JP5928240B2 - Three-dimensional shape interpretation apparatus and program - Google Patents

Description

  The present invention relates to a three-dimensional shape interpretation apparatus and program, and more particularly to a three-dimensional shape interpretation apparatus and program for interpreting a three-dimensional shape from three-dimensional point cloud data.

  2. Description of the Related Art Conventionally, a method is known in which a three-dimensional parametric shape model of an object is defined, an association that minimizes the distance from a point cloud is obtained by searching, and the position and orientation of the object are recognized (Non-Patent Document 1). ).

  A bird's-eye view image generation method is known in which a vehicle equipped with a camera and a laser scanner travels in an urban area and acquires a high-definition three-dimensional point cloud of a road (Patent Document 1).

  A point group data processing apparatus is known that calculates a contour of a three-dimensional object by performing plane fitting for each local region from a point group and connecting them (Patent Document 2).

  There is known an automatic three-dimensional shape generation apparatus for determining a shape by applying a triangular polygon to a point group and applying a plane equation or a curved surface equation to the point polygon (Patent Document 3).

  In addition, the shape to be fitted is expressed by an equation in the Conformal Geometric Algebra space, and by estimating the coefficients, different shapes such as “straight and circle” and “plane and sphere” are expressed by a common model, and the shape is determined based on the fitting result. Classification methods are known (Non-Patent Documents 2 and 3).

JP 2012-018170 A JP 2012-13660 A Japanese Patent Laid-Open No. 2004-272459

P.J.Besl and H.D.McKay, "A method for registration of 3-D shapes", Trans. PAMI vol.14, no.2, pp.239-256, 1992. D. Hildenbrand, “Geometric computing in computer graphics using conformal geometric algebra”, Computers and Graphics, vol.29 (2005), no.5, pp.802-810. Phanmintun, Kanta Tachibana, Daihiro Yoshikawa, Takeshi Furuhashi, "Proposal and Application of Approximation Method Using Conformal Geometric Algebra", Proceedings of the 1st Computational Intelligence Study Group, pp. 29-36, 2011.

  As described above, there are many prior literatures relating to object detection from a point cloud, but complicated calculation processing is required, and implementation is not easy.

  In the method described in Non-Patent Document 1, an object having a complicated shape can be associated with a point cloud. However, it is necessary to create a shape model individually for each object in advance, and geometric calculation and search processing are complicated. There is a problem of becoming.

  The method described in Patent Document 1 is a point cloud acquisition method and a road image generation method. However, in order to recognize a structure as a structure, there is a problem that a recognition process is separately required. .

  The technique described in Patent Document 2 has a problem in that it is necessary to determine noise for local plane fitting and local plane connection processing is necessary.

  The technique described in Patent Document 3 has a problem that the fitting error must be evaluated for each equation because the plane and the curved surface are defined by separate equations.

  In the techniques described in Non-Patent Documents 2 and 3 above, it is possible to extract a specific shape from a point cloud by using Conformal Geometric Algebra. However, since the least square method is used for shape fitting, when applied to a three-dimensional point group including noise, the accuracy of the position and the object identification performance deteriorate due to the influence of the outlier. When applied to a three-dimensional point group, there are two types of objects that can be represented by the common model described in Non-Patent Documents 2 and 3 above: a plane and a sphere, and straight lines, circles, cylinders, and the like have different expressions. (Conformal Geometric Algebra space requires higher-order coefficients), so there is a problem that it cannot be handled.

  The present invention has been made to solve the above-described problems, and provides a three-dimensional shape interpretation apparatus and program capable of interpreting a three-dimensional shape in three-dimensional point cloud data with simple processing. With the goal.

