JP6179913B2 - Columnar object extraction method, columnar object extraction program, and columnar object extraction device - Google Patents

Description

  The present invention relates to a columnar object extraction method for extracting a columnar object such as a streetlight, a power pole, and a sign from a measurement point group obtained by measuring an urban area, and a columnar object extraction program and columnar object extraction apparatus for automatically extracting a columnar object It is about.

  In recent years, the demand for topographic information (spatial information) has increased with the advancement of measurement technology. For example, for the purpose of managing facilities installed on roads or along roads to a higher degree, there are an increasing number of managers requesting facility space information such as the shape and installation position thereof. In particular, there are many pillar-shaped features such as streetlights, utility poles, signs, etc. (hereinafter referred to as “columnar objects”). However, not only the installation location but also spatial information such as the height of facilities, deposits (signboards, signals, etc.), and positional relationships with other surrounding features are also being demanded.

  In the past, when acquiring spatial information of a columnar object, a method of actually performing surveying or extracting from a plan drawing or the like has been employed. That is, surveying or the like was performed after a person determined in advance that the feature was a columnar object. Therefore, these operations performed on a large number of columnar objects not only require a considerable amount of time and labor, but also result accuracy is not necessarily high, such as extraction omissions and errors.

  On the other hand, as described above, recent measurement techniques have advanced remarkably, and it has become possible to simultaneously acquire a large amount of measurement data with higher accuracy than in the past. A typical measurement means is measurement using a laser scanner, which is a technique of measuring using laser reflection irradiated on a measurement object at several tens of thousands per second, and can simultaneously acquire a large amount of measurement data. Normally, the measurement is performed while moving a laser scanner on a moving body, and until now it has been the mainstream to measure the topography from the air by mounting it on an aircraft. In recent years, a method called a mobile mapping system (Mobile Mapping System: MMS) that is mounted on an automobile and measures while moving on a road is also frequently used.

  According to the mobile mapping system, measurement data of every feature on the road or along the road is acquired. Of course, the measurement result of the columnar object is also included in the measurement data measured by the mobile mapping system. By using this measurement data, spatial information (installation position, shape, etc.) of the columnar object can be obtained. . However, if a columnar object is extracted from a large amount of measurement data according to human judgment such as visual inspection, more time and labor are required than before, and many extraction omissions and errors may occur. is expected.

  If the measurement data of columnar objects can be automatically extracted from a large amount of measurement data, human work will be significantly reduced and human error will be eliminated. Has been.

  Therefore, Patent Document 1 proposes a technique for automatically detecting a columnar object (referred to as a “cylindrical object” in this document) by processing measurement point data obtained by laser scanner measurement with a computer.

JP 2010-286267 A

  In Patent Document 1, for example, when an object having the same cross-section in the vertical direction is measured, such as a utility pole, the measurement point (three-dimensional coordinates) is projected onto a horizontal plane, and is measured in a certain shape (that is, the cross-sectional shape of the utility pole). It is a technology that uses the feature that points are concentrated. Specifically, a cylindrical object is detected by collating a measurement point projected on a horizontal plane with a template pattern prepared in advance and judging the collation result.

  However, the method of Patent Document 1 requires labor for creating a template pattern of a cylindrical object to be detected, and there is instability that the detection accuracy differs depending on the quality of the template pattern. In addition, it can be pointed out that an object whose cross section changes in the vertical direction is difficult to detect, and that even a columnar shape is difficult to detect. Further, it is necessary to recognize the shape of the cylindrical object to be detected, and when it is not possible to grasp in advance which columnar object exists, the method of Patent Document 1 can be adopted. Can not.

  The subject of the present invention is the problem of the prior art, that is, it takes a lot of labor and time to obtain the spatial information of the columnar object, and it is difficult to extract a large number of columnar objects without omission. A columnar object extraction method, a columnar object extraction program for automatically extracting a columnar object from a large amount of measurement data accurately and in a short time, and The object is to provide a columnar object extraction apparatus.

  The columnar object extraction method of the present invention is a columnar object having a columnar shape among standing objects from a measurement point group that is a set of measurement points in a predetermined space including a plurality of standing objects standing on the ground. The segmentation process for obtaining segment data for each area independent from the measurement point group and the thinning process for clarifying the shape characteristics of the standing object for each segment data. The thinning process to obtain the feature shape data and the principal component analysis for each of the shape constituent points constituting the standing shape data to obtain eigenvalues and eigenvectors, and shape classification based on the feature values calculated from the eigenvalues And a columnar object determination step of determining the columnar shape of the standing object based on the feature quantity of the shape composing point. Here, the segmentation step generates an edge between an arbitrary measurement point and a nearby measurement point close to this point, and obtains one continuous component obtained as a result as one segment data. In the thinning step, The thinning process is to move the position of an arbitrary measurement point, and based on the thinning intensity obtained from the angle between the arbitrary measurement point, the nearby measurement point, and two neighboring measurement points at the arbitrary measurement point, An arbitrary measurement point position after movement is obtained.

  In the columnar object extraction method of the present invention, the segmentation step is performed by selecting a predetermined number k of neighboring measurement points in order from the arbitrary measurement point and generating an edge between the neighboring measurement point and the arbitrary measurement point. It is also possible to create a neighborhood connection graph and obtain one continuous component obtained from the k neighborhood connection graph as one segment data.

  In the columnar object extraction method of the present invention, the eigenvalues obtained by principal component analysis are first eigenvalue, second eigenvalue, and third eigenvalue in descending order of values, and the eigenvector corresponding to the first eigenvalue is the first eigenvector, A method corresponding to the second eigenvalue may be a second eigenvector, and a method corresponding to the third eigenvalue may be a third eigenvector. In this case, the feature amount is calculated based on the first feature value calculated based on the first eigenvalue and the second eigenvalue, the second feature amount calculated based on the second eigenvalue and the third eigenvalue, and the third eigenvalue. The shape classification is composed of two or more classifications including “points on columnar objects”. In addition, a “point on a columnar object” is assigned as a shape classification to a shape composing point where the first feature amount has the maximum value among the feature amounts.

  In the columnar object extraction method of the present invention, the ratio of the shape component points classified as “points on the columnar object” out of the shape component points constituting the standing object shape data, or / and “points on the columnar object” The columnar object degree of standing object shape data is obtained on the basis of the ratio of the shape component points in which the first eigenvector is vertically or nearly perpendicular to the shape component points categorized into 3 and the columnar object degree is evaluated. It can also be set as the method of determining the columnar shape of a standing thing.

  In the columnar object extraction method of the present invention, the columnar object determination step may include a neighboring connection graph generation step, a basic component formation step, a basic component classification step, and a columnar object segment creation step. In this case, the adjacent connection graph generation step obtains an adjacent point adjacent to each shape component point based on the eigenvector of each shape component point, and generates an edge between the adjacent point and the shape component point, One continuous component obtained as a result is obtained as one adjacent connection graph. In addition, the basic component forming step obtains an adjacent point assigned the same shape classification as each shape component point for each shape component point constituting the adjacent connection graph, and between the adjacent point and the shape component point. To generate an edge and obtain one continuous component as a basic component. In the basic component classification step, components corresponding to the shape classification are classified with respect to the basic component, and the basic component whose shape classification is “a point on a columnar object” is classified as a “post component”. In the columnar object segment creation step, an adjacent connection graph including a column component is extracted, and one or more basic components included in the adjacent connection graph are combined to create a columnar object segment. As the adjacent point at this time, the shape composing point closest to the shape composing point is selected within the spatial range expanded from the shape composing point toward the eigenvector direction.

