JPH0755469A - Multi-viewpoint stereoscopic image measuring method - Google Patents

Abstract

PURPOSE:To determine an initial value in a repetitive computation with good accuracy and to determine a relative positional relationship and actual spatial coordinates with a good accuracy in a multi-viewpoint stereoscopic image measuring method wherein stereoscopic image of a plurality of objects are measured from a plurality of viewpoints whose relative positional relationship is unknown and the relative positional relationship between the viewpoints and the actual spatial coordinates of the individual objects are determined on the basis of measured results. CONSTITUTION:A stereoscopic image is measured (Step 101), the probability density distribution of positional measuring errors of individual objects in individual viepoint coordinate systems is found (Step 103), the total product of the probability density distribution is found, and a composite probability density distribution is obtained (Step 104). An adaptable degree on the basis of the composite probability density distribution is defined, and an initial value is found by a hereditary algorithm including a selection process in such a way that a survival probability becomes higher the higher the adaptable degree is (Step 105). A value as a goal is computed finally by a most rapid descent method using the initial value (Step 106).

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a stereo image measuring method, and in particular, it performs stereo image measurement of a plurality of objects from a plurality of viewpoints whose relative positional relationship is unknown, and based on this measurement result, the relative position between the viewpoints. It relates to a relationship and a method for determining the real space coordinates of each object.

[0002]

2. Description of the Related Art In explaining the background art of the present invention, first, the principle of a stereo image measuring method will be explained, and then the measurement results (stereo data) obtained from a plurality of viewpoints whose positional relationships are unknown. A multi-view stereo image measurement method for integrating and determining the relative positional relationship between viewpoints and the real space coordinates of each object with high accuracy will be described.

<< Principle of Stereo Image Measurement >> The principle of stereo image measurement will be described with reference to FIG. Stereo image measurement projects an image of a three-dimensional space onto a two-dimensional space (image surface) by the left and right two cameras placed at the viewpoint, and measures the position of the measured point by the parallax between the two cameras. Is the way. Using x, y, and z as real space coordinates, the x axis and the y axis are set in the horizontal plane, the z axis is set in the vertical direction, and the origin of this coordinate system is the viewpoint O. Principal point F _L of the right and left cameras, F _R, respectively (-a, 0,0), in (a, 0,0). Origin O _L in the left image plane 2 and the right image plane 3, O _R, respectively corresponding principal point F _L, is spaced apart from F _R by f (focal length). The horizontal u-axis and the vertical v-axis on each image plane 2 and 3 are parallel to the x-axis and z-axis in the real space, and the optical axis 1 of each camera is parallel to the y-axis. Is. Therefore, the origins O _L and O _R of the image planes 2 and 3 are (-a, f, 0) and (a, f, 0), respectively, when expressed in real space coordinates.

Now, the point P (x, y, z) in the real space is
Point I on image planes 2 and 3_L(u_L, v_L), I _R(u_R, v_R) Projected on
Suppose At this time, v depending on the epipolar condition_L= V_R
Therefore, in the following, v_LAnd v_RExpressed as v without distinguishing
You △ I_LI_RP and △ F_LF_RBecause it is similar to P

[0005]

[Outer 1] The conversion to can be written as:

[0006]

[Equation 1] Therefore, by obtaining the image plane coordinates of the projection point of the object in the two stereo images, the real space coordinates of the object can be obtained using the equation (1).

An example of the configuration of a device for performing stereo image measurement is shown in FIG. In the stereo image measurement, two TV cameras 5 _L and 5 _R are used. The images captured by the television cameras 5 _L and 5 _R are photoelectrically converted while being scanned for each scanning line, and are converted into temporally continuous video signals. This video signal is sampled into millions of pixels per screen by the sampler in the digitizer 6, and the A / D converter in the digitizer 6 samples the intensity of light incident on each pixel. Is quantized into several hundred gradations as a density corresponding to. in this way,
The image is digitized via the digitizer 6, and is input and stored in the image memory 7 as information indicating the coordinates indicating the pixel position and the density there. The stored image information is input from the image memory 7 to the computer (CPU) 8,
Using features such as edges as clues

[0008]

[Outside 2] Is required. Then, from these values, three-dimensional coordinate values are obtained based on the above-mentioned principle.