  In order to achieve the above object, a three-dimensional shape interpretation apparatus according to the present invention includes a point cloud acquisition unit that acquires point cloud data including a plurality of points having three-dimensional coordinates, and the plurality of points of the point cloud data. Is converted into a CGA (Conformational Geometric Algebra) space, and any three or four of the plurality of points and the infinite base of the point cloud data converted onto the CGA space. Creating means for generating geometric structure data representing any one of a plurality of predetermined three-dimensional shapes by selecting a combination consisting of the above and combining the selected combinations using an outer product; For each of the plurality of points of the point cloud data converted to, the distance between the point and the geometric structure data created by the creating means is obtained by inner product calculation Fitness calculation means for calculating the fitness of the geometric structure data based on the distance obtained for each of the plurality of points; and the geometric structure based on the fitness of the geometric structure data. An update means for updating the combination included in the data, and a three-dimensional representation represented by the geometric structure data by repeating the update by the update means and the calculation by the fitness calculation means so that the fitness is maximized And an optimization means for outputting the shape.

  The program according to the present invention includes a point cloud acquisition unit that acquires point cloud data including a plurality of points having three-dimensional coordinates, and the plurality of points of the point cloud data are stored on a CGA (Conformational Geometry Algebra) space. A space converting means for converting to the CGA space, selecting a combination of any three or four of the plurality of points of the point cloud data converted to the CGA space and a base at infinity, and the selected combination Creating means for creating geometric structure data representing any of a plurality of predetermined three-dimensional shapes by combining outer products, and the plurality of points of the point cloud data converted on the CGA space For each of the points, the distance between the point and the geometric structure data created by the creating means is obtained by an inner product operation, and each of the plurality of points is assigned. A fitness calculation means for calculating the fitness of the geometric structure data based on the distance determined in the step, and updating the combination included in the geometric structure data based on the fitness of the geometric structure data. Update means that performs the update, and the optimization means that outputs the three-dimensional shape represented by the geometric structure data by repeating the update by the update means and the calculation by the fitness calculation means so that the fitness is maximized. It is a program to make it function.

  According to the present invention, the point cloud acquisition unit acquires point cloud data including a plurality of points having three-dimensional coordinates. The plurality of points of the point cloud data are converted into a CGA (Conformational Geometry Algebra) space by a space conversion means.

  Then, the creation unit selects any three or four combinations of the plurality of points and the infinity base of the point cloud data converted into the CGA space, and the selected combination is selected. The geometric structure data representing any of a plurality of predetermined three-dimensional shapes is created by combining the outer products. For each of the plurality of points of the point cloud data converted on the CGA space by the fitness calculation means, a distance between the point and the geometric structure data created by the creation means is obtained by inner product calculation, The fitness of the geometric structure data is calculated based on the distance obtained for each of the plurality of points.

  Then, the combination included in the geometric structure data is updated by the updating means based on the fitness of the geometric structure data. The updating by the updating unit and the calculation by the fitness calculating unit are repeated so that the fitness is maximized by the optimization unit, and the three-dimensional shape represented by the geometric structure data is output.

  In this way, a plurality of points of point cloud data are converted into CGA space, a combination of a plurality of points or infinite bases is selected and combined using an outer product, and geometric structure data representing a three-dimensional shape To calculate the fitness of the geometric structure data by calculating the distance between the point of the point cloud data and the geometric structure data by the inner product operation, and the geometric structure data based on the fitness of the geometric structure data And updating the geometric structure data and calculating the fitness so that the fitness is maximized, and outputting the three-dimensional shape represented by the geometric structure data. The three-dimensional shape in the three-dimensional point cloud data can be interpreted.

  The creating means according to the present invention selects any three or four of the plurality of points and the infinite base of the point cloud data transformed into the CGA space as a combination of genes, A plurality of chromosomes including a combination of four or four genes is created, and for each chromosome, the combination of the genes of the chromosome is combined using an outer product to create the geometric structure data, and the fitness calculation means includes: For each of the plurality of chromosomes, the fitness of the geometric structure data of the chromosome is calculated, and the updating means is based on the fitness of each of the plurality of chromosomes calculated by the fitness calculation means. The plurality of chromosomes can be updated by selecting two of the chromosomes and crossing the genes of the two selected chromosomes. In this way, the three-dimensional shape in the three-dimensional point cloud data can be accurately interpreted using the genetic algorithm.

