coordinate point having the maximum numerical value obtained by adding the square value of the integrated value of the cosine value and the sine value from the candidate coordinate points, and uses the detected candidate coordinate point as a reference A program for causing a computer to execute recognition target object determination processing for detecting a rotation angle corresponding to the coordinate point as a rotation angle of the input image with respect to the registered image.