<Multi-view stereo image measuring method> Multi-view stereo for integrating stereo data obtained from a plurality of viewpoints whose relative positional relationship is unknown to determine relative positional relationship between viewpoints and real space coordinates of each object An image measuring method is disclosed in Japanese Patent Application No. 4-238586. Here, the technique will be briefly described.

A model of multi-view stereo image measurement is shown in FIG. In this example, m viewpoints and n objects are provided. Let the true coordinates of the j-th object in the i-th viewpoint coordinate system (origin is O _i , and each coordinate axis is represented by x _i , y _i , z _i )

[0011]

[Outside 3] The measurement coordinates of the j-th object measured from the origin O _i of this coordinate system as the viewpoint

[0012]

[Outside 4] Express with. A rotation matrix representing the conversion from the first viewpoint coordinate system to the i-th viewpoint coordinate system, using the first viewpoint coordinate system (origin is represented by O ₁ ) as the reference coordinate system.

[0013]

[Outside 5] The matrix that represents the translation vector

[0014]

[Outside 6] And these are defined as follows.

[0015]

[Equation 2] Rotation matrix

[0016]

[Outside 7] Is a 3 × 3 matrix and is composed of 9 elements, of which the three independent variables are α _i , β _i , and γ _i .
Therefore, as a matrix showing these three rotation parameters,

[0017]

[Equation 3] Is defined. By this definition,

[0018]

[Outside 8] The relation of is expressed as in equation (2).

[0019]

[Equation 4] As is clear from the equation (1), the coordinate measurement accuracy of the object in the real space depends on the measurement error in the image plane coordinates. The measurement error in the image plane

[0020]

[Equation 5] Assuming that a normal distribution with independent standard deviation σ is assumed as its statistical property,

[0021]

[Outside 9] The probability density distribution s _{i, j}

[0022]

[Equation 6] It is represented by.

Measurement error on the image plane

[0024]

[Outside 10] If is small, the measurement error on the image plane and the measurement error in the real space caused by it

[0025]

[Equation 7] The relationship with and is expressed by equation (4).

[0026]

[Equation 8] here,

[0027]

[Equation 9] And specifically,

[0028]

[Equation 10] Is. The probability density distribution s _{i, j} of the measurement error in the image plane coordinates expressed by the equation (3) is calculated by the equations (2), (4) and (5) as follows in the real space coordinates. It can be converted into probability density distribution of measurement error.

[0029]

[Equation 11] The probability density distribution of the measurement error in the real space coordinates of the j-th object at the i-th viewpoint is obtained by equation (7), and this probability density distribution is combined for all m viewpoints and all n objects, the combined probability The density distribution S is

[0030]

[Equation 12] Is represented. Where E is

[0031]

[Equation 13] Is.

The condition for maximizing the composite probability density distribution S expressed by the formula (8) is a condition for minimizing E expressed by the formula (9), and this E is minimized.

[0033]

[Outside 11] Is the relative position of the viewpoint to be sought and the (first
This is the position (expressed in the viewpoint coordinate system). To calculate these values analytically,

[0034]

[Equation 14] Must be solved simultaneously, but its calculation is difficult,
The steepest descent method is used instead of the analytical one.

In the steepest descent method, the function

[0036]

[Equation 15] As opposed to

[0037]

[Equation 16] Point sequence

[0038]

[Outside 12] Is systematically generated, and the minimum point of the function e is actually numerically obtained by this repetition. In this method, the current point (kth point)

[0039]

[Outside 13] Vector of steepest descent at

[0040]

[Equation 17] Ask for this vector

[0041]

[Outside 14] Using a certain step width ε,

[0042]

[Equation 18] Produces the next new point, the k + 1 th point.

As a result, the value of the evaluation function e can be reduced most each time the iterative calculation progresses by one. At the minimum point of the function e,

[0044]

[Formula 19] Therefore, iterative calculation is performed until this condition is satisfied, and the minimum point of e is actually calculated numerically.