  The geometric structure data includes geometric structure data representing a plane obtained by combining a combination of the selected three points and an infinite base with an outer product, the selected three points, and the selected 3 Geometric structure data representing a curved surface obtained by combining combinations of points added based on two points with an outer product, and three points that are changed so that the height components of the selected three points match Geometric structure data representing a cylinder obtained by combining combinations by outer products, and geometric structure data representing straight lines obtained by combining combinations of the selected two points and an infinite base by outer products. .

  The fitness calculation means of the present invention obtains the distance between the point and the geometric structure data created by the creation means for each of the points of the subset of the point cloud data, and calculates the points of the subset And calculating the fitness of the geometric structure data on the basis of the distance obtained for each of the parameters, and determining the subset to be different each time the calculation by the fitness calculation means is repeated. Can do. As a result, it is possible to interpret the three-dimensional shape in the three-dimensional point group data while suppressing an increase in the amount of calculation.

  The point cloud acquisition means according to the present invention acquires point cloud data consisting of a plurality of points having three-dimensional coordinates around the vehicle, and the optimization means updates the update so that the fitness is maximized. The update by the means and the calculation by the fitness calculation means can be repeated to output a three-dimensional shape represented by the geometric structure data existing around the vehicle.

  The three-dimensional shape interpretation apparatus according to the present invention further includes positioning means for positioning the position information of the vehicle, and the three-dimensional coordinates of each point of the point cloud data are the position information of the vehicle determined by the positioning means. The three-dimensional coordinates of the real coordinate system obtained using can be used.

  The three-dimensional shape interpretation apparatus may further include a map registration unit that registers the three-dimensional shape represented by the geometric structure data output by the optimization unit on a map.

  The creation means according to the present invention repeatedly creates the geometric structure data, and the plurality of points of the point cloud data excluding the combination of the points included in the geometric structure data output last time by the optimization means. The geometric structure data can be generated by selecting any three or four combinations from the points and the base at infinity.

  As described above, according to the three-dimensional shape interpretation apparatus and program of the present invention, a plurality of points of point cloud data are converted on the CGA space, a combination of a plurality of points or infinite bases is selected, and the outer product is selected. Are used to create geometric structure data that represents a three-dimensional shape, and the distance between the point cloud data points and the geometric structure data is obtained by the inner product operation to calculate the fitness of the geometric structure data. The geometric structure data is updated based on the fitness of the geometric structure data, and the update of the geometric structure data and the calculation of the fitness are repeated so that the fitness is maximized. By outputting the three-dimensional shape to be expressed, it is possible to obtain an effect that the three-dimensional shape in the three-dimensional point cloud data can be interpreted with a simple process.

It is a block diagram which shows the three-dimensional shape interpretation apparatus which concerns on embodiment of this invention. It is a figure for demonstrating CGA space. It is a figure for demonstrating the inner product between bases. (A) A figure showing a sphere, (B) A figure showing a plane, (C) A figure showing a circle, and (D) A figure showing a straight line. It is a figure for demonstrating the process at the time of generation change. It is a flowchart which shows the content of the three-dimensional shape interpretation process routine in the three-dimensional shape interpretation apparatus which concerns on embodiment of this invention. (A) The figure which shows the example of the one image input with the stereo camera, (B) The figure which shows the example which displayed the map where the structure was registered.

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

  As shown in FIG. 1, a three-dimensional shape interpretation apparatus 10 according to an embodiment of the present invention includes an imaging apparatus 12 that captures an image of the front of the host vehicle including a road area and outputs an image, and a GPS that measures the position of the host vehicle. A computer that executes a process of extracting the three-dimensional shape of the structure ahead of the host vehicle and registering it on the map data based on the positioning device 14 including the image obtained from the imaging device 12 and the positioning information by the positioning device 14 16.

  The imaging device 12 is mounted on a vehicle, images the front of the host vehicle including a road area, and outputs a stereo image. The imaging device 12 captures the front of the host vehicle, generates two image signals (not shown), and an A / D conversion unit that A / D converts the image signals generated by the two imaging units. (Not shown) and an image memory (not shown) for temporarily storing the A / D converted image signal. The imaging device 12 outputs the two captured images to the computer 16 as stereo images.

