JP6086491B2 - Image processing apparatus and database construction apparatus thereof - Google Patents

Description

  The present invention relates to an image processing apparatus that identifies a recognition target from an input image acquired by an imaging unit such as a camera, and estimates a relative position and orientation relationship between the recognition target and the camera, and a database construction apparatus thereof.

  In recent years, AR (augmented reality) technology for processing a real-space image by a computer and adding further information has been realized in an information terminal equipped with a camera module such as a mobile phone. In these AR technologies, even in a terminal with few processing resources, recognition objects in an input image acquired by an imaging means such as a camera are identified, and the relationship between their positions and postures (hereinafter referred to as posture parameters) Need to be estimated in real time.

  Patent Document 1 discloses a technique for estimating a posture parameter with respect to a reference marker in an input image by image recognition processing. Patent Document 2 discloses a technique for identifying an arbitrary object in an input image in a terminal with few processing resources by image recognition processing using feature point matching between a server and an information terminal. Patent Document 3 discloses a technique for estimating a posture parameter for an arbitrary object in an input image in a terminal with less processing resources by a hybrid type image recognition process combining initialization and tracking.

JP 2012-212460 A JP 2012-203669 A Special table 2013-508844 gazette

M. Ozuysal, M. Calonder, V. Lepetit, and P. Fua, "Fast keypoint recognition using random ferns," Pattern Analysis and Machine Intelligence, IEEE Transactions on, vol. 32, no. 3, pp. 448 -461, march 2010.

  In the prior art described in Patent Document 1, there is a problem that an arbitrary object such as a print with a complicated design cannot be handled as a recognition target object.

  In the prior art described in Patent Document 2, there is a problem of a decrease in identification speed due to communication with a server and a lack of robustness with respect to a change in the distance and direction of photographing an object.

  In the prior art described in Patent Document 3, there is a problem of lack of robustness with respect to changes in the distance and direction in which an object is imaged at the time of initialization.

  In Patent Documents 2 and 3, it is assumed that a technique called Random Ferns described in Non-Patent Document 1 that can be executed by a terminal with a small processing resource is used. In this case, offline processing required before execution is used. The problem was that it took a long time and the database size required for execution was large.

  On the other hand, when local features such as SURF, ORB, BRISK, etc. that do not require time for offline processing are used, there is a lack of robustness against changes in the processing load at the time of execution and the distance and direction of shooting the object. There was a problem of becoming more prominent.

  An object of the present invention has been made in view of the above-described problems of the prior art, and is an image processing apparatus capable of improving robustness against changes in shooting distance and direction without increasing processing load and further reducing offline processing time. And providing a database construction apparatus thereof.

  (1) The image processing apparatus of the present invention includes representative image generation means for generating a representative image specific to each viewpoint by sampling the recognition target from a plurality of viewpoints, and feature points and local features thereof from the recognition target camera image. Corresponding feature point detecting means for detecting the amount, corresponding point candidate detecting means for detecting the corresponding point candidate and its local feature amount from each representative image, and the local feature amount of the corresponding point candidate and the back-projected coordinates on the recognition target A database to be managed, and a recognition means for recognizing corresponding points based on a matching result between local feature amounts of each feature point detected from the camera image and local feature amounts of each corresponding point candidate managed in the database; And posture parameter calculation means for estimating a posture parameter when the camera image is taken based on the recognition result.

  (2) The image processing apparatus according to the present invention associates the local feature amount of each corresponding point candidate detected from the representative image generated by sampling the recognition target from a plurality of viewpoints and the back-projected coordinates on the recognition target. A database to be managed, feature point detection means for acquiring feature points and their local feature amounts from camera images to be recognized, local feature amounts of feature points detected from camera images, and correspondences managed in the database A recognition unit for recognizing a corresponding point based on a matching result with a local feature amount of a point candidate and a posture parameter calculation unit for estimating a posture parameter when the camera image is taken based on the recognition result are provided.

