JP2008225785A - Image recognition device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an image recognition device, recognizing the presence/absence of an object with good accuracy. <P>SOLUTION: This image recognition device 1 includes two cameras for imaging the object and a processing unit. The processing unit restores the surface shape of the object as a set of a number of dots on a three-dimensional space according to the pick-up images of the object by the respective cameras, divides the three-dimensional space including these restoring dots into voxels to generate a three-dimensional voxel group, and further slices the three-dimensional voxel group to be converted to a plurality of two-dimensional voxel groups. The processing unit counts restoring dots included in each voxel in the respective two-dimensional voxel groups, determines the probability of existence of the object in each voxel from the count, and recognizes the presence/absence of the object according to the result. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to an image recognition device that recognizes an object using a captured image of the object.

As a conventional image recognition device, for example, as described in Non-Patent Document 1, a surface shape of an object is restored on a three-dimensional space as a set of points, and a point exists in each voxel of the three-dimensional space. Those that detect whether or not are known.
FANG, S. AND CHEN, H.2000.Hardware AcceleratedVoxelization.Computers & Graphics 24,3 (June), 433-442.

  However, the following problems exist in the prior art. That is, when an image recognition device is applied to, for example, a robot and the movement path of the robot is obtained according to the recognition result of the image recognition device, if one point is erroneously restored due to the influence of noise or the like, it originally exists. It is judged that there is an obstacle in the place that should not be done. In this case, a correct movement path of the robot cannot be obtained.

  An object of the present invention is to provide an image recognition apparatus that can accurately recognize the presence or absence of an object.

  The present invention provides an image recognition apparatus for recognizing an object using a captured image of the object acquired by an imaging unit, based on the captured image of the object, a plurality of points corresponding to the surface of the object in a three-dimensional space. 3D restoration means for restoring upward, space division means for dividing a 3D space including a plurality of points restored by the 3D restoration means into a plurality of voxels, and each voxel obtained by dividing by the space division means And existence probability determining means for counting the number of points included in the object and determining the existence probability of the object in each voxel.

  In such an image recognition device, a three-dimensional space including a plurality of restored points is divided into a plurality of voxels, the number of points included in each voxel is counted, and the existence probability of an object in each voxel is determined. To do. At this time, as the number of restoration points included in the voxel increases, the existence probability of the object in the voxel is increased. When the existence probability of the target object in the voxel is lower than the predetermined value, it is determined that the restoration point included in the voxel is erroneously restored due to the influence of noise or the like, and the position corresponding to the voxel is the target. It is considered that the thing does not exist. Thereby, it is possible to accurately recognize the presence or absence of an object by excluding the influence of noise or the like.

  Preferably, the apparatus further comprises means for converting a three-dimensional voxel group composed of a plurality of voxels obtained by dividing by the space dividing means into a plurality of two-dimensional voxel groups, and the existence probability determining means is provided for each two-dimensional voxel group. Count the number of points contained in each voxel. By converting a three-dimensional voxel group into a two-dimensional voxel group in this way, the processing for counting the number of points included in each voxel can be performed at high speed using a GPU or the like suitable for processing a two-dimensional image. It becomes possible.

  Preferably, the space dividing unit divides the three-dimensional space including the restored points into a plurality of voxels so that the size of the voxel increases as the distance from the imaging unit increases. In this case, since the number of voxels decreases as the distance from the imaging unit increases, the processing time of the existence probability determining means can be shortened, and the memory size required for storing the existence probability data can be reduced. In addition, since the position accuracy is low at a restoration point far from the imaging unit, even if the size of the voxel far from the imaging unit is increased, there is almost no problem in the recognition process of the presence / absence of the object.

  According to the present invention, the presence or absence of an object can be accurately recognized. Thereby, for example, when the robot is moved to an arbitrary position using the recognition result of the image recognition apparatus, it is possible to obtain a correct movement path of the robot.

  DESCRIPTION OF EMBODIMENTS Hereinafter, a preferred embodiment of an image recognition apparatus according to the present invention will be described in detail with reference to the drawings.

  FIG. 1 is a schematic configuration diagram showing an embodiment of an image recognition apparatus according to the present invention. In the figure, the image recognition apparatus 1 of this embodiment is mounted on, for example, an industrial robot (not shown).