  In the columnar object extraction method of the present invention, when two or more basic components whose shape classification is “points on columnar objects” are included in one adjacent connection graph, the shape classification is “points on columnar objects”. A basic component having a minimum elevation value may be classified as a “post component”.

  The columnar object extraction method of the present invention may be a method of creating a columnar object segment by combining one or two or more basic components after excluding the “ground component”. In this case, “point on the plane” is assigned as the shape classification to the shape constituent point having the maximum second feature amount, and the basic component whose shape classification is “point on the plane” is classified as “plane component”. . Then, a substantially horizontal planar component is extracted as a “ground component” based on the normal direction of the surface formed by the shape composing points constituting the planar component.

  The columnar object extraction method of the present invention may be a method of creating a columnar object segment by combining one or more basic components after excluding the “connected object component”. In this case, a basic component whose shape classification is “a point on a columnar object” and whose first eigenvector of the shape composing point constituting the basic component is substantially horizontal is extracted as a “connected component”.

  The columnar object extraction method of the present invention can further classify columnar objects based on the degree of attribution. In this case, out of the basic components that make up the columnar object segment, the basic components other than the column components are extracted as parts and the combinations of the basic components connected to each other are extracted as parts. Then, the attribution degree of the columnar object segment is obtained based on the evaluation of the height of the columnar object segment, the number of parts included in the columnar object segment, and the part type.

  The columnar object extraction method of the present invention can further classify columnar objects based on the degree of belonging after extracting the strut composing points. In this case, the shape composing point within a predetermined distance from the strut candidate point and the strut candidate point are collectively set as a strut composing point. From the standing object shape data, a continuous component composed of shape composing points other than the strut composing point is extracted as a part, and this part is classified into a part type based on the shape classification, the height of the standing object shape data, Based on the number of parts included in the standing object shape data and the evaluation regarding the part type, the degree of attribution of the standing object shape data is obtained.

  The columnar object extraction method of the present invention can further classify columnar objects based on the context feature quantity. In this case, a plurality of meshes are configured with a predetermined grid, the degree of attribution is assigned to the nearest mesh, an area configured with a predetermined number of meshes is set, and a standard is set for each type of columnar object based on this area. Create a distribution map. Further, a peripheral distribution map is created for each columnar object segment or each standing object shape data based on the area, and the columnar object segment or standing object shape data is generated based on the standard distribution map and the peripheral distribution map. To calculate the context feature value.

  The columnar object extraction program of the present invention is a columnar object having a columnar shape among standing objects from a measurement point group that is a set of measurement points in a predetermined space including a plurality of standing objects standing on the ground. This is a columnar object extraction program that allows a computer to execute the function of extracting the segment, segmentation function to create segment data for each area independent from the measurement point group, and the shape characteristics of standing objects for each segment data A thinning function to create standing shape data by performing thinning processing, and a principal component analysis function to calculate eigenvalues and eigenvectors by performing arithmetic processing for principal component analysis on each shape component point, and Furthermore, a shape classification function for assigning a shape classification according to the feature amount calculated based on the eigenvalue to each shape constituent point, and the feature amount of the shape constituent point Those having a columnar object determination function to perform the columnar determination process of standing 設物 based. Here, the segmentation function generates an edge between an arbitrary measurement point and a nearby measurement point close to this point, and uses one continuous component obtained as a result as one segment data. The thinning process to be performed is to move the position of an arbitrary measurement point, and calculates an included angle between two adjacent measurement points at an arbitrary measurement point, Based on this, the position of the arbitrary measurement point after movement is calculated.

  The columnar object extraction program of the present invention uses a segmentation function to select a predetermined number k of neighboring measurement points in order from the nearest measurement point and generate an edge between the neighboring measurement point and the arbitrary measurement point, thereby making the k neighborhood It is also possible to create a connection graph and use one continuous component obtained from this k-neighbor connection graph as one segment data.

  In the columnar object extraction program of the present invention, the eigenvalues obtained by the principal component analysis function are set to the first eigenvalue, the second eigenvalue, and the third eigenvalue in descending order of the values, and the first eigenvalue corresponding to the first eigenvalue is also set as the first eigenvalue. A program corresponding to the eigenvector, the second eigenvalue corresponding to the second eigenvalue, and the third eigenvector corresponding to the third eigenvalue may be used. In this case, the feature value is based on the first feature value calculated based on the first eigenvalue and the second eigenvalue, the second feature value calculated based on the second eigenvalue and the third eigenvalue, and the third eigenvalue. The shape classification is composed of two or more classifications including “points on columnar objects”. In addition, a “point on a columnar object” is assigned as a shape classification to a shape composing point where the first feature amount has the maximum value among the feature amounts.

  The columnar object extraction program of the present invention is the ratio of the shape component points classified as “points on columnar objects” out of the shape component points constituting the standing object shape data, or / and “points on columnar objects”. The columnar object degree of standing object shape data is calculated on the basis of the proportion of the shape constituent points whose eigenvectors are vertical or close to vertical among the shape constituent points classified as, and this columnar object degree is evaluated. It is also possible to perform columnar determination of standing objects.

  In the columnar object extraction program of the present invention, the columnar object determination function may include an adjacent connection graph generation function, a basic component formation function, a basic component classification function, and a columnar object segment creation function. In this case, the adjacent connection graph generation function selects an adjacent point adjacent to each shape component point based on the eigenvector of each shape component point, and generates an edge between the adjacent point and the shape component point. One continuous component obtained as a result is taken as one adjacent connection graph. In addition, the basic component forming function selects an adjacent point to which the same shape classification as each shape component point is assigned to each shape component point constituting the adjacent connection graph, and Edges are generated between them, and one continuous component obtained as a result is used as one basic component. The basic component classification function assigns a component classification corresponding to the shape classification to the basic component. For a basic component whose shape classification is “a point on a columnar object”, the component classification is “ A “post component” is assigned. The columnar object segment creation function extracts an adjacent connection graph including a column component and combines one or more basic components included in the adjacent connection graph to create a columnar object segment. As the adjacent point at this time, the shape composing point closest to the shape composing point is selected within the spatial range expanded from the shape composing point toward the eigenvector direction.

  In the columnar object extraction program of the present invention, when two or more basic components whose shape classification is “points on columnar objects” are included in one adjacent connection graph, the shape classification is “points on columnar objects”. It is also possible to give a “post component” as a component classification to a certain basic component having the minimum elevation value.

  The columnar object extraction program of the present invention can also create a columnar object segment by combining one or more basic components, excluding the “ground component”. In this case, “point on the plane” is assigned as the shape classification to the shape component point where the second feature value has the maximum value, and the component classification is applied to the basic component whose shape classification is “point on the plane”. Assign "Plane component". Then, a substantially horizontal planar component is extracted as a “ground component” based on the normal direction of the surface formed by the shape composing points constituting the planar component.

  The columnar object extraction program according to the present invention may be configured to create a columnar object segment by combining one or more basic components after excluding the “connected object component”. In this case, a basic component whose shape classification is “a point on a columnar object” and whose first eigenvector of the shape composing point constituting the basic component is substantially horizontal is extracted as a “connected component”.

  The columnar object extraction program of the present invention may be provided with a columnar object type recognition function for further classifying columnar objects based on the degree of attribution. In this case, out of the basic components that make up the columnar object segment, a basic component other than the column component, and a part extraction function that extracts a combination of interconnected basic components as parts, and parts based on shape classification A part type classification function that classifies the object into a type, and an attribute calculation function that calculates the attribute of the columnar object segment based on the evaluation of the height of the columnar object segment, the number of parts included in the columnar object segment, and the part type Also prepare.