  (3) The database construction device of the present invention includes representative image generation means for generating a representative image specific to each viewpoint by sampling a recognition target from a plurality of viewpoints, and corresponding point candidates and their local feature amounts from each representative image. And a database for associating the local feature amount and the back-projected coordinates to the recognition target for each corresponding point candidate.

  (4) The representative image generation means determines the density of image sampling for each parameter such as distance, angle, and rotation that changes the viewpoint according to the robustness to the viewpoint change of the algorithm that detects feature points and / or feature point candidates. I decided to decide.

According to the present invention, the following effects are achieved.
(1) The representative image itself with different projection parameters is not stored in the database, but the local feature amount of the corresponding point candidate detected from each representative point and the coordinate value obtained by back projecting the corresponding point candidate to the recognition target In addition, each local feature is handled uniformly regardless of the representative image from which the corresponding point candidate is detected, thereby reducing the database capacity and simplifying the database configuration. The speed is improved.

  (2) The density of the viewpoint arrangement when acquiring the representative image by changing the projection parameter value is determined to compensate for the weakness according to the robustness to each viewpoint change of the feature point detection and corresponding point candidate detection algorithm Is done. Therefore, densely arrange viewpoints for highly robust viewpoint changes and wastefully accumulate corresponding point candidates for a large number of representative images that do not contribute to accuracy improvement, or conversely, roughly arrange viewpoints for viewpoint changes with low robustness. The processing load can be reduced and the database capacity can be reduced without degrading the accuracy by performing the process while ensuring sufficient robustness against changes in viewpoint.

1 is a block diagram showing a configuration of an AR system to which the present invention is applied. 3 is a block diagram showing a configuration of a main part of posture parameter estimation apparatus 2. FIG. It is the figure which showed the example of the representative image produced | generated from the recognition object. It is the figure which showed the method of acquiring the coordinate and gradient direction of the recognition target corresponding to each corresponding point candidate by back-projecting each corresponding point candidate of a representative image to a recognition target. It is the figure which showed the setting method of a virtual viewpoint. It is the figure which showed the example which changes the registration content of a feature point database according to the robustness of a feature point detection algorithm. It is the block diagram which showed the structure of the database construction apparatus of an image processing apparatus.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 is a block diagram showing a configuration of an AR system to which the present invention is applied, and is used by being mounted on an information terminal such as a mobile phone, a smartphone, a PDA, or a notebook PC.

  The imaging device 4 is a camera module or WEB camera device mounted on a portable terminal or the like, captures the recognition target 5 and outputs a camera image Ica to the display device 1 and the posture parameter estimation device 2. The recognition target 5 is an arbitrary three-dimensional object whose shape and pattern are known, and includes a two-dimensional object (image) represented by a printed matter or a print.

  The posture parameter estimation device 2 is based on the camera image acquired from the imaging device 4 and features between each corresponding point candidate (feature point) of the recognition target 5 and each feature point of the camera image, as will be described in detail later. Point matching is performed, and the relationship between the relative position and orientation of the recognition target 5 and the AR system or the imaging device 4 is estimated based on the feature points associated with each other and the internal parameters of the camera.

  In general, the relationship between the relative position and orientation is expressed in the form of a matrix called an orientation parameter or a camera external parameter, and includes information on the position and orientation in a three-dimensional space. The appearance of the object on the screen is determined by this attitude parameter, the camera's intrinsic parameters, the focal length inherent to the camera, the matrix containing information on the principal axis position, and other optical distortion parameters. The

  In the present embodiment, it is assumed that internal parameters and distortion parameters are acquired in advance by calibration or the like and distortion is removed, and the posture parameter estimation results are provided to the display device 1. When the AR system targets a plurality of types of objects as recognition targets, a set of posture parameters and target object IDs are provided to the display device 1 by the number of recognitions.