Using this steepest descent method, a parameter vector that minimizes equation (9) is calculated. The parameter vector to find is

[0046]

[Outside 15] Therefore,

[0047]

[Equation 20] Becomes Also,

[0048]

[Equation 21] You should say In addition,

[0049]

[Equation 22] Is. It converged by this iterative calculation

[0050]

[Outside 16] Of the most certain point of view

[0051]

[Outside 17] And

In various iterative methods including the steepest descent method, it is necessary to give an initial value in advance. In the case of the multi-view stereo image measuring method, for example, the value obtained by the method disclosed in Japanese Patent Laid-Open No. 4-118504 can be used as the initial value. The one disclosed in Japanese Patent Laid-Open No. 4-118504 obtains a relative positional relationship between two viewpoints using a coordinate conversion parameter that minimizes the sum of squares of the Euclidean norm of the shift amount vector in the real space. It is something to try.

[0053]

As described above, in the conventional multi-view stereo image measurement method, the relative positional relationship between the viewpoints and the actual object are calculated by iterative calculation under the condition that the equation (9) is minimized. Seeking a spatial difference table. Here, the problems of this method will be examined. Equation (9) is a multivariable nonlinear function, and it is difficult to infer what shape the curved surface of the function is. Here, assuming a function having a shape as shown in FIG. 4, a problem in iterative calculation will be described.

For example, when p1 in FIG. 4 is set as an initial value and the iterative calculation is started from here, instead of the global optimum solution q2,
It converges to the local optimum solution q1. The same is true when the initial value is p3, and in order to converge to the global optimum solution q2, iterative calculation must be performed from a position such as p2 in the figure. That is, in the iterative calculation such as the steepest descent method, the convergent value depends on the initial value of the iterative calculation,
There is a problem that it may fall into a local optimal solution instead of the global optimal solution. As a method of solving this problem, there is a method of calculating approximate values of a viewpoint position and a direction and an object position using sensor data or the like, and using the approximate values as initial values for repeated calculation. However, this method has a problem in that the size and cost of the device are increased.

An object of the present invention is to determine an initial value that does not fall into a local optimum value for an initial value of repeated calculation which becomes a problem in multi-view stereo image measurement, and use this to determine the relative positional relationship between viewpoints. Another object of the present invention is to provide a method capable of determining the real space coordinates of each object with high accuracy.

[0056]

A multi-view stereo image measuring method according to the present invention captures stereo images of a plurality of objects at a plurality of viewpoints whose relative positional relationships are unknown, and arithmetically processes each stereo image. For each of the viewpoints, the real space coordinate values of the objects described by the coordinate system at the viewpoint are obtained, and the relative positional relationship between the viewpoints is determined using the real space coordinate values of the coordinate system at the viewpoint. In a multi-view stereo image measurement method that determines the coordinate transformation parameters to be determined and the real space coordinates of each object in the reference coordinate system by an iterative method, the initial value used for the iterative method should be optimized corresponding to the initial value. A one-dimensional array in which each parameter is assigned to each element is called an individual, and a value is determined for each of the viewpoints by using the real space coordinates based on the viewpoint and the parameters. When defining the fitness, a value of each of the parameters to be assigned to each element is given by a random number, and a first step of generating a plurality of individuals as parents of the first generation to form an individual group; Select any two individuals from the parent population formed in the step of, and replace each element between the two individuals with respect to the parameters randomly selected from the above parameters The second step of generating new individuals and repeating the process of generating the two new individuals to form a child population, and the new population formed in the second step, A third step of adding a random number to a parameter randomly selected from the parameters to form a new child population, the child population formed in the second step, and Third work In the population selected as a child and the individuals selected from the parent population in the descending order of fitness, a selection process in which the higher the fitness is, the higher the survival probability is, By performing a fourth step of leaving the same number of individuals as the parent individual group and deleting other individuals, by the fitness to the individuals left in the fourth step and a predetermined convergence condition. A fifth step of determining convergence, and when it is determined not to converge in the fifth step, the individual left in the fourth step is set as a new parent individual, and the second to It corresponds to the individual left in the fourth step when it is determined by the genetic algorithm consisting of the sixth step of repeating the fourth step and the convergence in the fifth step. The parameter is used as the initial value.