  The columnar object extraction program of the present invention may have a columnar component point extraction function and a columnar object type recognition function for further classifying columnar objects based on the degree of belonging. In this case, the column configuration point extraction function is a configuration component point classified as “point on a columnar object” in the standing object shape data, and the shape configuration point whose first eigenvector is substantially in the vertical direction. As the candidate points, the shape constituent points within a predetermined distance from the pillar candidate points and the pillar candidate points are collectively set as the pillar constituent points. In addition, a part extraction function that extracts continuous components composed of shape constituent points other than the pillar constituent points from the standing object shape data as parts, and a part type classification function that classifies parts into part types based on shape classification, In addition, it has an attribute calculation function that calculates the attribute of the standing object shape data based on the evaluation of the height of the standing object shape data, the number of parts included in the standing object shape data, and the part type. .

  The columnar object extraction program of the present invention may be provided with a columnar object type recognition function for selecting the type of columnar object based on the context feature value. In this case, a plurality of meshes are configured by a predetermined grid, an attribution score distribution function for distributing the attribution to the nearest mesh, an area setting function for setting an area configured by a predetermined number of meshes, and an area based Standard distribution map creation function to create a standard distribution map for each type of columnar object, and peripheral distribution map creation function to create a peripheral distribution map for each columnar object segment or each standing object shape data based on area And a context feature amount calculation function for calculating a context feature amount for the columnar object segment or the standing object shape data based on the standard distribution map and the peripheral distribution map.

  The columnar object extraction device of the present invention is a columnar object whose shape is a columnar shape among standing objects from a measurement point group that is a set of measurement points in a predetermined space including a plurality of standing objects standing on the ground. Is a columnar object extraction apparatus that measures a predetermined space and acquires a measurement point group, a storage unit that stores the measurement point group acquired by the measurement unit, and a standing unit based on the measurement point group An arithmetic processing means for extracting a columnar object from among the objects is provided. Further, the arithmetic processing means has a segmentation function, a thinning function, a principal component analysis function, a shape classification function, and a columnar object determination function. Here, the segmentation function creates segment data for each independent region from the measurement point group, and selects a predetermined number k neighboring measurement points in order from the arbitrary measurement point. The k neighborhood connection graph is created by generating an edge between and an arbitrary measurement point, and one continuous component obtained from the k neighborhood connection graph is used as one segment data. For each segment data, the standing line shape data is created by thinning processing to clarify the shape characteristics of the standing object, and the principal component analysis function is applied to each shape composing point. The eigenvalues and eigenvectors are obtained by performing arithmetic processing for principal component analysis, and the shape classification function is calculated based on the eigenvalues for each shape composing point. The columnar object determination function is used to perform columnar determination processing for standing objects based on the feature quantities of the shape composing points, and the thinning performed by the thinning function. The process is to move the position of an arbitrary measurement point, calculate an included angle between two adjacent measurement points at the arbitrary measurement point, and move based on the included angle, the arbitrary measurement point, and the adjacent measurement point The later arbitrary measurement point position is calculated.

The columnar object extraction method, columnar object extraction program, and columnar object extraction apparatus of the present invention have the following effects.
(1) Since a columnar object is automatically extracted from a large number of measurement points without any human judgment such as visual inspection, human error is eliminated and extraction can be performed without omission, and the extraction result can be obtained in a very short time. Can be obtained.
(2) Since the thinning process is performed based on the thinning strength obtained from the angle between two adjacent measurement points at an arbitrary measurement point, the end points and branch points of the columnar object are retained (point movement is small) and the thin line As a result, the shape feature can be clarified while accurately expressing the shape. Therefore, a columnar object can be extracted accurately.
(3) When segmentation is performed, if k neighboring measurement points are selected in order from the nearest measurement point and a k neighborhood connection graph is created, the segmentation processing time is shortened, and as a result, columnar objects are extracted. The processing can be further speeded up.
(4) Columnar objects of various thicknesses and inclinations can be extracted by using concepts such as principal component analysis and shape classification according to feature values calculated based on eigenvalues.
(5) By extracting parts from columnar objects and classifying them into parts, the columnar objects can be further classified into detailed classifications such as utility poles, street lamps, and signs.

(A) is the whole model figure showing the topography of the whole city area, (b) is the columnar object model figure showing only the extracted columnar object. The flowchart which shows the procedure of this invention, and its flow. The flowchart which shows the procedure of 1st Embodiment, and its flow. Explanatory drawing which shows the view which detects the vicinity measurement point of arbitrary measurement points. The model figure explaining the included angle which consists of an arbitrary measurement point and a nearby measurement point. The model figure for demonstrating the element used in order to produce a dispersion | distribution covariance matrix. The model figure which shows the point distribution state and 1st eigenvector of a neighborhood shape constituent point set. The model figure which provided the shape classification | category with respect to the shape composition point, when a standing thing is a guide sign comprised with a vertical pillar, a horizontal pillar, and a flat plate. FIG. 6 is a model diagram in which when a standing object is a guide sign composed of a vertical pillar, a horizontal pillar, and a flat plate, a shape classification is given to the shape constituent points and further classified with a first eigenvector. The flowchart which shows the procedure and flow of 2nd Embodiment. The model figure which demonstrated the condition which searches the adjacent point in a shape composition point concretely. (A) is a model diagram showing an eigenvector selected when the shape classification is “a point on a columnar object”, and (b) is a model diagram showing an eigenvector selected when the shape classification is “a point on a plane”. (C) is a model figure which shows the eigenvector selected when shape classification is "the point on another object." The model figure which shows the state by which multiple different basic components were produced | generated. The model figure which shows a connection thing component. The model figure which shows that the cross-sectional shape of a support | pillar component forms a part of substantially circle shape. The flowchart which shows the procedure which classify | categorizes a columnar object. The model figure which shows the example of the part type classified based on the shape classification of the basic component, and the combination. Explanatory drawing which shows the calculation method of an attribution degree for every telephone pole, a streetlight, and a sign. The flowchart which shows the procedure which classifies | categorizes the type of a columnar object based on a context feature-value. The model figure which shows the mesh formed in the set area range centering | focusing on the columnar object segment to focus on. The model figure which shows three types of periphery distribution maps. The model figure which shows the calculation method of a context feature-value. The model figure which shows the specific example which calculated the context feature-value.

  Exemplary embodiments of a columnar object extraction method, a columnar object extraction program, and a columnar object extraction apparatus according to the present invention will be described with reference to the drawings.

1. Overall Overview FIG. 1 is a model diagram showing a state in which a columnar object is extracted from an entire terrain obtained by measuring a city area, (a) is an overall model diagram showing the terrain of the entire city area, and (b) is an extracted columnar object. It is a columnar object model figure showing only. The overall model diagram of FIG. 1A is a bird's eye view showing a set of measurement points (hereinafter referred to as “measurement point group”) obtained by measuring a predetermined range of an urban area. As described above, the present invention extracts a columnar object from features of the entire terrain.

  The measurement points constituting the measurement point group can be obtained by, for example, measuring while moving a predetermined range in an urban area with an automobile equipped with a laser scanner. This measurement point is a three-dimensional coordinate (X, Y, Z) composed of a plane coordinate and a height. Of course, this measurement point can also be expressed by latitude, longitude, and altitude. In addition, the method for obtaining the measurement point is not limited to the method using an in-vehicle laser scanner (so-called MMS), and a method of measuring by mounting a laser scanner on an aircraft can be used, and TS can be used without using a laser scanner. A conventional measurement method using (total station) or the like can also be adopted. Further, it is not always necessary to perform measurement for the present invention, and it is needless to say that existing measurement points can be used.