  The additional information database 3 is a storage device configured by a hard disk drive, a semiconductor memory module, or the like. When the AR system recognizes the position of the recognition target 5, the additional information database 3 is displayed on the display device 1 in a superimposed manner on the recognition target 5. A two-dimensional image is held, and additional information regarding the recognition target 5 corresponding to the camera posture parameter estimated by the posture parameter estimation device 2 is output to the display device 1.

  The display device 1 is a monitor device that can post a camera image continuously acquired by the imaging device 4 and additional information acquired from the additional information database 3 to the user, and may be a display of a portable terminal. In addition, a form such as a head-mounted display (HMD) may be used. In particular, in the case of a see-through type HMD, it is possible to display only the additional information superimposed on the field of view without displaying a camera image. When the display device 1 is a display, the additional information input from the additional information DB is superimposed on the camera image. At this time, the additional information is displayed with its position and orientation corrected by the posture parameter input from the posture parameter estimation device 2.

  FIG. 2 is a block diagram showing the configuration of the main part of the posture parameter estimation device 2, which includes a feature point detection unit 21, a matching unit 22, and a posture parameter calculation unit 23.

  In the feature point detection unit 21, the feature point detector 201 detects a feature point from a camera image obtained by capturing the recognition target 5. The feature quantity extractor 202 extracts a local feature quantity from each feature point. As the feature point detector 201 and feature quantity extractor 202, algorithms such as Harris, Hessian, SIFT, SURF, FAST, BRIEF, ORB, BRISK, and FREAK can be used.

  In general, these algorithms have advantages and disadvantages. Algorithms such as SIFT and SURF that are robust to shooting distance, angle, and rotation have a large processing load, while algorithms such as FAST and BRIEF that have a low processing load have distance and angle.・ It is not robust against rotation. As a method that can achieve both processing load and robustness, there is an algorithm represented by Random Ferns that pre-learns classifiers for classifying feature points. However, long-term learning and large databases (classifiers) ) Is required.

  In the matching unit 22, the learning unit 203 includes a representative image generation unit 203a, a corresponding point candidate detection unit 203b, and a coordinate association unit 203c.

  The representative image generation unit 203a projects the recognition target 5 by photographing it from a plurality of positions and directions (from a plurality of viewpoints) or using a plurality of projection parameters with different viewpoints (hereinafter referred to as image sampling and expression). In some cases, a representative image is generated. In the present embodiment, a representative image is generated using a plurality of projection parameters. Here, the viewpoint includes a distance between the recognition target 5 and the imaging device 4, an angle of the imaging device 4 with respect to the recognition target 5, a rotation around the optical axis of the imaging device 4 with respect to the recognition target 5, and the like. The representative image generation unit 203a converts a plurality of different viewpoints into projection parameters. By projecting the recognition target 5 in two dimensions using the projection parameters, a plurality of images having different distances, angles, and rotations are generated as representative images.

  FIG. 3 is a diagram showing an example of a representative image generated from the recognition target 5. Here, a representative image is represented by using projection parameters converted from combinations of a plurality of types of angles (horizontal axis) and distances (vertical axis). An image has been generated.

  The representative image can be acquired by photographing the recognition target 5 from a plurality of viewpoints. However, from the viewpoint of simplicity of offline processing, if the recognition target 5 is a planar object, the reference image is subjected to projective transformation ( It is desirable to use a method of artificially creating a 3D model by projecting a 3D model if the conversion parameter is calculated from a projection parameter) or if it is a complex 3D object.

  The corresponding point candidate detection unit 203b detects a feature point from each of the representative images by using a method similar to the method in the feature point detection unit 21, and extracts a local feature amount of each corresponding point candidate. .

  In order to increase the accuracy of corresponding point candidates, a method for associating only local feature amounts whose distance is equal to or less than a preset threshold, a method for excluding corresponding point candidates that deviate from the overall tendency of corresponding points, and the like are known. These methods are also applicable to the matching unit 22 of the present invention.