[0057]

By applying a method generally called a genetic algorithm to the initial value search problem, an initial value that gives a global optimum solution can be obtained, and the relative positional relationship between each viewpoint and the real space coordinates of the object are increased. You will be able to determine the accuracy.

Here, the generally used genetic algorithm will be outlined. The genetic algorithm is a method considered as a model that imitates the inheritance and evolution of an organism. First, the set corresponding to the population of a certain species

[0059]

[Outside 18] think of. here

[0060]

[Outside 19] Indicates the individual number. Each individual

[0061]

[Outside 20] Is composed of a one-dimensional array, and each element of this array is coded by "0" or "1", that is, a binary representation. In genetic algorithms, this coding is generally the most difficult problem, and therefore various methods are used for coding. The number of iterations of the algorithm is called "generation" and is represented by t. T generation population

[0062]

[Equation 23] Is given to each individual

[0063]

[Outside 21] , A function that is regarded as the fitness of these individuals to the environment

[0064]

[Outside 22] To calculate. At this time, the basic operations of the genetic algorithm are performed, that is, crossing, mutation and selection,
+1 generation population

[0065]

[Equation 24] To form.

Next, the basic operation of the genetic algorithm will be described. The operation contents of crossover and mutation are schematically shown in FIGS. 5 (a) and 5 (b). In this figure,
Each individual consists of 6 elements, and the shaded squares and the outlined squares are

[0067]

[Outside 23] Represents the element of. Crossing is an operation in which two individuals are selected from a group of individuals, a point at which the two individuals are to be intersected is randomly selected, and symbols (elements) are exchanged before and after the intersection. In the example shown in FIG. 5 (a), the center of the array is an intersection, and the elements of one individual and the elements of the other intersect with each other with this intersection as the front and rear.

Mutation is an operation in which elements are randomly selected for the original individual and the contents of the selected elements are rewritten randomly. In FIG. 5 (b), p3 and p5
Are selected and rewritten as r3 and r5, respectively.

The selection is an operation in which the higher the fitness of an individual represented by the above-mentioned function is, the higher the survival probability is, and the individual to be left in the next generation is selected. Various survival probabilities are used depending on the problem.

By repeating these operations, the population is artificially "evolved" and the fitness as a whole is improved. Then, this operation is repeated until a certain end condition such as desired accuracy is satisfied.

The method based on the genetic algorithm can be applied to a problem such that an analytical solution cannot be obtained and a problem where the gradient of the objective function cannot be obtained. On the other hand, this method has a problem that this method is not efficient in terms of convergence. Therefore, in the present invention, instead of converging only by the genetic algorithm, the genetic algorithm is applied to the initial value search of the optimization method, and the result is applied to the iterative method to obtain the final convergence result. I have to. We also proposed a new method for each process of genetic algorithm coding, crossover, and mutation.

In the present invention, the steepest descent method can be used as an iterative method for finally determining the relative positional relationship between viewpoints and the real space coordinates of the object. Regarding the probability density distribution of the position measurement error of each object in the coordinate system based on each viewpoint, when the total product of the probability density distributions for each viewpoint and each of the objects is the combined probability density distribution, each should be optimized. A composite probability density distribution that explicitly includes a parameter can be the fitness in the present invention.

[0073]

Embodiments of the present invention will now be described with reference to the drawings. FIG. 6 is a flowchart showing the procedure of the multi-viewpoint stereo image measuring method according to the embodiment of the present invention.
Is a flowchart showing an initial value searching procedure in this procedure.