  When a laser scanner is used, a large number of measurement points are usually obtained by one measurement. As described, the set of many measurement points is a “measurement point group”. That is, the overall model diagram of FIG. 1A can be said to be a diagram in which the measurement point group of the entire city area is displayed as a bird's eye view. In addition, although the space made into object here is the predetermined range of an urban area, when implementing this invention, if the space contains a some standing structure, the object space does not need to be an urban area.

  FIG. 2 is a flowchart showing an outline of the present invention. Each procedure (step) and its flow will be described with reference to this figure. First, a measurement point cloud that is a measurement result of an urban area is prepared. This group of measurement points includes a “cluster” of measurement points that are somewhat independent. This “chunk” is considered to be a feature standing on the ground (standing on the ground) (hereinafter referred to as “standing feature”), and a predetermined area including this standing feature is divided into segments. It is the segmentation (S02). For convenience, a set of measurement points divided for each predetermined area is referred to as “segment data”.

  Next, thinning processing is performed on each segment data (S03). This is to clarify the characteristics of the shape (outer shape) of standing objects. For example, when measuring a power pole, measurement points other than the power pole and measurement points with errors are included. This is a process to bring it close to the shape of the utility pole. Here, the result obtained by thinning the segment data is called “standing structure data”, and the measurement points are adjusted by thinning processing (that is, each point constituting the standing structure data) Is referred to as a “shape component point”.

  Next, the principal component analysis is performed on each shape component point, the orientation (axial direction) of the standing object corresponding to the standing object shape data is obtained, and each shape is based on the eigenvalue obtained as a result of the principal component analysis. The feature amount of the constituent point is calculated (S04). This feature amount is a value representing the shape feature of the standing object, and a predetermined shape classification can be given according to the feature amount (S05).

  When a shape classification is given to each shape component point, it becomes possible to recognize a columnar object using this shape classification, and as a result, only a columnar object can be extracted from the features of the entire city area. It becomes possible.

  The method of recognizing a columnar object using shape classification is roughly divided into three types. Here, “first embodiment”, “second embodiment”, and “third embodiment”, respectively. I will explain. Hereinafter, each embodiment will be described in detail for each element.

2. First Embodiment An example of a first embodiment of the present invention will be described with reference to the drawings. FIG. 3 is a flowchart showing an outline of the first embodiment. Hereafter, each procedure (step) is demonstrated along this flowchart.

(Removal of ground points)
When the measurement point group includes a measurement point obtained by measuring the ground (hereinafter referred to as “ground point”), the ground point is removed from the measurement point group (S01). This is because a group of measurement points that are independent (that is, not continuous with others) is required when performing segmentation described later. Therefore, when the ground point is not included in the measurement point group, it is not necessary to perform the processing here.

  The process of removing ground points from the measurement point group can be performed by a conventionally used method. One example is region growing (region expansion). In this method, when an arbitrary ground point is specified, a point having a predetermined gradient (vertical gradient) from that point is selected as a ground point. It is a technique to identify. The set of measurement points obtained by removing the ground points from the measurement point group can be considered as a set of measurement points obtained by measuring the standing object, and this is called the “standing object measurement point group”. And

(segmentation)
Segmentation is a process of separating (segmenting) each standing object with respect to the standing object measurement point group arranged randomly (S02). This is based on the idea that a plurality of measurement points obtained by measuring one standing object are close to each other, and a measurement point (neighbor measurement point) in the vicinity from an arbitrary measurement point (arbitrary measurement point i). A segment is formed by detecting j) and deriving it one after another. As a result, what is obtained is segment data, and basically, as many segment data as the number of standing objects included in the target space are obtained.

The segmentation will be described in more detail. FIG. 4 is an explanatory diagram showing the concept of detecting a measurement point j in the vicinity of an arbitrary measurement point i. Note that because of space limitations, FIG. 4 looks two-dimensional, but only shows a three-dimensional concept. The detection of the nearby measurement point j can be performed at all points within a predetermined distance from the arbitrary measurement point i, but considering the subsequent processing speed, the number of the nearby measurement points j should be limited to k or less. . In FIG. 4, the number of neighbors k is eight, and the neighbor measurement points j _{1 to} j ₈ are detected. In this case, the number k of neighboring measurement points j is selected in order from the arbitrary measurement point i. Of course, the number k of neighbors can be determined by designing as appropriate.

  In addition, a limit distance can be provided as a condition for selecting the nearby measurement point j. That is, as shown in FIG. 4, a sphere having a radius r is drawn around an arbitrary measurement point i, and the number of neighboring measurement points j corresponding to the number k of neighborhoods (eight in the figure) in the closest order among the measurement points within this range. To detect. Therefore, when there are few measurement points within the range of the limit distance (within the range of the sphere having the radius r), the nearby measurement points j less than the number k of neighborhoods are detected.

  When detecting a nearby measurement point j with respect to an arbitrary measurement point i, conditions can be collated for all measurement points. However, in order to shorten the detection processing time here, this is called a kd tree that has been used conventionally. Techniques can also be adopted.

When the nearby measurement point j at the arbitrary measurement point i is detected, the arbitrary measurement point i is associated with all the nearby measurement points j. Specifically, an edge (link) is generated at an arbitrary measurement point i and a nearby measurement point j. In FIG. 4, eight edges (i, j ₁ ) to (i, j ₈ ) are generated.

  When the edge generation processing is completed for one arbitrary measurement point i, the same processing is performed using the other measurement points as arbitrary measurement points. If this is repeated, measurement points in the vicinity gather and can be obtained as one continuous component. That is, one continuous component obtained here is segment data corresponding to one standing object. In addition, the result of obtaining a plurality of segment data from the standing object measurement point group in this way is called a connection graph. In particular, the result obtained by detecting the neighborhood measurement points j of the number k of neighborhoods is “ k neighbor connection graph ".

(Thinning process)
The thinning process is also called smoothing and clarifies the characteristics of the shape of the standing object (S03). The measurement points constituting the segment data include measurement points that are measured other than standing objects and measurement points with errors. The process of adjusting the measurement points to bring them closer to the original shape of the utility pole is the thinning process. The measurement points constituting the segment data are gradually moved and adjusted.

Thinning processing is a technique that has been performed so far, but the present invention focuses on the point of extracting columnar objects, and one feature is that it incorporates the concept of thinning strength (smoothing strength). is there. Specifically, it is a process of gradually moving the measurement points according to the following formula.

In the above formula, an arbitrary measurement point i, which is an arbitrary measurement point, is adjusted (moved) to i ′, p _i is the coordinate of the arbitrary measurement point i before thinning processing (before movement), and p _j is an arbitrary The coordinates of the measurement point j in the vicinity of the measurement point _i and p ′ _i are the coordinates after the thinning process (after movement). Further, i * is a set of measurement points j in the vicinity of the arbitrary measurement point i (hereinafter referred to as “neighboring point set”), q is a measurement point in the vicinity of the arbitrary measurement point i and is different from j, and V _i is It is a point pair set of combinations (j, q) of neighboring measurement points.