  In the present embodiment, it is desirable to use the same algorithm as that of the feature point detection unit 21 as the corresponding point candidate detection unit 203b of the matching unit 22, but it is not necessary to match even fine parameters such as resolution and threshold value, rather processing. From the viewpoint of load, the feature point detection unit 21 that executes matching for each camera image sets a parameter with priority on execution speed, and performs corresponding matching only at the beginning of execution or during offline learning. Then, it is desirable to set the parameter setting with priority on accuracy.

  As shown in an example in FIG. 4, the coordinate association unit 203 c backprojects each corresponding point candidate detected from each representative image onto the recognition target 5, so that the recognition target 5 corresponding to each corresponding point candidate is displayed. A coordinate value (three-dimensional coordinate) in the object coordinate system is acquired.

  In the feature point database 204, the local feature amount of each corresponding point candidate detected from each representative image and the coordinate value after backprojection acquired by the coordinate correspondence unit 203c are associated with each other and each corresponding point candidate is stored. Remembered. Each corresponding point candidate includes a local feature amount detected from a different image (representative image), but in the subsequent processing, all the corresponding point candidates are treated as a corresponding point candidate detected from the recognition target 5.

  The recognizing unit 205 performs corresponding point matching between the local feature amount of each feature point detected from the camera image of the recognition target 5 after learning and the local feature amount of each corresponding point candidate stored in the database 204. . In this embodiment, local feature quantities of each feature point and each corresponding point candidate are expressed in a vector format, and a pair of a feature point and a corresponding point candidate having the closest Euclidean distance or Hamming distance is associated with each other. It is considered as a corresponding point.

  Here, there is a possibility that a pair of corresponding feature points and corresponding point candidates (corresponding points) includes a partial error, and the corresponding points are input before the posture parameter calculation unit 23 is input. By performing the selection process, the reliability of corresponding points can be improved.

  As an example of this selection processing, for example, when the recognition target is a planar object, the feature points constituting each corresponding point are compared with the gradient directions of the corresponding point candidates, and the corresponding points having a unique difference in the gradient direction of each corresponding point are deleted. Process to do. Here, when the recognition target is a planar object, the difference between the gradient directions is expected to be approximately the same.

  Here, when the selection process using the gradient direction described above is performed, when the coordinate association unit 203c performs back projection, the coordinate values (three-dimensional coordinates) in the object coordinate system of the recognition target 5 corresponding to each corresponding point candidate. ) As well as the gradient direction in the object coordinate system of the recognition target 5 corresponding to each corresponding point candidate must be simultaneously acquired and stored in the database 204.

  The posture parameter calculation unit 23 estimates a posture parameter based on the correspondence between the feature point and the corresponding point candidate and the internal parameters of the camera. Conventionally, a method for estimating a posture parameter that explains the relationship from a match between a two-dimensional coordinate and a three-dimensional coordinate has been studied. The relationship between a three-dimensional coordinate and a two-dimensional coordinate is generally expressed by the following formula ( It is represented by 1).

  Here, [u, v] and [X, Y, Z] represent a two-dimensional pixel coordinate value and a three-dimensional coordinate value, respectively, and [•] ^ T represents a transposed matrix. A and W represent camera internal parameters and posture parameters, respectively. The internal parameters of the camera are obtained in advance by camera calibration.

  Attitude parameter W = [R, t] = [r1, r2, r3, t], which is represented by a rotation matrix R and a translation vector t. The posture parameter W can be estimated using the match between the three-dimensional coordinates [X, Y, Z, 1] ^ T and the two-dimensional coordinates [u, v, 1] ^ T and the internal parameters of the camera.

  Here, since the corresponding two-dimensional coordinates and the corresponding points of the three-dimensional coordinates include some erroneous corresponding points (pairs), the corresponding points input by the sampling method represented by RANSAC and PROSAC are used. It is common to extract only correct corresponding points (inliers) and estimate posture parameters that minimize the reprojection error of three-dimensional coordinates by an iterative method such as the Levenberg-Marquard method.