In this embodiment, stereo images of a plurality of objects are measured from a plurality of viewpoints whose relative positional relationship is unknown, and the relative positional relationship between the viewpoints and the real space coordinates of each object are determined from the measurement results. However, up to the point where the measurement result of the stereo image measurement is obtained, it is the same as the above-mentioned conventional example. That is, assuming that the number of viewpoints is m and the number of objects is n, stereo image measurement is performed at each viewpoint (step 101), and the measured value of the real space coordinates of the object represented by the coordinate system for each viewpoint. To get Then, of the measurement results obtained from different viewpoints, the one corresponding to the same object is found (step 102). As a method of associating objects between viewpoints, for example, Japanese Patent Laid-Open No. 5-12
There is one described in Japanese Patent No. 430. Then, from the equation (7), the probability density distribution s _{i, j} of the measurement error in the real space coordinates of the j-th object at the i-th viewpoint is obtained (step 10
3) According to equation (8), this probability density distribution is combined for all m viewpoints and all n objects to obtain a combined probability density distribution S (step 104).

Obtaining the relative positional relationship between viewpoints and the real space coordinates of each object (represented by the reference coordinate system) is to obtain the value of each parameter that maximizes the composite probability density distribution S. This is to find the value of each parameter that minimizes E expressed by the equation (9).

Using the first viewpoint coordinate system as the reference coordinate system,

[0077]

[Outside 24] Extreme conditions for

[0078]

[Equation 25] From, the object position that minimizes E in equation (9)

[0079]

[Outside 25] Is the unknown parameter

[0080]

[Outside 26] Including the form, it is given by the following equation.

[0081]

[Equation 26] This equation (11) represents that if the above-mentioned unknown parameter relating to the viewpoint position is given, the object position at that time is obtained, and by using this condition,
It is shown that the parameter in (9) can be reduced to only the parameter related to the viewpoint position.

In this embodiment, the parameters relating to the viewpoint position

[0083]

[Outside 27] A target value is searched for by using a genetic algorithm (step 105), and then a final value is calculated using the steepest descent method as in the above-described conventional example (step 106).

Next, the initial value search by the genetic algorithm will be described with reference to the flowchart of FIG.

<< Coding >> First, coding is performed to generate an individual (step 111). Parameters to optimize

[0086]

[Outside 28] Is assigned to an element of a one-dimensional array to generate an individual. An area (range) that each parameter can take is set, and a uniform random number is generated in that area to generate a first-generation parent individual. Here, the previously used "0" and "1"
Instead of giving each element a value in its binary representation by
A value is given to each element by a real number. By doing so, the value of each element and the parameter value can be made to coincide with each other, and at the time of crossing, it is possible to cross each parameter independently. 8 (a) is 1
A schematic example of the arrangement of one individual is shown, and the subscript m represents the total number of viewpoints. Then, it is judged whether or not the number of generated individuals has reached a predetermined number (step 112), and if not, step 111 is repeated to finally form an individual group consisting of Q individuals.

<< Crossover >> Next, two individuals are appropriately selected from the Q individuals forming the individual group, and intersections are performed to replace the elements between the two individuals with respect to the randomly selected parameters (step 113). Get two children. This process is repeated to generate a population of Q individual children. Here, since each parameter is independent, a method of randomly selecting whether or not to replace each parameter was used instead of the intersection in a general genetic algorithm. When optimizing a multivariable function in which each parameter is independent of each other as in the present embodiment, the method described here is considered to be more effective than the intersection using the general method described above. . By doing so, the intersection can be performed only by changing the combination without breaking the parameter values. FIG. 8B schematically shows the processing of this intersection, and the second element and the fourth element are selected as the objects of the intersection.

<< Mutation >> Next, for all the individual children obtained by crossing, elements are randomly selected, and a random number is added to the selected elements to perform mutation, and Q new children are newly added. The individual population of is generated (step 114).
As a method of mutation, ± 10% of the area of each parameter
The method used is to generate a random number in the range of and to add the random number to the original parameter value. Compared with the new rewriting method in the general genetic algorithm described above, there is an effect of leaving the effect on the original parameter value and an effect of leaving the effect of the parameter having a good value. On the other hand, the effect of bad parameters will remain, but we used this method, paying attention to the effect of good parameters.

