JPS62284479A - Device for recognizing stereoscopic shape of object - Google Patents

Abstract

PURPOSE:To recognize a local recessed part and to improve the accuracy of recognition of the whole shape by combining a conical body penetration method with a stereoscopic vision method based upon both eyes. CONSTITUTION:The processings of steps 101-105 are shape recognition based upon the conical body penetration method and the processings of steps 106-109 are shape correction based upon the stereoscopic vision of both eyes. Images are photographed by TV cameras 1-8 and the envelope surface of the shape of an object is recognized by using the images 1-4 picked up by the cameras 1-4 based on the conical body penetration method. Then, shape correction based upon stereoscopic vision is executed by using the images 1, 5; 2, 6; 3, 7 and 4, 8. Namely stereoscopic vision can be executed by using a geometric restricting condition obtained by the conical body penetration method, so that an accurate stereoscopic shape can be recognized. Consequently, a local recessed part which can not be recognized by the conical body penetration method can be recognized and the accuracy of the whole shape recognition can be improved.

Description

Detailed Description of the Invention 3. Detailed Description of the Invention [Industrial Field of Application] The present invention relates to a V & Kasumi and method for recognizing the three-dimensional shape of an object using a camera system, and particularly relates to data processing in a computer graphics system. The present invention relates to an apparatus and method for recognizing the three-dimensional shape of an object, which is suitable as an input method.

[Conventional technology]

As a conventional method for recognizing the three-dimensional shape of an object, Information Processing Society of Japan Computer Vision Study Group Materials, 37-2 (1
As discussed in 985), objects are photographed from multiple directions with a TV camera.

Based on the contour line information of the image, a three-dimensional shape is created in the computer as a voxel (VOXEL) model using the cone interaction method.
The method was to reconstruct it as dimensional coordinates. but,
The shape obtained by this method is the envelope surface of the object, and no consideration was given to the case where the object has a concave portion.

In addition, as a method that can recognize local depressions and input them into the computer, a binocular stereoscopic (stereoscopic) method as discussed in the proceedings of the 24th 5ICE Academic Conference, pp. 847 to 848 (1985) (visual) law is known.

[Problem that the invention seeks to solve]

The conventional technique using the cone interoperability method described above was a method of reconstructing a three-dimensional shape from image contour information.

If an object has a local depression that does not appear on the contour line, this part cannot be recognized, and the shape input into the computer has problems in terms of accuracy.

On the other hand, in the recognition method using binocular stereopsis, if you try to enlarge the disparity of each eye to obtain sufficient accuracy, the correspondence between the left and right images becomes a chore, and the solution becomes inaccurate. There was a problem.

An object of the present invention is to solve the problems of the prior art, to enable the recognition of one local recess, and to improve the accuracy of overall shape recognition.

[Means for solving problems]

The present invention that achieves the above object is characterized by: a camera system that photographs an object from multiple viewpoints; and a system that calculates geometric constraints regarding the position and/or shape of the object from the output of the camera system. A first calculating means calculates whether or not the intersection position of lines of sight from two viewpoints of the camera system satisfies the geometrical constraint, and determines the intersection point that satisfies the geometrical constraint. and a second calculating means for calculating three-dimensional coordinates of corresponding points of the object based on the position and the above-mentioned geometric constraint conditions.

In a preferred embodiment of the present invention, in addition to the cone interaction method used as the principle of shape recognition in the prior art, a binocular stereoscopic method that can recognize local depressions is combined.

[Effect]

Recognize the envelope surface of an object shape using the cone-transfer method.

Since binocular stereoscopic vision is performed using the geometrical constraint conditions obtained from this, the parallax between both eyes can be increased, and sufficient accuracy can be obtained to recognize more precise three-dimensional shapes. As a result, it is possible to recognize local depressions that could not be recognized when using the cone interaction method alone, and the accuracy is further improved compared to when the binocular stereoscopic method is used alone.