  Next, a method for generating a representative image by the representative image generating unit 203a will be described in more detail. In the present invention, in order to generate a large number of representative images so that the viewpoints to be projected are not biased, virtual viewpoints are arranged evenly around the recognition target 5 and images observed from each viewpoint are generated. . More specifically, a projection image is calculated from each viewpoint, and a representative image is generated by projecting the recognition target 5 or its three-dimensional model in two dimensions.

  In this embodiment, as illustrated in FIG. 5, a polygonal vertex that approximates a hemisphere that covers the periphery of the recognition target 5 is treated as a virtual viewpoint. In this case, the rotation matrix R can be obtained from the viewpoint by calculating the following equations (2), (3), and (4).

  However,

  Here, P represents the three-dimensional coordinates of the virtual viewpoint, and rod (•) represents the Rodrigues transformation. Further, an arbitrary translation parameter t is added to each rotation matrix, and a projection parameter P = A [R | t] is created together with the internal parameter A of the camera. Finally, the representative image generation unit 203a generates the representative image (FIG. 3) by projecting the recognition target 5 using the projection parameters.

  Note that the representative image generation unit 203a may generate each representative image without creating a bias in the viewpoint, but in a situation where the likelihood of shooting from a specific viewpoint is high, The range may be limited. In general, if the recognition target 5 is a planar object, it is assumed that the likelihood of photographing from an angle close to the front is high. In that case, the representative image generation unit 203a can generate a representative image using only a virtual viewpoint close to the front, thereby realizing reduction in learning time and improvement in reliability.

  In addition, when the shooting distance is assumed to be limited within a certain range, for example, on the road surface, the number of representative images can be reduced by narrowing the range of the translation parameter, thereby reducing learning time and reliability. It is possible to improve the performance. On the other hand, when the recognition target 5 has a three-dimensional shape and may be photographed from the entire periphery, it is desirable to arrange the virtual viewpoints in the entire periphery.

  By the way, when the feature point detection algorithm employed by the feature point detection unit 21 and the corresponding point candidate detection unit 203b has, for example, high robustness against a change in angle, the viewpoint interval is set close to the shooting angle. Increasing the number of representative images does not contribute much to the improvement of matching accuracy.

  Therefore, if an algorithm that is excellent in robustness against any change in distance, angle, and rotation (hereinafter may be collectively referred to as “a parameter that changes the viewpoint”), such as the SIFT algorithm, is used. If the interval is rough with respect to any of distance, angle, and rotation, it is possible to reduce the offline processing load and the database capacity while ensuring sufficient robustness against changes in the viewpoint when photographing the recognition target 5. become.

  On the other hand, if an algorithm with low robustness against any change in distance, angle, and rotation, such as the FAST algorithm or the BRIEF algorithm, is adopted, the viewpoint interval is fine with respect to any distance, angle, or rotation. By doing so, it becomes possible to improve the matching accuracy.

  However, even if the viewpoints are arranged more densely than necessary, the matching point candidate detection unit 203b of the present invention does not change the viewpoints (distance change, angle change, rotation change, etc.). ) To generate the minimum necessary number of representative images to compensate for robustness. For example, in the case of the BRIEF algorithm, matching can be performed without any problem if the rotation is about 0 to 10 degrees. Therefore, with respect to the rotation, the viewpoint interval (resolution of image sampling) is set to 20 degrees (18 patterns) or By setting the angle to 30 degrees (12 patterns), it is possible to reduce the off-line processing load and the database capacity while ensuring sufficient robustness against changes in rotation when the recognition target 5 is photographed.

  In addition, if you adopt an algorithm that is less robust to changes in distance and angle compared to robustness against changes in rotation, such as the ORB algorithm and BRISK algorithm, the viewpoint interval will be coarse with respect to rotation, If the distance and angle are made dense, the processing load can be reduced and the database capacity can be reduced while securing sufficient robustness against changes in the viewpoint when the recognition target 5 is photographed.