As shown in the above formula, p ′ _i after the thinning process is based on the coordinates of the arbitrary measurement point i, the coordinates of the nearby measurement point j, the thinning intensity λ, and the weight ω _ij (using these as variables). Desired. Further, the thinning strength λ is | V _i | which is the number of point pairs (number of combinations of (j, q)) of the point pair set V _i , the neighboring measurement point j centered on the arbitrary measurement point i, and the neighboring measurement point q. It is calculated based on angle (j, i, q) which is an included angle. The weight ω _ij is a positive value obtained based on the reciprocal of || p _j −p _i ||, which is the distance between p _i and p _j , and suppresses the influence of nearby measurement points at a distant position. ing.

FIG. 5 is a model diagram for explaining the included angle θ (that is, angle (j, i, q)) composed of an arbitrary measurement point and a nearby measurement point. As shown in this figure, it can be seen that the closer to the end point located at the end of the standing object, the smaller the included angle θ becomes. In other words, the thinning intensity λ takes a small value. As a result, the influence of the arbitrary measurement point p _i is strongly left, and the influence of the neighboring measurement point p _j is suppressed. Therefore, the end points and branch points constituting the standing object are not moved much by the thinning process, that is, the thinning process is performed while the end points and the branch structure are maintained.

  Basically, the thinning process is performed on all the measurement points constituting the segment data, and is performed on all the segment data. Also, as shown in the flow of FIG. 3, when the thinning process is performed once for all segment data (m segment data in the figure) and this is repeated a plurality of times (n times in the figure), This is preferable because the shape feature can be further clarified. The number of repetitions n of the thinning process can be designed as appropriate.

As described above, the shape composing point I (p ′ _i ) is obtained by thinning the measurement points, that is, the standing object shape data (a set of the shape composing points I) by thinning the segment data. Is obtained.

(Grant shape classification)
The shape component point I obtained by the thinning process constitutes the shape of the standing object, and the degree of influence on this shape differs depending on the local point distribution state. The local point distribution, by creating a variance-covariance matrix M _I shape configuration point I by the following equation, it is possible to quantitatively evaluate (S04).

In the above formula, P _J is the coordinate value of the neighboring shape component point J that is the vicinity of the shape component point I (selected in the same manner as the method for selecting the neighboring measurement points described in the segmentation), and I ^* is the shape near the shape component point I. A set of component points J (neighboring shape component point set), _PIO is a barycentric coordinate value of I ^* . FIG. 6 is a model diagram for explaining the elements P _I , I ^* and P _IO used to create the variance-covariance matrix M _I , where P _I is the coordinate of the shape component point I Value.

By creating the variance-covariance matrix M _I , the eigenvalue γ ^I of the neighboring shape constituent point set I ^* and the eigenvector e ^I corresponding thereto are obtained. Further, since the neighborhood shape component point set I ^* represents a three-dimensional distribution, three eigenvalues, that is, a first eigenvalue γ ^I ₁ , a second eigenvalue γ ^I ₂ , and a third eigenvalue γ ^I ₃ (γ ^I _1). ≧ γ ^I ₂ ≧ γ ^I ₃ ), and the first eigenvector e ^I ₁ , second eigenvector e ^I ₂ , and third eigenvector e ^I ₃ corresponding to each eigenvalue are obtained. Of these, the first eigenvector e ^I ₁ indicates the main direction (main axis direction) of the neighboring shape composing point set I ^* . FIG. 7 is a model diagram showing the point distribution state of the neighboring shape composing point set I ^* and the first eigenvector e ^I ₁ (main direction). In this figure, only the first eigenvector e ^I ₁ and the second eigenvector e ^I ₂ are shown, and the third eigenvector e ^I ₃ shown in the vertical direction of the drawing is omitted.

When the eigenvalue γ ^{I of the} neighboring shape composing point set I ^* is obtained, three feature values S ^I , that is, a first feature value S ^I ₁ , a second feature value S ^I ₂ , and a third feature value are then obtained based on this. Calculate S ^I ₃ . The respective feature amounts are obtained as S ^I ₁ = γ ^I ₁ −α · γ ^I ₂ , S ^I ₂ = γ ^I ₂ −γ ^I ₃ , and S ^I ₃ = β · γ ^I ₃ . Here, α and β are correction coefficients corresponding to the measuring means, and can be set as appropriate. As an experimental example, both α and β are integers, and when α is a value smaller than β, favorable results (high recognition rate of columnar objects) were obtained.

When obtaining three feature quantity S ^I from these magnitude relation, it is possible to impart a shape classified in shape configuration point I (S05). For example, if the first feature amount S ^I ₁ is the maximum, the shape classification point I is given the shape classification “point on the columnar object”, and if the second feature amount S ^I ₂ is the maximum, the shape component point I is assigned. For example, the shape classification “point on the plane” is given, and if the third feature amount S ^I ₃ is the maximum, the shape classification “point on other object” is given to the shape constituent point I.

FIG. 8 is a model diagram in which a shape classification is assigned to the shape component point I when the standing object is a guide sign composed of a vertical column, a horizontal column, and a flat plate. In the first area shown in this figure, the first feature quantity S ^I ₁ is the maximum, and the shape classification point “I” here is given the shape classification “point on the columnar object”. In the second area, the second feature amount S ^I ₂ is maximized, and the shape classification point “I” here is given the shape classification “point on the plane”. Basically, the shape classification is given to all the shape composing points I constituting the standing object shape data, and to all the standing object shape data.

(Vertical columnar judgment of standing objects)
By the procedure so far, the measurement points of the standing object become the segment data, become the standing object shape data, and the shape classification is given to each shape constituent point I. Based on this result, the columnar shape of the standing object is finally determined.

FIG. 9 shows that when the standing object is a guide sign composed of a vertical column, a horizontal column, and a flat plate, a shape classification is given to the shape component point I, and the first eigenvector e ^I ₁ is added. It is the classified model figure. In this figure, Ss indicates a set of shape composing points I of a guide sign, Cs indicates a set of shape composing points I to which a shape classification of “points on a columnar object” is assigned, and Ds. Indicates a set of shape composing points I in which the first eigenvector e ^I ₁ is vertical or nearly vertical in the columnar point set Cs. Note that the “direction close to vertical” means that the angle between the first eigenvector e ^I ₁ and the vertical direction is within a threshold (for example, 30 degrees).

The columnar determination of the standing object can be performed by evaluating the columnar object degree obtained based on the constituent point set Ss, the columnar point set Cs, and the vertical point set Ds (S06). The columnar object degree can be determined by the following equation as an example.

In the above equation, f _s is the columnar object degree, | Ss | is the number of shape component points I included in the component point set Ss, | Cs | is the number of shape component points I included in the columnar point set Cs, and | Ds | Is the number of shape composing points I included in the vertical point set Ds, and w ₁ and w ₂ are weights. In this equation, the columnar object degree f _s is calculated based on the ratio of | Cs | in | Ss | and the ratio of | Ds | in | Cs |. It is possible to calculate only from the ratio occupied by | or only from the ratio occupied by | Ds | of | Cs |. In the experiment example in which both the weights w ₁ and w ₂ are integers and the weight w ₁ is smaller than the weight w ₂ , a favorable result (a high recognition rate of the columnar object) is obtained.

According to the above formula, the columnar object degree f _s is expressed as a percentage, and it can be understood that the larger the value, the more the columnar object can be evaluated (S07). For example, a columnar object degree threshold τ can be set, and a standing object whose columnar object degree f _s is larger than the columnar object degree threshold τ can be recognized as a columnar object. A columnar object can be extracted from a large number of measurement point groups by columnar determination based on the columnar object degree f _s . In addition, at the stage of extracting the columnar object, the height of the standing object shape data is a predetermined height (for example, 2 m) or the number of shape components | Ss | constituting the standing object shape data is a predetermined value (for example, 50 points) The following can be determined as columnar depending on the situation, for example, not recognized as a columnar object regardless of the columnar object degree f _s .