〔Example〕

Hereinafter, one embodiment of the present invention will be described in detail with reference to the drawings.

FIG. 1 is a flowchart showing an outline of the processing procedure when the embodiment of the present invention is realized by computer software, and FIG. 2 shows a part of the procedure in detail. Further, FIG. 3 is a diagram illustrating the positional relationship between the cameras 1 to 8 and the large blade object 9. still.

Of course, a dedicated circuit may be used instead of computer software.

Processing at each step will be explained based on the flowchart in FIG. In this diagram, steps 101 to 10
The processes up to step 5 are shape recognition using the cone interaction method, which corresponds to the first calculation means, and steps 106 to 5
Reference numeral 109 indicates shape recognition correction using binocular stereoscopic vision, which corresponds to the second calculation means.

First, in step 101, an image is photographed by a camera system consisting of TV cameras 1 to 8, and the output data is inputted into a computer. The positional relationship of the viewpoints at this time is as shown in FIG. 3, and the cameras 1 to 4 are arranged in four orthogonal directions with the large blade object 9 as the center. Furthermore, xYZ fixed in space
It is assumed that cameras 1 and 3 are located in the ±Z-axis direction of the orthogonal coordinate system, and cameras 2 and 4 are located in the ±X-axis direction. Further, camera 5 is located at a position where camera 1 is moved horizontally by a predetermined distance to the right toward the origin, and the other cameras 6 to 8 are located in the same way as cameras 2 to 4.
The position is horizontally moved.

The two-dimensional image data taken by each camera and input to the computer are shown in FIGS. 4(a) to (h), where the image data of camera 1 is shown in image 1 shown in FIG. 4(a). The same applies to the following. In each two-dimensional image, the area where the object is projected is called an object image, and the other area is called a background image.

The cone interoperability method is a method of finding a set of envelope surfaces of an object from image contour information and using this as an approximate shape.The zero-order step 102 is the process of the following steps 103 to 105. is repeated for two-dimensional images 1 to 4 in FIG. 4.

Images 5 to 8 from cameras 5 to 8 are not used in the shape input using the cone-penetrating method.

In step 103, each two-dimensional image is subjected to threshold processing and binarized to extract an object image.

In step 104, a hypothetical existence area that becomes part of the geometrical constraints regarding the position and/or shape of the object is calculated. The hypothetical existence area is the area where camera 1 is located as shown in Figure 5.
It is a cone-shaped area whose apex is the projection center of , and whose cross-sectional shape is the object image of one image. This means that the large blade object will always exist inside this area. In order to calculate this region, this cone-shaped shape is described using a voxel model. A voxel model is a model that describes a three-dimensional shape in terms of the presence or absence of three-dimensional, equally spaced grid points, and the space in which voxels are defined is called voxel space. Place it in the same position.

Each voxel is a cubic element of equal size, and a voxel model is a collection of cubes that expresses a three-dimensional shape. The method of calculating the hypothetical existence area when using this model is shown in Figure 6. The hypothetical existence area in Figure 6 corresponds to the line of sight direction of the camera l, and the voxel space in Figure 5 is transformed into the XZ plane. It was cut by. To simplify the explanation, the voxel space is 10X
It is assumed that the image is composed of 10XIO voxels.

The cross section of the hypothetical existence region on the XY plane is similar to the shape of the object image, and its size is proportional to the distance on the Z axis from the projection center. Therefore, X in voxel space
The Y cross section can be obtained by scaling the object image obtained from the camera according to the distance. By sequentially substituting the scaled object images into the voxel space, a hypothetical existence area indicated by hatching can be obtained as a voxel model as shown in FIG. As shown in Figure 5, the hypothetical existence area expands as you move away from the projection center, so the 6th
Also in the figure, the region where Z is large has a wide shape.