<< Selection >> Next, from the total of 2Q individual children obtained by crossing and mutation, and the individual selected from the parent population, survival according to fitness is performed. By the selection process with changed probability, Q individuals are left and the others are deleted (step 115). As a method of selecting individuals from the parent group of individuals, a method of selecting a predetermined number of individuals in descending order of fitness calculated from the value of the evaluation function E is used. Here, 10% of the individuals with good fitness from the parent Q individuals were used for this process.

The fitness will be described. For the r-th individual z ^{r in} the population of R individuals to be selected, the value of the evaluation function E given by equation (9) is E (z
^r ). At this time, the fitness w (z ^r ) of the r-th individual is defined by the following equation.

[0091]

[Equation 27] Then, the higher the fitness, the higher the probability of survival, and the surviving Q individuals are set as a new parent population. Here, note that the total sum of the fitness values of the R individuals is 1 from the equation (11), and the section from 0 to 1 is divided into individual individuals with a width proportional to the fitness, and 0 to 1 is divided. We used a method to generate uniform random numbers in the interval and leave individuals with a width at the place of the random number. Individuals with higher fitness have intervals [0,
Since there is a large width within 1), the probability of remaining in the next generation, that is, the survival probability, increases. In addition, by using this method, it is possible that an individual with low fitness is not unconditionally deleted due to low fitness, but rather contains a parameter with a good value. There is a possibility that even such individuals will remain in the next generation.

An example of selection is included in FIG. 8 (c). FIG. 8 (c) shows how 4 individuals are selected from 6 individuals. When four uniform random numbers were generated in the interval from 0 to 1, one was generated at the width of w (z ¹ ), two at the width of w (z ⁴ ), and 2 at w (z ⁶ ). Since one random number was generated in each position in the width, in the next generation, ^one z ¹ individual, 2 z ⁴ individuals, 1 z ⁶ individual survived, and other individuals were weeded out. , Erased.

<< Convergence Judgment >> When the selection process is completed, a convergence judgment is made (step 116). If the convergence condition is satisfied, the process goes to the next step. If the convergence condition is not satisfied, the process returns to step 113, The steps 113 to 115 of crossing, mutation, and selection are repeated. Here, it is assumed that the convergence condition is satisfied when 10% of the individuals remaining in the selection from the one with the highest fitness become the same individuals.

When the convergence condition is satisfied, each element is read out from the individual having the highest fitness at that time, and as described above, the element value is used as an initial parameter and the final value is determined by a method such as the steepest descent method. Then, the relative positional relationship between each viewpoint and the real space coordinate value of each object are determined.

<< Specific Example >> Next, an example of an experiment specifically carried out on this embodiment will be described. Camera interval (2a)
Was 40 mm and the focal length (f) was 4 mm, and two-dimensional multi-stereo image measurement was performed based on this example. In this experiment, the positions of five objects were measured from the first and second viewpoints, and the total number of individuals Q in the genetic algorithm for initial value search was set to 50. Table 1 shows the measurement results of five objects in the stereo image measurement from each viewpoint.

[0096]

[Table 1] Also, in the process of initial value search by the genetic algorithm, the rotation matrix

[0097]

[Outside 29] The process of convergence of the error of the parameter γ ₂ of is shown in FIG.

[0098]

[Outside 30] FIG. 10 shows a process in which the errors of the parameters t _x2 and t _y2 , which are the components of, converge. 9 and 10, the individuals with the highest fitness in each generation are shown. Table 2 shows the true value of each parameter, the initial value obtained by the genetic algorithm in this embodiment, and the value obtained by the steepest descent method based on this initial value.

[0099]

[Table 2] As shown in Table 2, it is confirmed that the initial value of each parameter is obtained with sufficient accuracy by using the algorithm of the present embodiment, and that the value is further improved by the steepest descent method.

Although the embodiment of the present invention has been described above, the genetic algorithm in this embodiment has various advantages as described above as compared with the conventional general genetic algorithm, and has a multi-view stereo image. In addition to the initial value search in measurement, it can be widely applied to the initial value search problem when the optimization condition of the evaluation function is obtained by iterative calculation because an analytical solution is not generally obtained.