  In FIG. 1, the image recognition apparatus 1 includes cameras 2 and 3 that image the front of the robot, and an arithmetic processing unit 4 connected to these cameras 2 and 3. The cameras 2 and 3 are, for example, CCD cameras, and are provided on both eyes (not shown) of the robot so as to image the object A from two different viewpoints. When the object A is imaged by the cameras 2 and 3, images of the object A are projected onto the imaging surfaces 2a and 3a of the cameras 2 and 3, respectively.

  The arithmetic processing unit 4 inputs the captured image of the object A acquired by the cameras 2 and 3, performs predetermined image processing, and determines whether an obstacle, a gripping object, or the like actually exists in front of the robot. . The details of the arithmetic processing procedure by the arithmetic processing unit 4 are shown in FIG.

  In the figure, first, for example, by using stereo vision using the principle of triangulation, as shown in FIG. 3A, the surface shape (unevenness) of the object A imaged by the cameras 2 and 3 is displayed at a number of points. Restoration on the three-dimensional space as a set of P (step S51). By associating the same point P on the image planes 2a and 3a of the cameras 2 and 3, the depth can be known by the principle of triangulation.

  Subsequently, as shown in FIG. 3B, a three-dimensional array is obtained by dividing the three-dimensional space including all the points (restoration points) P obtained in step S51 into small spaces (voxels) Q. A three-dimensional voxel group R composed of the plurality of voxels Q is generated (procedure S52). As a result, all the restoration points P are included in any voxel Q.

  Subsequently, as shown in FIG. 3C, by slicing the three-dimensional voxel group R obtained in step S52, a plurality of two-dimensional voxel groups S in which a plurality of voxels Q are arranged in a matrix are generated. (Procedure S53). That is, the three-dimensional voxel group R is converted into a plurality of two-dimensional voxel groups S.

  Subsequently, the number of restoration points P included in each voxel Q in the two-dimensional voxel group S is counted, and the probability (existence probability) that the object A exists in each voxel S is determined according to the count number (procedure) S54). At this time, the existence probability is increased, for example, linearly or exponentially as the number of restoration points Q included in the voxel S increases.

  The two-dimensional voxel group S can be processed as a single two-dimensional image. Therefore, this processing is executed using, for example, a GPU (Graphic Processing Unit) that can process a two-dimensional image at high speed. At this time, the number of restoration points P included in each voxel Q is represented as a gradation value of luminance, for example, as shown in FIG.

  Subsequently, the existence probability data of each voxel Q obtained in step S53 is stored in a memory (not shown) together with the position data of each voxel Q (step S55).

  Subsequently, it is determined whether or not the existence probabilities of the objects A in all the voxels P have been obtained (step S55). If the existence probabilities of the objects A in all the voxels P have not yet been obtained, other two-dimensional voxels are obtained. The above steps S54 and S55 are repeated for the group S. On the other hand, when the existence probabilities of the objects A in all the voxels P are obtained, the existence probability data of each voxel Q stored in the memory (not shown) is read, and the obstacles are within the imaging range of the cameras 2 and 3. Or whether there is a gripping object or the like (step S57). The determination is specifically made as follows.

  That is, at a position where an obstacle or the like actually exists, the number of restoration points P included in the voxel Q corresponding to the position increases. However, in the above step S51, if the correspondence between the same points is not taken well between the cameras 2 and 3, the restoration point P appears as noise or the like in the voxel Q corresponding to the position where no obstacle or the like originally exists. May come. However, the restoration point P due to this noise or the like does not come together.

  Therefore, when the existence probability of the object in the voxel Q is higher than a preset threshold value, the restoration point P included in the voxel Q is a normal one (actual restoration point), and the position corresponding to the voxel Q. It is determined that there are obstacles. On the other hand, when the existence probability of the object in the voxel Q is lower than a preset threshold value, the restoration point P included in the voxel Q is erroneously generated due to the influence of noise or the like. It is determined that there is no obstacle or the like at the corresponding position.

  In the above, the procedure S51 constitutes a three-dimensional restoration unit that restores a plurality of points corresponding to the surface of the object on the three-dimensional space based on the captured image of the object. Step S52 constitutes a space dividing unit that divides a three-dimensional space including a plurality of points restored by the three-dimensional restoration unit into a plurality of voxels. Step S53 constitutes means for converting a three-dimensional voxel group composed of a plurality of voxels obtained by dividing by the space dividing means into a plurality of two-dimensional voxel groups. Step S56 constitutes existence probability determining means for counting the number of points included in each voxel obtained by dividing by the space dividing means and determining the existence probability of the object in each voxel.