  In general, there are many algorithms with low processing load that are excellent only in robustness against changes in rotation, and if robustness against changes in distance and angle is added, there is a tendency to increase the processing load. Therefore, the ORB algorithm and BRISK algorithm, which are not robust against changes in distance and angle, are used to generate representative images to compensate for the robustness against changes in distance and angle. It is desirable to achieve both robustness against

  As described above, the corresponding point candidate detection unit 203b according to the present invention changes each viewpoint of the algorithm employed by the feature point detection unit 21 and the corresponding point candidate detection unit 203b (change in distance, change in angle, change in rotation, etc.). According to the robustness to the image, the representative image is generated so as to compensate for the weakness of the robustness.

  FIG. 6 shows different tendencies of representative images in which corresponding point candidates are registered in the feature point database 204 according to the robustness to each viewpoint change of the algorithm adopted by the feature point detecting unit 21 and the corresponding point candidate detecting unit 203b. It is the figure which showed the example, and here shows the comparison with the case where the robustness with respect to the change of the angle is equal and the algorithm with high robustness with respect to the change of the distance is adopted .

  When a recognition algorithm that is highly robust with respect to a change in distance is used, the same matching accuracy can be obtained regardless of the shooting distance, and it is not necessary to prepare a large number of representative images having different shooting distances for each angle. In other words, since the distance between viewpoints can be made coarse with respect to distance, the number of representative images in which feature points are registered can be reduced.

  On the other hand, when an algorithm having low robustness with respect to a change in the shooting distance is mounted, matching accuracy greatly changes depending on the shooting distance, so it is necessary to prepare a large number of representative images having different distances for each angle. In other words, the distance between viewpoints must be close with respect to the distance, and the number of representative images increases.

  Although not shown, when a recognition algorithm that is robust against changes in angle and rotation is used, the same matching accuracy can be obtained regardless of the angle and rotation. There is no need to prepare a large number of different representative images. That is, the viewpoint intervals may be roughly arranged with respect to the angle and rotation.

  On the other hand, if an algorithm with low robustness against changes in angle and rotation is installed, the matching accuracy changes greatly depending on the angle and rotation, so a large number of representative images with different angles and rotation are prepared for each shooting distance. There is a need to. That is, the interval between viewpoints must be set closely with respect to angle and rotation.

  As described above, according to the present embodiment, the density of the viewpoint arrangement when the representative image is acquired by changing the projection parameter value is weak in the robustness to each viewpoint change of the feature point detection unit and the corresponding point candidate detection unit. Is determined to compensate. Therefore, densely arrange viewpoints for highly robust viewpoint changes and wastefully accumulate corresponding point candidates for a large number of representative images that do not contribute to accuracy improvement, or conversely, roughly arrange viewpoints for viewpoint changes with low robustness. The processing load can be reduced and the database capacity can be reduced without degrading the accuracy by performing the process while ensuring sufficient robustness against changes in viewpoint.

  In the above embodiment, the posture parameter estimation apparatus 2 of the AR system is described as including the learning unit 203 and constructing the feature point database 204 in advance, but the present invention is not limited to this. 7, a learning function 10 equivalent to the learning unit 203 is provided outside the posture parameter estimation device, and a database constructed offline by the learning function 10 is transplanted to the learning unit 203. May be.

  In the learning function 10, the representative image generation unit 101 generates a representative image by projecting the recognition target 5 two-dimensionally with various projection parameters. Corresponding point candidate detection unit 102 detects feature points from each representative image as a corresponding point candidate by a method similar to the method in feature point detection unit 21, and further sets the local feature amount of each corresponding point candidate to each representative image. Extract from

  The coordinate correlating unit 103 back-projects each corresponding point candidate detected from each representative image onto the recognition target 5 to thereby obtain a coordinate value (three-dimensional coordinate) in the object coordinate system of the recognition target 5 corresponding to each corresponding point candidate. ) And get the gradient direction. In the database 104, the local feature amount of each corresponding point candidate detected from each representative image, the coordinate value after back projection and the gradient direction are stored in association with each other for each corresponding point candidate. To the feature point database 204.