  The positioning device 14 measures the position of the host vehicle in the real coordinate system using a GPS device mounted on the host vehicle, and outputs the measurement result to the computer 16.

  The computer 16 includes a CPU, a RAM, a ROM that stores a program for executing a later-described three-dimensional shape interpretation processing routine, and a bus that connects these. When the computer 16 is described in terms of functional blocks divided for each function realization means determined based on hardware and software, a point cloud acquisition unit 18, a CGA conversion unit 20, a chromosome creation unit 22, a fitness calculation unit 24, the optimization part 26, the update part 28, the map registration part 30, and the memory | storage part 32 which memorize | stores the map data showing the road plane of a real coordinate system.

  The point group acquisition unit 18 extracts edge points from the stereo image captured by the imaging device 12, and obtains the three-dimensional position of the edge points with reference to the imaging device 12 from the parallax obtained by stereo matching. In addition, the point cloud acquisition unit 18 obtains the positioning information obtained by the positioning device 14 corresponding to each of the plurality of frames from the plurality of three-dimensional points obtained for the stereo images of the plurality of frames within a predetermined distance. By using and integrating, a three-dimensional point group having three-dimensional coordinates in the real space is obtained.

  Note that the edge points in the image may be extracted using a conventionally known method, for example, by the Canny algorithm.

  Further, the position of the three-dimensional point with reference to the imaging device 12 can be obtained by a method described in the following document using a calibrated camera.

References: M Bertozz, A Broggi, “Stereo Inverse Perspective Mapping: theory and applications”, image Vis. Comput. , Vol. 8, no. 16, pp. 585-590, 1998.

  Note that the point cloud may be acquired using a laser radar instead of the imaging device 12 which is a stereo camera.

The CGA conversion unit 20 converts each point of the three-dimensional point group into a conformational geometrical geometry (CGA) space. First, a three-dimensional point (x ₁ , x ₂ , x ₃ ) is expressed by the following formula (1) using _three bases (e ₁ , e ₂ , e ₃ ).

Then, each point of the three-dimensional point group is expressed by the following equation (2) on the five-dimensional CGA space to which two bases (e _∞ , e _O ) are added (see FIG. 2).

  In the converted CGA space, there is an advantage that a figure passing through the point group is obtained by outer product calculation (Outer product), and a distance between figures is obtained by inner product calculation (inner product). By using this property, it is easy to fit a geometric structure to a point cloud.

  Details of the conformational geometry are described in the following documents.

  Reference: Dorst, D.C. Fontijne, and S.M. Mann, Geometric Algebra for Computer Science-An Object-oriented approach to Geometry (Elsevier-Morgan Kaufmann Publishers, Sanco France 7)

  Here, an inner product operation, an outer product operation, and a dual operation in the CGA space will be described.

  First, the inner product between bases is expressed by the following equation (3).

  The inner product operation of the above equation (3) is represented in a table as shown in FIG.

The dual operation ( ^* ) is expressed by the following equation (4).

  In the outer product calculation (演算), the following equations (5) to (9) are established for arbitrary CGA points a, b, c and scalar β.

  Further, the inner product of the result of the outer product operation of the CGA points a and b and the base is expressed by the following equation (10).

By applying the outer product operation (外) and the dual operation ( ^* ) to the CGA point, it becomes possible to express various figures.

For example, as shown in FIG. 4A, a sphere that passes through the points P ₁ , P ₂ , P ₃ , and P ₄ is represented by the following equation (11).

As shown in FIG. 4B, the plane passing through the points P ₁ , P ₂ , and P ₃ is expressed by the following equation (12).

As shown in FIG. 4C, a circle passing through the points P ₁ , P ₂ , and P ₃ is expressed by the following equation (13).

As shown in FIG. 4D, a straight line passing through the points P ₁ and P ₂ is expressed by the following equation (14).

The chromosome creation unit 22 selects any three or four genes as a gene from the set of the three-dimensional point group transformed into the CGA space and the base e _∞ corresponding to the infinity point, A plurality of chromosomes including combinations of the selected genes are created, and for each chromosome, the points or bases of the chromosome genes are connected by an outer product to calculate the geometric structure represented by the chromosome. By estimating the optimal chromosome among all combinations of points using a genetic algorithm, it is possible to estimate the three-dimensional shape of an object that falls within a three-dimensional point group.