[0101]

As described above, the present invention obtains an initial value with sufficient accuracy by searching the initial value used in the iterative method, which is an optimization method in multi-view stereo image measurement, with a genetic algorithm. Thus, there is an effect that the relative positional relationship between viewpoints and the real space coordinates of each object can be determined reliably and with high accuracy.

[Brief description of drawings]

FIG. 1 is a diagram illustrating the principle of stereo image measurement.

FIG. 2 is a block diagram showing an example of an apparatus used for stereo image measurement.

FIG. 3 is a diagram illustrating a multi-view stereo image measuring method.

FIG. 4 is a diagram illustrating a convergence value of an evaluation function according to the steepest descent method.

FIG. 5 is a diagram illustrating a basic operation in a general genetic algorithm, and FIGS. 5A and 5B are diagrams illustrating crossover and mutation, respectively.

FIG. 6 is a flowchart showing a procedure of a multi-viewpoint stereo image measuring method according to an embodiment of the present invention.

FIG. 7 is a flowchart showing an initial value search procedure.

FIG. 8 is a diagram for explaining the genetic algorithm in the procedure of FIG. 7, wherein (a), (b), and (c) are diagrams for explaining coding, intersection, and selection, respectively.

FIG. 9 is a characteristic diagram showing a convergence process of a rotation parameter.

FIG. 10 is a characteristic diagram showing a process of converging a translation parameter.

[Explanation of symbols]

1 optical axis 2 left image plane 3 right image plane 5 _L , 5 _R camera 6 digitizer 7 image memory 8 computer 101-106, 111-116 steps

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Mitsuhiro Tachida 1-1-6 Uchisaiwaicho, Chiyoda-ku, Tokyo Nihon Telegraph and Telephone Corporation

Claims (2)

[Claims] 1. Stereo images of a plurality of objects are photographed at a plurality of viewpoints whose relative positional relations are unknown, and each stereo image is arithmetically processed to describe each viewpoint by a coordinate system at the viewpoint. The real space coordinate value of each of the objects is obtained, and the real space of each object in the reference coordinate system and the coordinate conversion parameter that determines the relative positional relationship between the viewpoints using the real space coordinate value of the coordinate system of each of the viewpoints. In a multi-view stereo image measurement method for determining coordinates and an iterative method, an initial value used for the iterative method is a one-dimensional array in which each parameter to be optimized corresponding to the initial value is assigned to each element as an individual. When defining the fitness whose value is determined using the real space coordinates with respect to the viewpoint and the parameters for each viewpoint, it should be assigned to each element. The first step of generating a plurality of individuals of the first generation parent as an individual group by giving the values of the respective parameters by random numbers, and the parent individual group formed in the first step 2 arbitrary individuals are selected from the parameters, and each element is exchanged between the 2 individuals with respect to the parameters randomly selected from the respective parameters to generate 2 new individuals. A second step of repeating the generation process to form a child population, and a new population formed in the second step with respect to parameters randomly selected from among the parameters A third step of adding a random number to form a new child population, a child population formed in the second step, and a child population formed in the third step And the individual that is the parent For individuals selected from the group in the descending order of fitness, a selection process is performed in which the higher the fitness is, the higher the survival probability is, and the same number of individuals as the parent population is left. A fourth step of erasing other individuals, a fifth step of judging convergence based on the fitness and predetermined convergence conditions for the individuals left in the fourth step, and a fifth step When it is judged that it has not converged,
The individual left in the fourth step is set as a new parent individual, and the sixth step of repeating the second to fourth steps is performed, and the sixth step is determined by a genetic algorithm. A multi-view stereo image measurement method, wherein a parameter corresponding to an individual left in the fourth step when it is determined to be converged is set as the initial value. 2. The iterative method is a steepest descent method, and relates to a probability density distribution of position measurement error of each object in a coordinate system based on each viewpoint, with respect to each of the viewpoint and each object. The multi-view stereo image measurement method according to claim 1, wherein when the total product of the probability density distributions is a combined probability density distribution, the fitness is based on the combined probability density distribution explicitly including the parameters. .