JP2005346391A - Object recognition method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an object recognition method for improving discrimination capability about an object of a simple shape. <P>SOLUTION: From the image of an acquired object (a step S1), an internal edge and an external edge are detected (a step S2), and featured values are detected on the basis of the extracted external/internal outlines (a step S3) (steps S3 to S6), and the calculation of its similarity with a registered pattern (a step S7) and the specification of the pattern based on this is carried out (steps S8, S9). When any pattern can not be specified from the external/internal outlines, closed curves based on the edges are successively added, and the featured value detection and the similarity calculation are carried out (steps S10 to S21), so that the pattern can be specified. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  The present invention relates to an object recognition method for recognizing an object extracted from an image.

  As an example of a control technique using the vision of a robot, development of a technique for recognizing an object from image information obtained by a visual device is recommended. As a recognition method for such an object, a knowledge-based object recognition technique for recognizing an object based on similarity to a pattern registered in advance is known. In such knowledge-based object recognition technology, it is necessary to accurately select a pattern similar to the target object from registered patterns.

  The technique described in Patent Document 1 is a technique related to such graphic classification / retrieval. The shape feature that is invariant to the rotation, scaling, and movement of the figure and the corresponding shape parameter are obtained, the coordinate conversion parameter to the model figure with the same shape feature is calculated using the shape parameter, and the model identifier and coordinate conversion parameter Each group is voted and classified into models with many votes. Similarly, the technique of Non-Patent Document 1 extracts feature amounts from contours and compares them with models.

Further, the technique of Patent Document 2 extracts a closed curve that constitutes a contour line inside and outside an object by tracking points that form a contour, and performs object discrimination based on this.
JP 2000-215315 A Japanese Patent Laid-Open No. 7-220089 Amit Sethi, et al. "Curve and Surface Duals and the Recognition of Curved 3D Objects from their Silhouette", International Journal of Computer Vision Vol. 58, No.1, June 2004, pp73-86

  In the case of discriminating using the contour in this way, if the external contour of the recognition object is a simple shape such as a sphere, an ellipse, a rectangular parallelepiped, a cylinder, or a cone, it is difficult to determine the difference. For example, there is a problem that it is difficult to distinguish two objects having a cylindrical outer shape such as a can and a cup. In addition, it is difficult to discriminate between objects having similar shapes and different surface patterns.

  Therefore, an object of the present invention is to provide an object recognition method that improves discrimination power even for an object having a simple shape.

  In order to solve the above problems, an object recognition method according to the present invention extracts an edge of an image, extracts an edge of an object in the image, further extracts a closed curve from the edge, and based on the object In the object recognition method for recognizing the shape pattern, a closed curve is generated by combining the extracted internal edge and external edge, and the feature pattern is extracted from the generated closed curve to recognize the shape pattern. To do.

  In this way, by finding combinations of closed curves that can be created by combining internal edges and external edges, not only external contours and internal contours of objects, but also information such as internal edge boundaries that do not constitute closed curves by themselves are used. Extract features. Thereby, the dimension of the feature amount becomes higher than when only the contour information is used. The closed curve extraction combining the inner edge and the outer edge is performed by extending an object contour extraction algorithm.

  This feature amount extraction is performed on a group of closed curves composed of a plurality of closed curves. Each edge is prioritized, and a shape pattern cannot be specified with a predetermined accuracy from a closed curve obtained by combining high-priority edges. In this case, the shape pattern may be specified by sequentially adding edges with inferior priorities to generate a closed curve and calculate a feature amount.

  That is, instead of extracting feature values for all closed curve groups that can be extracted from the beginning, a closed curve is generated by adding an edge according to the recognition result, and additional calculation of feature values is performed.

  In the prioritization of edges, the priority of the edges constituting the contour line and the edge having high reliability among the internal edges is increased, and the priority of the edge having low reliability among the internal edges is decreased. This reliability may be determined by, for example, the ratio of noise components for each edge.

  Here, the predetermined accuracy means that the degree of similarity with any shape pattern exceeds a predetermined value.

  Alternatively, the object recognition method according to the present invention extracts an outline of an object in an image by extracting an edge of the image, extracts a feature amount based on the contour line, and based on the extracted feature amount In the object recognition method for recognizing the shape pattern of an object, when a shape pattern with a similarity exceeding a predetermined value cannot be found based on a feature amount derived from an outline, a shape pattern with a similarity exceeding a predetermined value is further found. From the closed curve consisting of one or more internal edges or external edges, the closed curve with the highest reliability among the closed curves not set in the closed curve group is sequentially added to the closed curve group, and the feature amount is calculated based on the closed curve group. Extracting and repeating object shape pattern recognition based on the extracted feature quantity. Here, a configuration is adopted in which a closed curve is added instead of an edge.