  Depending on the combination of selected genes, the type attribute of the object (geometric structure) is determined from one of (plane, curved surface, cylinder, straight line), and the geometric attribute of that type attribute is determined by the outer product operation and dual operation of the selected gene. Calculate the academic structure. The equations for obtaining the applicable geometric structure are summarized in Table 1 below.

  If the selected combination of genes does not correspond to the combination of genes in Table 1, three or four are again selected as genes.

  For some types of attributes, special measures are taken to simplify calculation.

Specifically, a curved surface is approximated by a sphere, and a fourth restriction that restricts the center of the sphere to be the same height as the imaging device 12 based on the _three points P ₁ P ₂ P ₃ . Add a point P _4, the outer product operation and dual operation, calculates the geometry.

The cylinder is approximated by a circle parallel to the road surface, and a point P ′ ₂ P ′ ₃ obtained by translating the point P ₂ P ₃ in the vertical direction so that the point P ₁ and the height component x ₂ coincide with each other is obtained. Used to calculate the geometric structure by outer product operation and dual operation.

The curved surface and the cylinder are distinguished depending on whether or not the _three points P ₁ P ₂ P ₃ selected as genes are within a predetermined distance range. At this time, the diameter of the utility pole (about 40 cm) is set as a threshold value, and when the maximum value of the distance between the points P ₁ P ₂ P ₃ is equal to or less than the threshold value, it is distinguished as a cylinder, and the point P ₁ P ₂ P _When the maximum value of the distance between ₃ is larger than the threshold, it is distinguished as a curved surface.

  The fitness calculation unit 24 calculates the distance between the geometric structure of the chromosome and the point by an inner product (·) in order to determine whether the geometric structure of the chromosome applies to the three-dimensional point cloud, and how many points are present. The fitness of a chromosome is determined by whether or not it is in close proximity to the geometric structure.

  However, the distance between the geometric structure S of the chromosome and the point P is calculated according to the following equation (15).

  Since the cylinder is expressed as a circle, the distance between the point and the circle is calculated on the same plane as the circle.

  The number N of point groups (neighboring points) whose distance is equal to or less than a predetermined threshold is defined as fitness. When calculating the fitness, the distance from all the point groups is calculated, and the amount of calculation becomes enormous when the number of point groups and the number of generational changes of the genetic algorithm are large. Therefore, in the present embodiment, a small subset of the three-dimensional point group is determined, and the number N of neighboring points is obtained for the point group included in the subset.

  For example, an index is assigned to each point of the three-dimensional point group, and in the k-th generation, only the points whose remainder is k when the index is divided by the generation alternation number K until the optimum individual is extracted are included in the subset. .

  However, in the K generation for extracting the optimum individual, the number of neighbors is obtained for all three-dimensional point groups, and the fitness is calculated.

  The optimization unit 26 extracts the three-dimensional shape of the structure on the road ahead of the host vehicle by obtaining a chromosome that maximizes the fitness based on a genetic algorithm.

  Specifically, as illustrated in FIG. 5, the optimization unit 26 selects two individuals (chromosomes) that leave offspring at the time of generation change from the individual set (chromosome set). Individuals are selected by roulette selection weighted by fitness.

  The update unit 28 crosses the genes (point and type attributes) of the two selected chromosomes as shown in FIG. At this time, crossover is performed so that any one of the four genes does not overlap in the next-generation chromosome to be newly generated. It is also effective to perform mutation with a predetermined low probability at the time of generation change and initialize a part of genes of the selected chromosome with random numbers.

  When the generation unit is changed a predetermined number of times K, the optimization unit 26 extracts the optimum individual (chromosome) having the maximum fitness based on the fitness of each chromosome, and the geometric structure of the extracted chromosome Is the three-dimensional shape of the structure of the runway ahead of the host vehicle. Since there is a possibility that a plurality of structures are included in the target range, after retrieving the optimum individual, the process returns to the process of the chromosome creation unit 22 and the above process is repeated. Each time, the point group included in the geometric structure represented by the extracted optimum individual (point group whose distance from the geometric structure is below the threshold value) is excluded, and is included in the interpretation of any geometric structure. Processing is interrupted when there are not enough point clouds. The process is also interrupted when the number of points near the optimum individual falls below a predetermined threshold.