  DESCRIPTION OF SYMBOLS 1 ... Display apparatus, 2 ... Attitude parameter estimation apparatus, 3 ... Additional information database, 4 ... Imaging apparatus, 5 ... Recognition object, 21 ... Feature point detection part, 22 ... Matching part, 23 ... Attitude parameter calculation part, 203 ... Learning Part 203a ... representative image generation part 203b ... corresponding point candidate detection part 203c ... coordinate correspondence part 204 204 feature point database 205 ... recognition part

Claims (10)

In an image processing apparatus that estimates posture parameters for a recognition target in a captured camera image,
Representative image generation means for sampling a recognition target from a plurality of viewpoints and generating a representative image unique to each viewpoint;
Feature point detection means for detecting a feature point and its local feature amount from a camera image to be recognized; and
A corresponding point candidate detecting means for detecting a corresponding point candidate and its local feature amount from each representative image;
For each of the corresponding point candidates, a database for managing the local feature amount and the backprojected coordinates on the recognition target in association with each other,
Recognizing means for recognizing corresponding points based on matching results between local feature amounts of each feature point detected from the camera image and local feature amounts of each corresponding point candidate managed in the database;
Posture parameter calculation means for estimating a posture parameter at the time of shooting the camera image based on the recognition result ,
The representative image generating means determines an image sampling density for each parameter for changing a viewpoint according to robustness against a viewpoint change of an algorithm for detecting the feature points and / or feature point candidates. Processing equipment.   The image processing apparatus according to claim 1, wherein the representative image generation unit performs image sampling more densely for a parameter having lower robustness.   The image processing apparatus according to claim 1, wherein the representative image generation unit performs image sampling more coarsely for a parameter having higher robustness. In an image processing apparatus that estimates posture parameters for a recognition target in a captured camera image,
For each corresponding point candidate detected from a representative image generated by sampling a recognition target from a plurality of viewpoints, a database for managing the local feature amount and the back-projected coordinates on the recognition target in association with each other,
Feature point detection means for acquiring a feature point and its local feature amount from a camera image to be recognized; and
Recognizing means for recognizing corresponding points based on matching results between local feature amounts of each feature point detected from the camera image and local feature amounts of each corresponding point candidate managed in the database;
Posture parameter calculation means for estimating a posture parameter at the time of shooting the camera image based on the recognition result,
The image processing apparatus according to claim 1, wherein the representative image is image-sampled at a density corresponding to a robustness against a viewpoint change of an algorithm for detecting the feature point for each parameter for changing the viewpoint. Parameter for changing the viewpoint, distance, angle and the image processing apparatus according to any one of 4 to claims 1, characterized in that it comprises at least one rotation. The representative image, the image processing apparatus according to any one of claims 1, characterized in that an image to be observed from virtually arranged viewpoint evenly around the recognized 5. The representative image is a projection image obtained by projecting the recognition target in a two-dimensional image processing apparatus according to any one of claims 1 to 6 parameter for changing the viewpoint, characterized in that a projection parameters. In a database construction device of an image processing device that estimates posture parameters for a recognition target in a captured camera image,
Representative image generation means for sampling a recognition target from a plurality of viewpoints and generating a representative image unique to each viewpoint;
Corresponding point candidate detecting means for acquiring a corresponding point candidate and its local feature amount from each representative image;
A database for correlating the local feature amount and the backprojected coordinates to the recognition target for each corresponding point candidate,
The database construction of the image processing apparatus, wherein the representative image generation means determines the density of image sampling for each parameter for changing the viewpoint according to robustness to the viewpoint change of the algorithm for detecting the corresponding point candidate apparatus. 9. The database construction apparatus for an image processing apparatus according to claim 8 , wherein the representative image generation unit performs image sampling more densely for parameters with lower robustness. 9. The database construction apparatus for an image processing apparatus according to claim 8 , wherein the representative image generation means performs image sampling more roughly as the parameter having higher robustness.