  According to the present invention, since the feature amount is extracted using a plurality of closed curves derived from the inner edge or the like even for an object having a simple shape that is difficult to recognize only by the outer contour line and the inner contour line, the dimension of the feature amount It is possible to improve the recognition accuracy.

  Also, instead of using all closed curves recognized from the beginning, feature quantities are calculated, and by sequentially adding edges or closed curves, feature quantity calculation and similarity determination are performed, the number of closed curves actually processed is reduced. Thus, optimization of the processing can be achieved, so that an increase in the amount of calculation processing can be suppressed, and both improvement in recognition accuracy and high speed can be achieved.

  In addition, by using an edge / closed curve with a high degree of reliability, it is possible to ensure the reliability of the feature amount and to ensure the recognition accuracy.

  DESCRIPTION OF EXEMPLARY EMBODIMENTS Hereinafter, preferred embodiments of the invention will be described in detail with reference to the accompanying drawings. In order to facilitate the understanding of the description, the same reference numerals are given to the same components in the drawings as much as possible, and duplicate descriptions are omitted.

  FIG. 1 is a schematic configuration diagram of an image processing apparatus that implements an object recognition method according to the present invention. The image processing apparatus 100 includes a camera 1 that captures an image of an object 200, an image processing unit 2 that receives a video signal from the camera 1 and performs image processing, and a monitor 3 that displays a processing result.

  The camera 1 is, for example, a video camera, and can transfer a moving image acquired at 30 frames per second to the image processing unit 2. The video signal at this time is, for example, a digital signal in which one screen is composed of 640 × 480 dots and each pixel is represented by 8 bits for each of RGB. In addition to the form in which the camera 1 converts it into a digital signal and transfers it to the image processing unit 2, an image processing unit that uses an analog video signal (for example, NTSC format) is used for the camera 1. A configuration in which an A / D converter is provided in 2 or a configuration in which an A / D converter is provided independently between the image processing unit 2 may be adopted. Here, a case where the camera 1 is a single unit will be described. However, the object 200 is stereoscopically viewed using a plurality of cameras, and information from a plurality of cameras is processed in parallel by one image processing unit 2. You may make it do.

  The image processing unit 2 includes an image processing calculation unit 20 configured by a CPU, a ROM, a RAM, and the like, and an information storage unit 21 including a storage unit such as a hard disk for accumulating and storing comparison pattern data. Yes. The image processing unit 2 may be configured as dedicated hardware specialized for object recognition processing, or uses a general-purpose computer such as a personal computer or a workstation, and software that operates on these computers. You may implement | achieve the object recognition method concerning.

  Next, the image processing method according to the present invention will be specifically described. FIG. 2 is a specific flowchart of this processing method, and FIGS. 3 to 7 are diagrams for explaining the details of individual processes.

First, an image of an object (target object) is acquired from the camera (step S1), and external / internal edges of the object are detected from the image by edge detection or the like (step S2). FIG. 3 is an example in which the extracted edges 201h _{1 to} 201h ₃ are superimposed and displayed on the image of the mug 201 with the handle, and FIG. 4 is an overlay of the extracted edges 202h _{1 to} 202h ₃ on the image of the cup 202. It is an example displayed. The edges extracted here are only edges that form a closed curve by singly or connecting to other edges.

Next, first, an outer contour line and an inner contour line are extracted from the extracted edges 201h _{1 to} 201h ₃ or 202h _{1 to} 202h ₃ (step S3). The external contour line at this time is a boundary line (outer boundary line) between the projected image when the object is silhouette-projected and the surrounding background or other object (outer boundary line), and the internal contour line is a projection inside the projected image. A boundary line (inner boundary line) between the image and the background or another object. Therefore, when the mug 201, edge 201h ₁ is an external contour, the edge 201h ₂ is an internal contour. On the other hand, in the case of the cup 202, only the outer contour lines 202h ₁ and 202h ₂ are detected. Both the outer contour line and the inner contour line are closed curves.