  The optimal population obtained as a result of the optimization unit 26 represents a structure of any one of a plane, a curved surface, a cylinder, and a straight line.

  Planes include not only guardrails and walls, but also road and sidewalk planes. In particular, since the road plane has a large absolute number of points, it is often extracted before other structures.

  Guardrails and curbs can be distinguished based on the height distribution of neighboring points with reference to the road plane. When the mode value of the height of the neighboring points exceeds a predetermined threshold value, it is determined as a guardrail, otherwise it is determined as a curbstone.

  The map registration unit 30 determines the type of the structure as described above, and registers the structure together with the type of the structure on the map data stored in the storage unit 32. At this time, genes (point groups) constituting the optimum individual may be registered on the map data, or may be registered after being converted into general graphic equations. In either case, the interpretation result can be registered on the map data and used in post-processing.

  Next, the operation of the three-dimensional shape interpretation apparatus 10 according to the embodiment of the present invention will be described. First, when imaging of the front of the host vehicle is started by the imaging device 12 and the position of the host vehicle is measured as needed by the positioning device 14, the three-dimensional shape interpretation processing routine shown in FIG. Executed.

  In step 100, positioning information for a predetermined distance range obtained by the positioning device 14 and stereo images of a plurality of frames for the same distance range captured by the imaging device 12 are acquired.

  Next, in step 102, edge points are extracted from the stereo image of each frame acquired in step 100. In step 104, a three-dimensional position is obtained for each point of the edge point group of the stereo image of each frame extracted in step 102, and the edge point group of each frame is integrated using the positioning information acquired in step 100. A three-dimensional point group is obtained.

  In step 106, each point of the three-dimensional point group obtained in step 104 is converted into a CGA space according to the equations (1) and (2). In step 108, the variable k indicating the generation is set to 1, which is an initial value.

  In the next step 110, three or four genes are selected from the three-dimensional point cloud converted to the CGA space and the base e∞, and a predetermined number of chromosomes made up of gene combinations corresponding to the type attribute of the object are created. . Then, for each chromosome, a point or base that is a gene is connected by an outer product and a dual operation is performed to calculate an expression for obtaining a geometric structure.

  In step 112, a subset of the three-dimensional point group is determined based on the generation k, and the distance from each point of the determined subset is calculated for each chromosome created in step 110 according to the above equation (15). The number of points at which the distance is less than or equal to the threshold is obtained as fitness. When the generation is K, the fitness is obtained for all points in the three-dimensional point group.

  In the next step 114, it is determined whether or not k is less than the generational change number K for extracting the optimum individual. If k is less than K, in step 116, k is incremented by one. In step 118, two chromosomes are selected based on the fitness calculated in step 112, and the genes of the two selected chromosomes are crossed to update a predetermined number of chromosomes. Then, for each chromosome, the points or bases that are genes are connected by an outer product and a dual operation is performed to calculate an equation for obtaining a geometric structure, and the process returns to step 112 above.

  On the other hand, when k reaches K, the process proceeds to step 120, and the chromosome that is the optimum individual is extracted based on the fitness calculated in step 112. Further, a point group included in the extracted geometric structure of the chromosome is obtained from the three-dimensional point group, and the obtained point group is excluded from the three-dimensional point group.

  In step 122, it is determined whether or not to interrupt the genetic algorithm processing. If the process is not interrupted, the process returns to step 108 above. On the other hand, when the process is interrupted, in step 124, the type of structure is determined for the geometric structure of the chromosome of each optimal individual extracted in step 120, and the geometric structure (three-dimensional shape equation) is determined. Alternatively, the point cloud constituting the three-dimensional shape) is registered on the map data together with the type of the structure, and the three-dimensional shape interpretation processing routine is terminated.