In step 105, a common hypothesis existence region is determined. 7th
The figure is a diagram for explaining a common hypothetical existence area, and the common hypothetical existence area is a logical product of hypothetical existence areas corresponding to each projection center, and is an area indicated by hatching in the figure. Figure 8 shows a cross-section of the common hypothetical existence region found in voxel space (
FIG. The logical product between hypothetical existence regions can be calculated by a logical operation for each voxel. That is, by leaving only voxels that are common to all projection directions.

A shape as shown in FIG. 8 is obtained. In this way.

If the common assumption existence region is found, the large blade object has a shape that is inscribed in this region. In other words, the common assumed existence region is a set of envelope surfaces of the large-blade object, and this is assumed to be the approximate shape of the first-order large-blade object. This becomes a geometric constraint regarding the position and/or shape of the object. FIG. 9 is a diagram showing this shape three-dimensionally using a display means (not shown).
This is a display of the results obtained by shape recognition using the cone interoperability method.

The processing from step 106 onwards is to correct the obtained approximate shape using the binocular stereoscopic method.
Processing is performed using a combination of two two-dimensional images with different viewpoints, 6, 3, 7. 4, and 8. This shows that the depth to a large blade object is determined from two two-dimensional images.

For example, in Figure 4, image 1 corresponds to the left image, and image 5 corresponds to the right image.0 The basis of binocular stereopsis is to find the depth to an object from corresponding points in the left and right images using the principle of triangulation. It is.

Here, like camera 1 and camera 5 in FIG.

A method for finding corresponding points in an image when two cameras are placed horizontally facing the same direction will be explained.

It is now assumed that the surface point P on the 10th bladed object is located at the position of point O on the left image.

At this time, if the point R corresponding to the dot can be determined in the right image, the straight lines τV and RP are known, and the depth distance to the point P can be calculated by triangulation. By the way, as shown in Figure 3, when two cameras are placed horizontally at a predetermined distance apart, the points P and R are on the same plane due to the geometrical relationship, and the correspondence between the points is It can be seen that point R always exists somewhere on the straight line at the same height = h) on both image planes. This straight line is generally called an epipolar line, and exactly the same relationship exists for points on the left image that correspond to points on the right image.

Therefore, the problem of the correspondence between the left and right images can be reduced to the problem of matching the left and right evipolar lines, which are -dimensional vectors. , and search for corresponding points using dynamic programming.

FIG. 11 is a diagram illustrating a method of matching left and right epipolar lines by focusing on the brightness values of pixels on the epipolar lines. The graph showing the brightness distribution on the left epipolar line on the upper side of Figure 11 is the same as the one on the left image of Figure 10, and the graph of the brightness distribution on the lower right side of Figure 11 is on the right side of Figure 10. It is the same as the one in the image. This brightness distribution assumes that the part closest to the viewpoint is the brightest. Also, step 1
Similarly to 03, one line is divided into background image and object image areas by threshold processing. Matching of the left and right epipolar lines is performed by selecting the minimum cost path connecting the starting point S and the ending point E on the correspondence search map. The difference in brightness between the left and right pixels is used for cost calculation. The path for which the sum of all costs is the minimum is the minimum cost path. The correspondence search map shows how to combine the left and right pixel columns.The upper left point S indicates the start point of the path, and the lower right point E indicates the end point.The correspondence between the left and right pixels is determined by the path connecting the two. The relationship is determined. The details will be explained in a later step, but paths that move diagonally on the map indicate areas where pixels cannot be matched, and paths that move vertically and horizontally indicate areas that could not be matched. For example, each pixel included in section A on the left evipolar line is divided into section B on the right line.
There is a one-to-one correspondence with the internal pixels.

Regarding pixels that are matched in this way.

Based on the positions of both lines in real space, the depth to the large blade object can be calculated.

Step 107 indicates that the subsequent processing is repeated for each Epipolar line, and the entire image is scanned to obtain information on the three-dimensional coordinates of the large blade object.