  As described above, in the present embodiment, the number of restoration points P included in each voxel Q is counted, and the actual restoration point in each voxel Q, that is, the existence probability of the object A is obtained from the counted number. It is determined whether or not the object A exists at a position corresponding to each voxel Q based on the probability. Accordingly, since the presence of the object A is determined without the influence of noise or the like, the presence or absence of the object A can be accurately recognized.

  Therefore, for example, when a robot moves, it is recognized that there is an obstacle in the course of the robot, but there is actually an obstacle, and a movement path is taken to avoid the position. Can be prevented. Or, when the robot grips some object, it is possible to prevent the malfunction that it is recognized that there is actually a gripping object and there is no gripping object and operates to grip it. Become.

  In addition, since the two-dimensional voxel group S is generated by slicing the three-dimensional voxel group R including a large number of restoration points P, the processing for determining the existence probability of the object A in each voxel Q using a GPU or the like is performed at high speed. Can be done. As a result, it is possible to reduce the time required for recognizing an obstacle or a gripped object.

  The present invention is not limited to the above embodiment. For example, in the above embodiment, the three-dimensional space including a large number of restoration points P is all divided into voxels Q of the same size, but the restoration points P located far from the cameras 2 and 3 are the cameras 2 and 3. The position accuracy is lower than that of the restoration point P at a position close to. For this reason, the size of the voxel Q located far from the cameras 2 and 3 and the size of the voxel Q located near the cameras 2 and 3 are not necessarily equal.

  Therefore, when dividing the three-dimensional space including each restoration point P into voxels Q in step S52 shown in FIG. 2, the size of the voxel Q is increased with increasing distance from the cameras 2 and 3, as shown in FIG. You may do it. At this time, the size of each voxel Q is set so that the voxel size projected on the image planes of the cameras 2 and 3 becomes a constant value (e × e) with reference to one of the image planes of the cameras 2 and 3. It is desirable to set.

  In this case, since the number of voxels Q included in one two-dimensional voxel group S decreases as the distance from the cameras 2 and 3 increases, the time required to recognize the object A can be further shortened. Further, since a small-capacity memory can be used as the memory (described above) necessary for the recognition processing of the object A, it is advantageous in terms of cost.

  Moreover, in the said embodiment, although the arithmetic processing unit 4 shall be mounted in the robot, you may install the arithmetic processing unit 4 in locations other than a robot. In this case, an image captured by the cameras 2 and 3 may be sent from the robot to the arithmetic processing unit by wireless communication, and the processing result by the arithmetic processing unit may be similarly sent to the robot by wireless communication.

  Furthermore, it goes without saying that the image recognition apparatus of the present invention can be applied to a vehicle other than a robot, for example.

1 is a schematic configuration diagram illustrating an embodiment of an image recognition apparatus according to the present invention. It is a flowchart which shows the detail of the arithmetic processing procedure by the arithmetic processing unit shown in FIG. FIG. 3 is a diagram conceptually illustrating a three-dimensional restoration process, a space division process, and a slice process illustrated in FIG. 2. It is a figure which shows notionally the existence probability determination process shown in FIG. It is a figure which shows notionally the modification of the space division process shown in FIG.

Explanation of symbols

  DESCRIPTION OF SYMBOLS 1 ... Image recognition apparatus, 2, 3 ... Camera (imaging part), 4 ... Arithmetic processing unit (three-dimensional reconstruction means, space division means, existence probability determination means), A ... Object, P ... Restoration point, Q ... Voxel , R ... three-dimensional voxel group, S ... two-dimensional voxel group.

Claims (3)

In the image recognition apparatus for recognizing the object using a captured image of the object acquired by the imaging unit,
Based on a captured image of the object, three-dimensional restoration means for restoring a plurality of points corresponding to the surface of the object on a three-dimensional space;
Space dividing means for dividing a three-dimensional space including a plurality of points restored by the three-dimensional restoration means into a plurality of voxels;
Image recognition comprising: presence probability determining means for counting the number of points included in each voxel obtained by the space dividing means and determining the existence probability of the object in each voxel apparatus. Means for converting a three-dimensional voxel group consisting of the plurality of voxels obtained by dividing by the space dividing means into a plurality of two-dimensional voxel groups;
The image recognition apparatus according to claim 1, wherein the existence probability determining unit counts the number of points included in each voxel for each two-dimensional voxel group.   The space dividing unit divides a three-dimensional space including the restored points into a plurality of voxels so that the size of the voxel increases as the distance from the imaging unit increases. Or the image recognition apparatus of 2.

Images