  Next, a point where the gradient is the same on the extracted contour (external contour line and internal contour line) is obtained (step S4). Here, the point where the gradient is the same means a point where the gradient (gradient) of the tangent of the contour is the same, that is, a point where the tangent of the contour has a parallel relationship (see FIG. 5). Here, the case of the mug 201 is shown as an example). Then, distances d1 to dm between the tangents of the same gradient point are calculated (step S5), and a value obtained by dividing each straight line distance d1 to dm by the longest straight line distance dmax and converting it to a dimensionless value is a feature amount. (Invariant) is set (step S6). Here, the order of d1-dm is set according to the order of the group of the corresponding same gradient point. Actually, the calculations in steps S4 to S6 are obtained for a large number of points and tangent groups having different gradients. At this time, the gradient point, tangent, and distance may be obtained while shifting the gradient by a predetermined angle (for example, by 5 degrees).

  When the feature amounts d1 to dm thus obtained are plotted in the m-dimensional space, a predetermined curve is drawn as shown in FIG. 6 (here, the case where the feature amount is cubic, ie, m = 3 is illustrated). ing). When this feature amount extraction is performed on images with different viewpoints and all the feature amounts are similarly plotted, the feature amount has a curved surface shape as shown in FIG. Feature amount data corresponding to the curved surface shape data is obtained for various patterns and stored in the information storage unit 21 in advance. Then, the similarity is calculated by comparing the feature amount data accumulated in the information storage unit 21 with the obtained feature amount data of the object (step S7). This corresponds to checking whether or not the curve of FIG. 6 is included in a part of the curved surface of FIG.

  After obtaining the similarity for various patterns and calculating the maximum similarity, the maximum similarity is compared with a preset similarity threshold (step S8). If the maximum similarity exceeds the similarity threshold, it is specified that the object has a pattern that is the maximum similarity (step S9), and the process ends.

  On the other hand, when the maximum similarity is less than or equal to the threshold value, a determination is made by adding a closed curve group. First, it is determined whether there is an edge that can form a closed curve group other than the outline (step S10).

  If it is determined that there is an edge that can form a closed curve group other than the contour, prioritization is performed for the closed curve formed by a combination of edges (step S11). The priority order of the closed curves may be based on the degree of reliability for each edge. In other words, the priority order of closed curves obtained by combining edges with less noise based on continuity, thickness, etc. may be set high.

  After the priority order is set, the number of closed curves that can be added in combination is set to a variable n, and an initial value 1 is set to a variable i indicating the priority order of the closed curve to be added (step S12). Then, loop processing is started.

  First, a closed curve having the i-th priority order is added to the closed curve group (step S13). For the closed curve group including the added closed curve, the feature amount is set by the same method as in steps S4 to S7, and the similarity is calculated (steps S14 to S17). When the similarity is obtained for various patterns and the maximum similarity is calculated, the maximum similarity obtained again is compared with a preset similarity threshold (step S18). When the maximum similarity exceeds the similarity threshold, the process proceeds to step S9, where it is specified that the object has a pattern that is the maximum similarity, and the process ends.

  On the other hand, if the maximum similarity is less than or equal to the threshold value, it is further determined whether or not all closed curves have been added by determining whether i is greater than or equal to m (step S20). If i is less than m, it means that the addition of all the closed curves that can constitute the closed curve group has not been completed, so 1 is added to i (step S21), and the process returns to step S13. A closed curve is added to calculate feature values and determine similarity.

  On the other hand, if i has reached m, it means that the addition of all the closed curves that can form the closed curve group has been completed. In this case, similarly to the case where it is determined in step S10 that there is no edge that can form the closed curve group other than the contour, it is determined that the object recognition based on the feature amount extraction from the closed curve group cannot be performed, and the process proceeds to step S22. And the process determines that the object is an unknown object and ends the process.

  According to the present invention, in the case of the mug 201, as shown in FIG. 5, the dimension of the feature amount increases to four dimensions. In the conventional method using only the outer contour, the feature amount is quadratic or cubic (see FIG. 8). Therefore, in the conventional method, the similarity (which is 100 when completely matched) is 84, and the similarity is 54 even when compared with the clover figure shown in FIG. Can be identified. For this reason, discrimination power improves.

Further, in the case of the cup 202, it is impossible to distinguish from the same shape can and cylinder only from the contour information (edges 202h ₁ and 202h ₂ ) as shown in FIG. Again, closed curve and consisting of edges 202h _1, 202h ₃ shown in FIG. 10 (b), further sequentially added to recognize a closed curve consisting of the edge 202h _2, 202h ₃ that shown in FIG. 10 (c) By doing so, identification becomes possible. In particular, even if the outer shape is simple, it is effective for identifying an object having a feature in the internal shape or pattern. In the case of the mug 201 as well, when identifying an object having a similar shape (for example, a mug having a different surface shape or pattern), in addition to the outer contour and the inner contour, a closed curve combining the inner edge and the contour is used. It is good to add and identify.