In step 108, depth calculation is performed. Details of this portion are shown in steps 201 and subsequent steps in FIG. 2.

In step 201, first, a full correspondence search map is created.

As shown in FIG. 11, this represents the correspondence relationship between all pixels on the left and right evipolar lines. This process, on a computer, determines which combination of left and right pixels the path in the map represents.

It is to define. Now, if the number of pixels on the left line is m and the number on the right is n, then the number of nodes in the all correspondence search map is (m
+1) ・(n+1) pieces. Here, the contact point (i
, j), then (i.

The diagonal path connecting j) and (i + 1, j + 1) represents that the i+1th pixel on the left line is matched with the j+1th pixel on the right line.

Step 202 is an operation of cutting out a partial correspondence search map from the total correspondence search map. The 0 partial correspondence search map is an area of the total correspondence search map that corresponds to the object image portion of the left and right lines.

In FIG. 11, the area surrounded by a-b-d-c corresponds to this. Subsequent depth calculations are limited to only the paths inside this area.

FIGS. 12(a), 12(b), and 12(c) are diagrams showing the state of processing at nodes in the partial correspondence search map, and are part of the full correspondence search map. L5 to L8 are the 5th to 8th pixels on the left line and have a brightness distribution, for example, as shown in the figure. Also.

R3 to R6 are similar pixels on the right line.

And 325 from Sl is the node number, and D8 from DO
indicates the number of nodes.

In step 203, a minimum attainment cost map is created. As shown in FIG. 12, this corresponds to node S1 to node 325 of the partial correspondence search map, and if the number of pixels of the object image on the left line is mJ and that on the right line is n', then m' This becomes a two-dimensional array consisting of +1 columns and Xn'+1 rows. Figure 12 shows m' = 4. This is the case when n'=4. For each array element, in a later step, the minimum cost to reach the node is calculated and stored here.

Step 204 indicates that the subsequent steps 205 to 209 are processed for all nodes in the partial correspondence search map.

Step 205 is an operation of assigning a node number to one node.
Assign node numbers of j' X (Ω+1)+i'+1 to the nodes in the row. In the 12th @, 31 to 825
is the node number.

FIG. 13 is a diagram illustrating straight lines connecting the viewpoint and pixels on the image plane and their intersections. It is now assumed that there are 11 pixels on each of the left and right image planes.

Qll from the straight line capsule 1 is the pixel on the left image plane and the left viewpoint 'r
1 to rll is a similar straight line on the right.

Among these, the straight lines corresponding to the object image are from Q5 to R8 and from r3 to r6. Incidentally, in binocular stereopsis, there is a property that the accuracy is dominated by the parallax between the two eyes. The distance between the left viewpoint and the right viewpoint is the parallax, and as this becomes larger, the distance between the front and back of the intersection becomes narrower, so accuracy can be improved.

In step 206, the intersection point of paths is calculated 1. For example, the path S7→S13 in FIG.
Indicates that they are matching.

The actual object positions are Q6 and r4 in Fig. 18.
It is the intersection position of Already, the external shape of the object to be recognized can be determined as a set of these intersection points. In this step, positions are calculated for intersections other than the * marks in FIG. 13. At least one of the left and right pixels of an autointersection is a background image and is unrelated to the position of the large blade object. The viewpoint position and the pixel position are known, and the intersection position can be numerically determined as the intersection of two straight lines.

Step 207 is the setting of a penalty pass. 1st
In Figure 3, there are combinations of intersections marked with O and 0 in addition to marks. The one marked 0 means an intersection where the left and right pixels are both object images.

Those marked O are internal intersections of the common assumption existence region, which is a further geometrical constraint.