(Method, program, device)
By performing the procedure described above, the columnar object extraction method, the columnar object extraction program, and the columnar object extraction apparatus according to the present invention can be implemented. Specifically, a “ground removal step” for removing ground points (may be omitted in some cases), a “segmentation step” for obtaining segment data, and a “thinning step” for obtaining standing object shape data by performing thinning processing. The columnar object extraction method can be implemented by sequentially performing a “shape classification step” for assigning a shape classification to the shape component point I and a “columnar object determination step” for determining the columnar shape of the standing object. Also, “ground removal function” to remove ground points (may be omitted depending on the case), “segmentation function” to obtain segment data, “thinning function” to obtain standing object shape data by thinning processing, shape configuration A "principal component analysis function" for obtaining eigenvalues and the like for the point I, a "shape classification function" for giving a shape classification to the shape component point I, and a "columnar object determination function" for determining the columnar shape of a standing object, By causing the computer to execute these functions, a columnar object extraction program can be obtained. Furthermore, the “measurement means” for acquiring the measurement means and the measurement point group, the “storage means” having a storage area such as a hard disk for storing the measurement point group, the program, etc., the computer for executing the columnar object extraction program, etc. By providing the “arithmetic processing means”, a “columnar object extraction device” can be obtained.

3. Second Embodiment An example of the second embodiment of the present invention will be described with reference to the drawings. FIG. 10 is a flowchart showing an outline of the second embodiment. Hereafter, each procedure (step) is demonstrated along this flowchart. In the present embodiment, it is not always necessary to perform segmentation (S02), and this step can be omitted. When segmentation (S02) is omitted, naturally the thinning process (S03) is not performed on m segments, but is performed on the entire target measurement point group. The segmentation (S02) to the assignment of shape classification (S05) are the same as those in the “first embodiment” in the present embodiment, and thus the description thereof is omitted here.

(Removal of ground points)
When the measurement point group includes a ground point, in this embodiment as well as the “first embodiment”, a process of removing the ground point from the measurement point group may be performed (S01). However, in the present embodiment, as will be described later, the process of removing the ground can also be performed, and therefore, it is not always necessary to perform the process of removing the ground point.

(Generation of adjacent connection graph)
Here, a process of obtaining adjacent connection graphs is performed by connecting the shape composing points I to which the shape classification is given. When connecting the shape composing points I, first, an arbitrary shape composing point I is focused, and an adjacent point at the shape composing point I is searched (S08). Then, an edge is generated from the searched adjacent point and the shape composing point I. By performing this processing for each shape composing point I, an adjacent connection graph is generated (S09).

The adjacent points at the shape composing point I are searched based on the eigenvector of the shape composing point I. FIG. 11 is a model diagram specifically illustrating a situation in which an adjacent point at the shape composing point I is searched. As shown in this figure, the adjacent points are searched for within a spatial range having a predetermined spread in the eigenvector direction with the shape composing point I as the center. Specifically, the shape configuration point I to the apex, in the eigenvector e ^I is the center line (in FIG theta) a predetermined angle of conical shape with the adjacent points is searched. Furthermore, in addition to the condition that it is within the distance threshold δ from the shape composing point I (that is, the shape composing point I is in the center and within the sphere having the radius δ), the other closest to the shape composing point I within this spatial range The shape composing point I can also be selected as an adjacent point.

When searching for adjacent points at the shape composing point I, a predetermined eigenvector is selected according to the shape classification. As described, shape classification is set based on the feature amount S ^I, for example, "point on cylindrical objects", are classified as "points on a plane", "other points on the object." FIG. 12 is a model diagram showing an eigenvector selected according to the shape classification of the shape component point I, and (a) is a model diagram showing an eigenvector selected when the shape classification is “point on a columnar object”; (B) is a model diagram showing an eigenvector selected when the shape classification is “a point on a plane”, and (c) is a model diagram showing an eigenvector selected when the shape classification is “a point on another object”. It is. As shown in this figure, when the shape classification is “a point on a columnar object”, only the first eigenvector e ^I ₁ (main direction) is selected, that is, two directions from the shape constituent point I (left and right directions in the figure). ), A conical space is set, and adjacent points (that is, a maximum of two adjacent points) are extracted from a predetermined range. Similarly, when the shape classification is “point on the plane”, the first eigenvector e ^I ₁ (main direction) and the second eigenvector e ^I ₂ are selected, that is, four directions from the shape component point I (up, down, left, right in the figure). 4 directions), conical spaces are set, and adjacent points (that is, a maximum of four adjacent points) are extracted from a predetermined range. When the shape classification is “a point on another object”, the first eigenvector e ^I ₁ (main direction), the second eigen vector e ^I ₂ , and the third eigen vector e ^I ₃ are selected, that is, the shape component point I Are respectively conical spaces in six directions, and adjacent points (that is, a maximum of six adjacent points) are extracted from a predetermined range.

(Formation of basic components)
Here, the process is performed by connecting the shape component points I to which the same shape classification is given, and obtaining one continuous component as a basic component (S10). When connecting the shape component points I, first, pay attention to an arbitrary shape component point I, and search for an adjacent point at the shape component point I. Build one area. Subsequently, the added adjacent point is focused on as a new shape composing point I, and the adjacent point is added to the region by the same processing. Then, the adjacent points are repeatedly added until the adjacent points with the same shape classification are not searched, and one region (continuous component) obtained as a result is a basic component. As described above, a method of expanding a region from an arbitrary point based on a predetermined condition is also referred to as a region growth method, that is, a basic component is generated using a region growth method.

  FIG. 13 is a model diagram illustrating a state in which a plurality of different basic components are generated. As shown in this figure, a basic component is generated for each region having a different shape classification, and even if the basic components are different, there is an adjacency relationship between the shape component points I that satisfy the condition of being an adjacent point. An edge is generated.

(Classification of basic components)
In the classification of the basic component, processing for classifying the basic component according to the characteristic is performed (S11a to S11d). Here, a ground component that forms the ground, a connected component that forms a connected object such as an electric wire, a post component that forms a support column, and a wall surface component that forms a wall surface will be described.

  The ground component based on the ground point does not constitute a columnar object. Therefore, when the columnar object is accurately extracted, it is desirable to remove the ground component. The ground component is recognized from the basic components based on the following conditions. Since the ground component is planar, a basic component whose shape classification is “point on the plane” is selected. And the thing whose basic component is substantially horizontal is recognized as a ground component (S11a). At this time, it can be added to the condition that the number of shape composing points I constituting the basic component is a certain number or more.

The determination as to whether or not the basic component is substantially horizontal can be made based on the shape composing point I constituting the basic component. That is, adjacent shape constituent points I are connected to form a large number of triangular surfaces. Then, there are a plurality of triangular surfaces around the shape composing point I, and the normal of the shape composing point I based on the normal direction of the triangular surfaces (for example, adopting an average value in the normal direction). The direction can be determined. Alternatively, the third eigenvector e ^I ₃ can be simply set as the normal direction of the shape component point I. When the normal direction is obtained for all the shape composing points I in the basic component, the normal direction is divided into a predetermined range and a histogram is created. If the number of shape component points I recorded in the maximum range of the histogram (a range close to 90 degrees) is the largest, it can be determined that the basic component is generally horizontal as a whole.