  Here, the embodiment in which closed curves are sequentially added has been described. However, when there are a large number of edges, instead of adding closed curves, the edges are added sequentially and closed curves including the added edges are included. May be obtained, and a feature amount may be calculated for that.

  In the above description, the distance between the tangent lines of the contour having the same gradient is used as the feature amount. However, the feature amount is not limited to this as long as the feature amount is invariable with respect to the rotation and expansion / contraction in the pattern screen. For example, curvature, color information, a histogram, etc. can be used.

  For example, FIG. 11 shows a case where the target image is a figure in which three circles 203 to 205 having different colors are arranged close to each other. The circle 203 is painted green, the circle 204 is painted yellow, and the circle 205 is painted red. In this case, as shown in FIG. 12, if only the external contours 203h to 205h are extracted and compared, it is not possible to identify a figure or the like having a different color / pattern. Therefore, as shown in FIG. 13, color information is added to the above-described distance information and set as a feature amount (here, the density is shown as the density of contours 203 h ′ to 205 h ′ instead of color). It is also possible to identify images with different colors and patterns. The color information in this case may be set based on information such as the color value, hue, and luminance of the contour position.

  The present invention can be applied to, for example, a robot visual recognition technique, a monitoring system, and the like.

It is a schematic block diagram of the image processing apparatus which implements the object recognition method which concerns on this invention. It is a flowchart which shows the process of the object recognition method concerning this invention. This is an example in which the extracted edge is superimposed and displayed on the image of the mug with the handle. This is an example in which the extracted edges are superimposed on the cup image. It is a figure which shows the mode of the feature-value extraction from a mug. It is a graph which shows a mode that the feature-value extracted from the image was plotted in three-dimensional space. It is a graph which shows a mode that the stored feature-value was plotted in three-dimensional space. It is a figure which shows the mode of the feature-value extraction from the mug by the conventional method. It is a figure which shows the clover figure of a comparison object. It is a figure explaining the mode of addition of the closed curve group in a cup. It is an illustration of the target image which has color information. It is a figure which shows the outline information extracted from FIG. It is a figure which shows the outline information containing the color information extracted from FIG.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Camera, 2 ... Image processing part, 3 ... Monitor, 20 ... Image processing calculating part, 21 ... Information storage part, 100 ... Image processing apparatus, 200 ... Object, 201 ... Mug, 202 ... Cup, 203-205 ... Circle, 201h _{1 to} 201h ₃ , 202h _{1 to} 202h ₃ ... Edge.

Claims (5)

In the object recognition method for extracting the edge of the image, extracting the edge of the object in the image, further extracting the closed curve from the edge, and recognizing the shape pattern of the object based on this,
An object recognition method, wherein a closed curve is generated by combining an extracted internal edge and an external edge, and a feature amount is extracted from the generated closed curve to recognize a shape pattern.   The feature amount extraction is performed on a group of closed curves including a plurality of closed curves. Each edge is prioritized, and a shape pattern cannot be specified with a predetermined accuracy from a closed curve obtained by a combination of edges having high priority. The object recognition method according to claim 1, wherein the shape pattern is specified by sequentially adding edges with inferior priorities to generate a closed curve and calculating a feature amount.   The prioritization of the edges is characterized by increasing the priority of the edges constituting the contour line and the edge having high reliability among the internal edges and decreasing the priority of the edge having low reliability among the internal edges. The object recognition method according to claim 2.   The object recognition method according to claim 3, wherein the predetermined accuracy is a case where the degree of similarity with any shape pattern exceeds a predetermined value. In an object recognition method for extracting an outline of an object in an image by extracting an edge of the image, extracting a feature amount based on the contour line, and recognizing an object shape pattern based on the extracted feature amount ,
If a shape pattern with a similarity exceeding the predetermined value cannot be found based on the feature value derived from the contour line, it is further possible to use one or more internal edges or external edges until a shape pattern with a similarity exceeding the predetermined value is found. From the configured closed curves, the closed curves with the highest reliability among the closed curves not set in the closed curve group are sequentially added to the closed curve group, and feature quantities are extracted based on the closed curve groups. An object recognition method characterized by repeating shape pattern recognition.

Images