The common hypothesis existence region is determined as a set of elements in the voxel space as shown in FIG. Therefore, it is possible to determine whether the intersection point found using this information is inside the intersection point. Since the large blade object is always inside,
If the result of the judgment is outside, a penalty will be imposed on the relevant pass. In FIG. 12, the passes marked with 0 are penalty passes. In this step,
It is determined whether a penalty path is included in the path starting from the node, and if there is a penalty path, the penalty value is substituted into the minimum arrival cost map of the node.

The penalty value is a sufficiently large constant value of 0.For example, if the brightness level of a pixel is 256 levels,
A value of 256 or more may be used as a penalty value.

Also, in the 12th @, the node that does not impose a penalty is SL,
87,813,519. S25 is the only step, and the minimum cost path is uniquely determined.However, in actual cases, the number of pixels is larger, so a plurality of candidates for the minimum cost path exist near the diagonal line.

In step 208, the number of stages of nodes is assigned. This is a process necessary to match the left and right lines by dynamic programming. For example, in the example of FIG. The node is Sl, and the node S in the D1 stage
2 and S6 belong to this category. That is, a series of nodes arranged diagonally on the one-part correspondence search map are assumed to be on a common level.

Here, a method for finding the minimum cost path using dynamic programming will be explained. Let the starting point be S, and the (1' y
Let us denote the minimum cost to reach the node at position Gh (1'ej' ), then the recurrence formula for reaching the end point with the minimum cost path is as follows.

Go (0, O) = 0, where the number of pixels of the object image on the left line is m',
Letting the right one be n', the coordinates of S are (0°0) and the coordinates of H are (m', n'), where dl is vertical.
At the cost of the horizontal moving path.

Calculate the average difference in brightness between all left and right pixels.

Given as a constant value. d- is calculated by the following formula.

da(i' e j')=II bird(i')-I
t(j')I.

Here, Ia(i') indicates the brightness of the 11th pixel on the left line, and Ir(j') indicates the brightness of the 52nd pixel on the right line. The calculation using this recurrence formula is repeated from S to E to find the minimum cost path.

Step 209 sets the order of node cost calculation, and it is necessary to calculate from the smallest number of stages.
In the case of the 12th @, DO→D1→D2→D3→D4→D
Set the calculation to proceed in the order of 5→D6→D7→D8.

Step 210 indicates that the following process is repeated for each stage from the DO stage to the D8 stage.

Step 211 indicates that the following process is repeated for the nodes in the same stage, starting from the node with the smaller node number. For example, in the D4 stage, S5 → S9 → 813 → 3
Cost calculation is performed in the order of 17→821.

Step 212 is the calculation of the minimum reaching cost, and in order to reach 6, for example, 513, the above-mentioned gradual formula is calculated, S8→513. S7→S13゜812→S13-3
There are two paths, and you want to find the one that can be achieved at the lowest cost. Now, S7= (1, 1), S8=
(2,1), 512= (1,2), 513= (2
, 2), so it becomes 1, that is, the smallest path is selected from these three paths.

Step 213 is a process of storing the minimum attainment cost obtained in step 212 in the map.

For example, the value GD4 (2,2) is stored at position 313 in the map.

In step 214, the node numbers of the start and end points of the minimum cost path to each node are stored in the connection list. The structure of the connection list is shown in FIG.

Step 215 is a process of inversely searching the created connection list from the end point side.0For example, the starting point of the end point S25 is 51
9, and that the starting point of S19 is S13 is determined by reverse search.

This process is repeated until the starting point S1 is reached. In other words, from the information in the connection list, it is known which other contact point should be used as the starting point to reach a certain node at the lowest cost. Therefore, the path obtained by the reverse search becomes the minimum cost path connecting the starting point and the ending point.

In step 216, the depth to the object is calculated again from the determined matching information of the left and right pixels. This is the same as in step 206, and is a calculation to find the intersection of two straight lines. Note that this intersection point calculation is performed only for the paths that constitute the minimum cost path.