  For example, when a stereo image as shown in FIG. 7A is input, as shown in FIG. 7B, the extracted geometric structure is registered on the map data.

  As described above, according to the three-dimensional shape interpretation apparatus of the embodiment of the present invention, each point of the three-dimensional point group acquired from the stereo image is converted into the CGA space, and a plurality of points or infinity is obtained. Select a base as a gene to create a chromosome, combine the genes of the chromosome using a cross product, and perform a dual operation to calculate a geometric structure that represents a 3D shape, Find the distance from the geometric structure by inner product calculation, calculate the fitness of the chromosome, update the chromosome by changing the generation of the chromosome based on the fitness of the chromosome, and maximize the fitness By repeating the generation change and the calculation of fitness, and outputting the 3D shape represented by the geometric structure of the chromosome with the maximum fitness, the 3D shape in the 3D point cloud can be interpreted with simple processing. be able to.

  In addition, when a 3D point cloud obtained from stereo images is projected onto the Conformal Geometric Algebra space, objects of different shapes such as planes and curved surfaces can be represented in the same format, so objects of different shapes such as utility poles and guardrails are mixed. Even in the environment, object extraction and shape classification can be performed easily. In the Conformal Geometric Algebra space, the process of extracting an object from a point cloud can be simply described using a point cloud (or basis) outer product operation, so that the object can be extracted by a uniform optimization framework. Therefore, a physical boundary (guardrail, curbstone, utility pole) on the road can be analyzed by simple processing from the three-dimensional point group obtained from the positioning information and the stereo image. Further, only by repeating the optimization process, the three-dimensional structure of the runway (road plane, sidewalk plane, guardrail, curb, utility pole) can be sequentially extracted and registered on the map.

  In addition, since knowledge about whether the road is a straight line or a curve (curvature, etc.) can be obtained from the optimization result, no prior knowledge or model regarding the road shape is required.

  Further, in the five-dimensional CGA space in the prior art, points, spheres, and planes can be expressed by the following common formula.

In the above Non-Patent Document 3, an attempt is made to automatically classify a sphere or a plane by estimating the five coefficients (x ₁ , x ₂ , x ₃ , x _∞ , x _O ) of the above equation by the least square method. . However, the technique of Non-Patent Document 3 has a problem that an object other than a sphere and a plane must take into account a higher-order coefficient (eg, e ₁ ∧e ₂ ) and cannot be handled in this manner. is there. Further, the least square method has a problem that it is easily influenced by an outlier. Since the stereo three-dimensional point group includes a lot of noise, there is a possibility that the shape classification is erroneous unless the noise is appropriately removed.

  On the other hand, in this embodiment, there is no need for coefficient estimation, and an object can be determined from only a small number of points.

  In the above embodiment, the case where the three-dimensional shape interpretation device is mounted on a vehicle has been described as an example. However, the present invention is not limited to this, and may be mounted on a device other than the vehicle, for example, a robot.

  In addition, the case where an optimal geometric structure is obtained using a genetic algorithm has been described as an example. However, by using a method other than the genetic algorithm, the geometric structure is applied to a three-dimensional point group to obtain an optimal geometric structure. May be requested.

  The program of the present invention may be provided by being stored in a storage medium.

DESCRIPTION OF SYMBOLS 10 3D shape interpretation apparatus 12 Imaging device 14 Positioning apparatus 16 Computer 18 Point cloud acquisition part 20 CGA conversion part 22 Chromosome preparation part 24 Fitness calculation part 26 Optimization part 28 Update part 30 Map registration part

Claims (9)