  As shown in FIG. 14, the connection component based on the measurement point obtained by measuring the connection object such as an electric wire does not constitute a columnar object, and therefore, when the columnar object is accurately extracted, it is desirable to remove the connection component. . To recognize the connected component from the basic components, it is performed based on the following conditions. Since the connected component is linear, a basic component whose shape classification is “point on columnar object” is selected. Then, a component whose basic component is substantially horizontal is recognized as a connected component (S11b). At this time, it can be added to the condition that the number of shape composing points I constituting the basic component is a certain number or more.

The determination as to whether or not the basic component is substantially horizontal can be made based on the shape composing point I constituting the basic component. That is, if the first eigenvector e ^I ₁ (main axis direction) of the shape composing point I is substantially horizontal (within the threshold range from the horizontal direction), it can be determined that the basic component is generally horizontal as a whole.

  The column component is a basic basic component constituting the columnar object. In order to recognize the strut component from the basic components, the following conditions are used. Since the connected component is linear, a basic component whose shape classification is “point on columnar object” is selected. Then, a component whose basic component is substantially vertical (within a threshold range from the vertical direction) is recognized as a column component (S11c). At this time, it can be added to the condition that the number of shape composing points I constituting the basic component is a certain number or more.

Instead of the condition that the basic component is substantially vertical, it is also possible to recognize the column component on the condition that the cross-sectional shape of the basic component is substantially circular. FIG. 15 is a model diagram showing that the cross-sectional shape of the column component forms a part of a substantially circular shape. The cross-sectional shape of the basic component is considered to be a cross-section cut perpendicular to the first eigenvector e ^I ₁ (major axis direction) of the shape component point I, and the shape component point I projected there is a circle (or a circular one). If it is in a position close to (part), it is recognized as a strut component. At this time, whether or not the position is close to a circle can be determined using a technique such as RANSAC.

  As shown in FIG. 13, when a plurality of different support components are generated, only the lowest basic component can be recognized as a support component. In this case, in addition to the above-described conditions, the ground component is excluded from the plurality of connected basic components, and the pillar component has a minimum elevation value (Z of the three-dimensional coordinates) among the plurality of pillar components (candidates). As a condition to recognize as

  A wall surface component based on a measurement point at which a wall surface is measured does not constitute a columnar object. Therefore, when a columnar object is accurately extracted, it is desirable to remove the wall surface component. Recognizing wall surface components from basic components is performed based on the following conditions. Since the wall surface component is planar, a basic component whose shape classification is “point on the plane” is selected. And the thing whose basic component is substantially vertical is recognized as a wall surface component (S11d). At this time, it can be added to the condition that the number of the shape composing points I constituting the basic component is a certain number or more, or that the basic component is a predetermined area or more.

  Whether the basic component is substantially vertical or not is determined using the method described for the ground component. If the number of shape component points I recorded in the minimum range of the histogram (a range close to 0 degrees) is the largest, It can be determined that the basic component is generally vertical overall.

(Create columnar object segment)
When the column component is recognized, a columnar object segment is created based on the column component (S12). An edge is generated even if the shape component point I adjacent to the shape component point I in the column component constitutes another basic component. In other words, it may be linked even different basic components. As shown in FIG. 13, basic components that are successively connected from the column components are gathered to create one columnar object segment. At this time, it is desirable to create the columnar object segment after removing the above-mentioned ground component and connected component. If there is no other basic component to be connected, a columnar object segment is created by a single column component.

  The columnar object segment created as described above can be recognized as a columnar object, that is, a columnar object can be accurately extracted from a large number of measurement point groups.

(Method, program, device)
By performing the procedure described above, the columnar object extraction method, the columnar object extraction program, and the columnar object extraction apparatus according to the present invention can be implemented. Specifically, "Segmentation process" to obtain segment data, "Thinning process" to obtain standing object shape data by performing thinning process, "Shape classification process" to give shape classification to shape constituent points, The columnar object extraction method can be implemented by sequentially performing the “columnar object determination step” for determining the columnar shape of a standing object. The “columnar object determination process” includes an “adjacent connection graph generation process” that obtains an adjacent connection graph, a “basic component formation process” that obtains a basic component, and a “basic component classification process” that is divided into various components such as a “post component”. The “columnar object segment creating step” for creating the columnar object segment is included. "Segmentation function" to obtain segment data, "Thinning function" to obtain standing object shape data by performing thinning processing, "Principal component analysis function" to obtain eigenvalues etc. for shape composition points, Shape composition It is equipped with a “shape classification function” that gives shape classification to points, a “columnar object determination function” that performs columnar determination of standing objects, and a “columnar object determination function” that also obtains an adjacent connection graph “adjacent connection graph generation” "Function", "Basic component formation process function" to obtain basic components, "Basic component classification function" to divide into various components such as "Post component", and "Columnar object segment creation function" to create columnar object segments Yes. By causing the computer to execute these functions, a columnar object extraction program can be obtained. Furthermore, the “measurement means” for acquiring the measurement means and the measurement point group, the “storage means” having a storage area such as a hard disk for storing the measurement point group, the program, etc., the computer for executing the columnar object extraction program, etc. By providing the “arithmetic processing means”, a “columnar object extraction device” can be obtained.

4). Third Embodiment The technology for recognizing a columnar object has been described so far in the first embodiment and the second embodiment. On the other hand, the “columnar object” is a relatively broad concept, and is generally recognized by a specific type such as a utility pole, a streetlight, and a sign rather than a columnar object. Here, a method for classifying objects recognized as columnar objects into more specific types will be described. Here, three types of poles, street lamps, and signs are taken as examples of columnar objects.

  FIG. 16 is a flowchart showing a procedure for classifying a columnar object. In order to classify columnar objects, the characteristics of utility poles, street lamps, and signs are used. Here, what remove | excluded the support | pillar (thing recognized as a pillar) from the structure of a columnar object is considered as a "part", and classification is performed paying attention to the characteristic of the part.

(Extraction of strut candidate points)
Based on the shape classification of the shape component point I given in the first embodiment and the second embodiment, and the first eigenvector e ^I ₁ (main direction) of the shape component point I, column candidate points are extracted ( S13 in FIG. As described, shape classification is set based on the feature amount S ^I, for example, "point on cylindrical objects", are classified as "points on a plane", "other points on the object." A shape component point I whose shape classification is “point on columnar object” and whose first eigenvector e ^I ₁ (main direction) is substantially vertical (within a threshold range from the vertical direction) is extracted as a column candidate point. The At this time, if the strut candidate points are projected onto a predetermined plane, only those close to a circle with a predetermined radius can be left as the strut candidate points. In this case, it can be determined using a technique such as RANSAC. In the second embodiment, the shape composing point I constituting the strut component may be extracted as a strut candidate point.

  Next, the shape composing point I close to the strut candidate point, specifically, the shape composing point I within a predetermined threshold (distance) from the strut candidate point is also extracted. The shape composing point I extracted here and the strut candidate point are combined and determined as a strut composing point (S14).

  Next, out of the shape constituent points I constituting the standing object shape data (first embodiment) and the columnar object segment (second embodiment), the ones that are the pillar constituent points are excluded. As a result, the shape composition point I constituting the part is obtained. The shape constituent points I constituting these parts are connected by edges (in the case of the second embodiment, basic components (FIG. 13)), and continuous components connected by the edges are extracted as parts. (S15).