The process of finding the minimum cost path from the all-correspondence search map will be explained again with reference to FIG. First, as shown in FIG. 13, on the left epipolar line, pixels L5 to L8
is an object image, and on the right line are pixels R3 to R6. The paths marked with * in the diagram indicate that at least one of them is a background image. These paths mean depth information about background parts other than the object to be recognized. Therefore, from the matching calculation of the left and right lines, we remove the 0-order remaining part, perform the cost calculation for matching, and calculate the intersection positions sequentially for the combination of 4 intersections that each pass means. Among the intersection points, it is further determined whether or not they are inside the co-hypothetical existence region.In Figure 1, those that are inside are marked with an O, and those that are outside are marked with a 0.

A path in which both the left and right sides are object images but whose intersection point is outside the common assumed existence region is a penalty path. When cost calculation is performed using dynamic programming, the minimum cost path is selected from the paths marked with an O. In this way, in this embodiment, the range for selecting the minimum cost path is narrowly limited in two steps by geometrical constraints. With conventional technology.

Since the selections were not limited in this way, if the parallax was increased to improve accuracy, there was a risk of erroneous matching. If the disparity is large, the true minimum cost path will pass through a region far away from the diagonal of the all-correspondence search map. In other words, the number of pixels that cannot be matched increases, and depending on this cost setting,
There were cases where incorrect matching occurred, causing a decrease in accuracy.

In the eighth snap 109, which returns to the flowchart of FIG. 1, the common assumption existing region is corrected from the information obtained by binocular stereoscopic vision. The cross section of the shape obtained by the cone interpenetration method is as shown in FIG. On the contrary,
The cross section after the modification is as shown in Figure 15. FIG. 16 shows a three-dimensional voxel model displayed by a display means (not shown).

The shape shown in FIG. 16 is the final shape obtained, and this information is continuously managed and processed within the computer.

According to this embodiment, the approximate shape of the object is recognized by the cone-penetration method. This method uses contour line information, which is resistant to the effects of noise, among image information, so it can stably obtain shapes. However, the obtained shape is a collection of envelope surfaces of the object, and local concavities cannot be input. Therefore, in this example, shape recognition was performed using binocular stereoscopic viewing. Parts are being corrected. At this time, the common assumed existence region obtained in the voxel space is used as a geometric constraint for matching the left and right images. With this condition, the range through which the minimum cost path passes can be limited to a narrow range, and the accuracy of shape recognition can be further improved to 1 for the set of envelope surfaces of the object.
Since the final shape is obtained by making corrections using binocular stereopsis, it is possible to recognize local concavities that could not be recognized with conventional technology, improving the accuracy of overall shape recognition, and displaying this recognition data. It can be effectively used in computer graphics systems by processing and managing it.

In addition, in the conventional binocular stereoscopic viewing method, there was a problem in that it became difficult to make correspondences when the disparity was enlarged in order to obtain sufficient accuracy. Using geometric constraints, the mapping narrowly limits the search range and can eliminate erroneous mappings.

Therefore, the accuracy is improved compared to the conventional case where binocular stereopsis is used alone.

〔Effect of the invention〕

According to the present invention, it is possible to recognize local recesses and improve the accuracy of recognition of the overall shape.

[Brief explanation of drawings]

Figure 1 is a flowchart of one embodiment of the present invention, Figure 2 is a detailed flowchart of a part of Figure 1, Figure 3 is an explanatory diagram of the positional relationship between the camera and the large blade object, and Figure 4 is an illustration of the positional relationship between the camera and the large blade object. An explanatory diagram of an image obtained by a camera. Figure 5 is an explanatory diagram of the hypothetical existence area, Figure 6 is an explanatory diagram of the hypothetical existence area in voxel space, Figure 7 is an explanatory diagram of the common hypothesis existence area, and Figure 8 is an explanatory diagram of the hypothetical existence area in voxel space. Figure 9 is an illustration of a three-dimensional model using voxels, Figure 10 is an illustration of binocular stereoscopic viewing, Figure 11 is an illustration of a method for matching left and right images, and Figure 12 is a partial correspondence search. An explanatory diagram of the map, Fig. 13 is an explanatory diagram of the intersections meant by buses, Fig. 14 is an explanatory diagram of the all correspondence search map, Fig. 15 is an explanatory diagram of the common assumption existence region after correction in voxel space, FIG. 16 is an explanatory diagram of the finally obtained three-dimensional model using voxels. 1.2, 3.4, 5, 6, 7.8...camera, 9...
・Large blade object.