Point cloud acquisition means for acquiring point cloud data consisting of a plurality of points having three-dimensional coordinates;
Space conversion means for converting the plurality of points of the point cloud data into a CGA (Conformational Geometric Algebra) space;
A combination consisting of any three or four of the plurality of points and the infinity base of the point cloud data transformed onto the CGA space is selected, and the selected combination is combined using an outer product. Creating means for creating geometric structure data representing any of a plurality of predetermined three-dimensional shapes;
For each of the plurality of points of the point cloud data converted on the CGA space, a distance between the point and the geometric structure data created by the creation means is obtained by inner product calculation, and each of the plurality of points Fitness calculation means for calculating the fitness of the geometric structure data based on the distance determined for
Updating means for updating the combination included in the geometric structure data based on the fitness of the geometric structure data;
Optimization means for repeating the update by the update means and the calculation by the fitness calculation means so as to maximize the fitness, and outputting a three-dimensional shape represented by the geometric structure data;
A three-dimensional shape interpretation apparatus. The creation means selects any three or four of the plurality of points and the infinite base of the point cloud data converted on the CGA space as a combination of genes, and the three or four A plurality of chromosomes including a combination of two genes is created, and for each chromosome, the combination of the genes of the chromosome is combined using an outer product to create the geometric structure data,
The fitness calculation means calculates the fitness of the geometric structure data of the chromosome for each of the plurality of chromosomes,
The update means selects two of the chromosomes based on the fitness of each of the plurality of chromosomes calculated by the fitness calculation means, and crosses the genes of the selected two chromosomes. The three-dimensional shape interpretation apparatus according to claim 1, wherein the plurality of chromosomes are updated.   The geometric structure data is obtained by combining a combination of three selected points and a base at infinity with a cross product, a geometric structure data representing a plane, the selected three points, and the selected three points. Geometric structure data representing a curved surface, which is a combination of points added based on points, and a combination of three points that have been modified so that the height components of the three selected points match. Or a geometric structure data representing a straight line obtained by combining a combination of the selected two points and an infinite base with an outer product. Item 3. The three-dimensional shape interpretation apparatus according to item 1 or 2. The fitness calculation means obtains a distance between the point and the geometric structure data created by the creation means for each of the points of the subset of the point cloud data by an inner product operation, and each of the points of the subset Calculating the fitness of the geometric structure data based on the distance obtained for
The three-dimensional shape interpretation apparatus according to any one of claims 1 to 3, wherein the subset is determined so as to be different each time the calculation by the fitness calculation means is repeated. The point cloud acquisition means acquires point cloud data consisting of a plurality of points having three-dimensional coordinates around the vehicle,
The optimization means repeats the update by the update means and the calculation by the fitness calculation means so that the fitness is maximized, and the three-dimensional representation represented by the geometric structure data existing around the vehicle. The three-dimensional shape interpretation apparatus according to any one of claims 1 to 4, which outputs a shape. It further includes positioning means for positioning the vehicle position information,
6. The three-dimensional shape interpretation apparatus according to claim 5, wherein the three-dimensional coordinates of each point of the point group data are three-dimensional coordinates in an actual coordinate system obtained using position information of the vehicle measured by the positioning means.   The three-dimensional shape interpretation apparatus according to claim 6, further comprising a map registration unit that registers a three-dimensional shape represented by the geometric structure data output by the optimization unit on a map. The creation means repeatedly creates the geometric structure data,
From any of the plurality of points of the point group data excluding the combination of the points included in the geometric structure data output last time by the optimization means, and three or four of the infinite bases The three-dimensional shape interpretation apparatus according to any one of claims 1 to 7, wherein the combination is selected and the geometric structure data is created. Computer
Point cloud acquisition means for acquiring point cloud data consisting of a plurality of points having three-dimensional coordinates;
Space conversion means for converting the plurality of points of the point cloud data onto a CGA (Conformational Geometric Algebra) space;
A combination consisting of any three or four of the plurality of points and the infinity base of the point cloud data transformed onto the CGA space is selected, and the selected combination is combined using an outer product. Creating means for creating geometric structure data representing any of a plurality of predetermined three-dimensional shapes,
For each of the plurality of points of the point cloud data converted on the CGA space, a distance between the point and the geometric structure data created by the creation means is obtained by inner product calculation, and each of the plurality of points Fitness calculation means for calculating the fitness of the geometric structure data based on the distance determined for
Update means for updating the combination included in the geometric structure data based on the fitness of the geometric structure data, and update by the update means and calculation of the fitness so that the fitness is maximized A program for functioning as an optimization unit that repeatedly calculates by means and outputs a three-dimensional shape represented by the geometric structure data.