Claims (1)

[Claims] 1. A camera system that photographs an object from a plurality of viewpoints, and a first calculation means that calculates geometric constraint conditions regarding the position and/or shape of the object from the output of the camera system. , Calculate whether the intersection position of the line of sight from the two viewpoints of the camera system satisfies the geometric constraint, and calculate the intersection position that satisfies the geometric constraint and the geometric A device for recognizing the three-dimensional shape of an object, comprising: second calculation means for calculating three-dimensional coordinates of corresponding points of the object based on a constraint condition. 2. On page 1 of the claims, the first calculation means calculates a closed region formed by a set of planes circumscribing an object as a set of envelope surfaces of the object, and the object is a set of envelope surfaces of the object. A three-dimensional shape recognition device for an object, characterized in that the device is an arithmetic means that uses a property existing inside as the geometric constraint condition. 3. In claim 2, for each viewpoint direction, a hypothetical existence region is determined in the shape of a cone with the object image on the photographed image as the bottom shape and the viewpoint as the apex, and within this region, A three-dimensional shape recognition device for an object, characterized in that a logical product part common to all viewpoint directions is determined as a common hypothetical existence region, and the shape of the common hypothetical existence region is set as the envelope set. 4. In claim 3, the shape of the hypothetical existing region is described by a voxel model expressing the presence or absence of spatial grid points, and the common hypothesis is determined by performing a logical product operation for each unit voxel. A device for recognizing the three-dimensional shape of an object, characterized in that it determines the region of existence and describes a set of envelope surfaces of the object using a voxel model. 5. In claim 1, the second calculation means performs a correspondence on a pixel-by-pixel basis as each pixel of the corresponding point photographed image, and selects an optimal one from among candidates that satisfy the constraint condition. The evaluation function for selecting a suitable correspondence is the sum of the differences in brightness of the corresponding pixels, and the corresponding points are determined by finding the correspondence that minimizes this evaluation function. 3D shape recognition device. 6. A three-dimensional object according to claim 1, comprising display means for displaying the calculation result of the first calculation means and/or the calculation result of the second calculation means. Shape recognition device. 7. The approximate shape of the object to be recognized is obtained as a set of envelope surfaces, the object is photographed using horizontally placed left and right camera systems, and the area where the object is projected in the photographed image is defined as the object image. The brightness values of pixels in the object image portion on the scanning line of the image are extracted, and for each pixel of the object image, a straight line connecting the position of the pixel in real space and the viewpoint position is determined separately for the left and right sides.
The combination of pixels where the intersection of the left and right straight lines is inside the envelope set is calculated as a selection candidate, and when the left and right pixels are matched among these selection candidates, the sum of the differences in brightness values is the minimum. A method for recognizing the three-dimensional shape of an object, characterized in that a certain combination is selected, the shape by a set of envelope surfaces is corrected from the intersection information of the combination, and this is used as the final three-dimensional shape of the object. 8. In claim 7, an object is photographed by a camera system from a plurality of viewpoint directions, and for each viewpoint direction, a hypothetical existence region is formed in the shape of a cone with the object image as the base and the viewpoint as the apex. , and within this hypothetical existence area, find the logical product and partial area common to all viewpoint directions as a common hypothetical existence area, and let this common hypothetical existence area be the approximate shape of the set of envelope surfaces of the above object. A method for recognizing the three-dimensional shape of a characteristic object.