  The part is a part obtained by removing the column from the standing object shape data (first embodiment) and the columnar object segment (second embodiment). A single standing object shape data or a columnar object segment (hereinafter collectively referred to as “columnar object segment etc.” in this embodiment) may have a plurality of parts, or only one There is also. Moreover, a part may be comprised from one individual element (in the case of 2nd Embodiment, a basic component), and a some individual element may be connected and comprised. Parts classification categorizes such various configurations into preset types (part types) (S16). FIG. 17 is a model diagram illustrating examples of part types classified based on the shape classification of individual elements and combinations of individual elements. For example, a part consisting only of individual elements whose shape classification is “point on a columnar object” is classified into a part type “A” (A), and consists only of individual elements whose shape classification is “point on a plane”. The parts are classified into a part type of “surface” (B), and parts consisting only of individual elements whose shape classification is “points on other objects” are classified into a part type of “solid” (C). In addition, a part constituted by an individual element whose shape classification is “point on a columnar object” and an individual element whose shape classification is “point on a plane” is classified into a part type of “bar + plane” ( D). Here, eight types of parts are set. That is, the part type A is “bar”, the part type B is “surface”, the part type C is “solid”, the part type D is “bar + surface”, the part type E is “bar + solid”, and the part type F is “Surface + solid”, part type G is “bar + plane + solid”, and part type H is “other”.

  When a part type is assigned to each of the extracted parts, the degree of attribution is calculated for each columnar object segment or the like (S17). The degree of attribution refers to the degree to which the columnar object segment belongs to the utility pole, streetlight, and sign. The largest value is calculated from the three types of utility pole attribution, streetlight attribution, and sign attribution. The type is determined according to the degree of attribution indicating (S18).

Each degree of attribution is calculated based on the evaluation of the height of the columnar object segment, the number of parts included in the columnar object segment, etc., and the part type.

The fi on the left side of the above equation is the degree of attribution. When i = 1, the utility pole degree is obtained, when i = 2, the street lamp degree is obtained, and when i = 3, the sign degree is obtained. Further, hs on the right side is the height of the columnar object segment, ns is the number of parts included in the columnar object segment, Ts is an evaluation regarding the part type, and f _Hi (hs) is the height of the columnar object segment. , F _Ni (ns) is a function depending on the number of parts, and f _Ti (Ts) is a function depending on the part type. f _Hi (hs), f _Ni (ns), and f _Ti (Ts) are respectively obtained by the following equations.

As shown in FIG. 18, h _{min in} equation (6-1) has a different value depending on the type. For example, h _min = 6.5 m for a power pole, h _min = 3.0 m for a streetlight, In this case, h _min = 2.0 m can be set. Similarly, h _max varies depending on the type. For example, h _max = 15.0 m for a power pole, h _max = 12.0 m for a streetlight, and h _max = 6.5 m for a sign.

  As shown in FIG. 18, the parameters a and b in the equation (6-2) have different values depending on the type. For example, a = 0.1 and b = 0 for a utility pole, and a = -0 for a streetlight. .2, b = 1.0, and in the case of a label, a = −0.3, b = 0.9.

The “number of corresponding part types” in the equation (6-3) is the number of parts corresponding to a part type predetermined for each type. For example, “part type C or part type H” for a power pole, “part type A or part type E” for a streetlight, “part type B, part type D, part type F, or part type for a sign” G ”can be set. In this case, if the number of parts ns is 8 in a columnar object segment, etc., 1 part type A, 1 part type B, 2 parts type C, 4 parts type H, f _T1 (Ts) = 6/8 = 0.75, street light f _T2 (Ts) = 1/8 = 0.125, and sign f _T3 (Ts) = 1/8 = 0.125.

Thus, for each columnar object segment, etc., f _Hi (hs), f _Ni (ns), and f _Ti (Ts) are calculated for each pole, street lamp, and sign, and by adding these, the degree of utility of the pole, street lamp The degree of attribution and the degree of attribution of the label are respectively determined. For example, when the utility column has the highest degree of attribution, the columnar object segment is evaluated as “electric pole”, and when the street lamp has the greatest degree of attribution, the columnar object segment is evaluated as “streetlight”. And the columnar object segment and the like are evaluated as “signs”.

(Type recognition based on context features)
In recognition of the type of columnar object (electric pole, streetlight, sign) based on the degree of attribution described, it may be possible to make an incorrect recognition in some cases. Here, a method for more accurately recognizing the type of columnar object using the context feature amount will be described.

  Normally, utility poles and street lamps are arranged at equal intervals. In addition, there may be regularity in the positional relationship between the utility pole and the streetlight. Focusing on such “arrangement”, more accurate type recognition is performed by comparing the standard arrangement model with the arrangement of the periphery of the columnar object segment or the like to be evaluated. FIG. 19 is a flowchart showing a procedure for classifying the type of columnar object based on the context feature amount. As described, for each columnar object segment, etc., the degree of attribution (three types of degree of belonging: utility pole, street lamp, sign) is calculated (S17).

  A grid is set in a planar shape with respect to a target range (for example, a target city area) (S19). The grid is set at an arbitrary interval so that the vertical axis and the horizontal axis intersect. At this time, the intersection of the vertical axis and the horizontal axis can be orthogonal (at an intersection angle of 90 degrees), or another oblique angle can be selected and intersected.

  When the grid is set, a large number of meshes (mesh) surrounded by the grid are formed. Since the columnar object segment or the like has a plane coordinate, it can be recognized which mesh the columnar object segment or the like exists in or which mesh is closest to. Therefore, it is possible to arrange columnar object segments or the like on the mesh, and it is possible to assign the degree of attribution of the columnar object segments or the like (three types of attribution of electric poles, street lamps, and signs) to the mesh (S20).

  Next, a range (area) to be evaluated is set (S21). That is, to what extent the peripheral range of the columnar object segment or the like to be evaluated is to be expanded is determined. Here, since the mesh is already formed, the area is set by determining the number of meshes (vertical number × horizontal number) as shown in FIG.

  When the area is set, a “standard distribution map” which is a standard arrangement is created (S22). This standard distribution map is created for a set area. The standard distribution map is created based on, for example, the degree of signage that is assigned around the utility pole. The standard distribution map of the utility pole is assigned to the utility pole, streetlight, and sign. In the same manner, three types of standard distribution maps of street lamps and three types of standard distribution maps of signs are generated. Note that the standard arrangement may be determined by a person, or may be obtained by calculation, for example, by integrating the arrangements having a high degree of belonging and taking an average value.

  Focusing on the columnar object segment or the like to be evaluated, an area is set around this, and three types of maps (peripheral distribution maps) are created in which the degrees of attribution of the electric poles, street lamps, and signs in the vicinity are assigned (S23). ). FIG. 21 is a model diagram showing three types of peripheral distribution maps.

  When the standard distribution map is created and the peripheral distribution map is created, a context feature amount is calculated (S24). FIG. 22 is a model diagram illustrating a method for calculating a context feature amount. As shown in this figure, the context feature amount is calculated by associating the peripheral distribution map with the standard distribution map.

  First, a standard distribution map of power poles is prepared on the assumption that the columnar object segment or the like of interest is “electric pole”. At this time, there are three types of peripheral distribution maps in which utility poles, street lamps, and signage are assigned, and three types of standard distribution maps of utility poles, street lamps, and signs are prepared. Since the meshes formed on the maps are arranged based on the set area, the arrangement and the number are the same. The surrounding distribution map of the utility pole attribution is compared with the standard distribution map of the utility pole attribution, and the utility pole attribution of the same mesh is multiplied. Next, the standard distribution map of street lamp attribution and the surrounding distribution map are compared, the utility pole attribution of the same mesh is multiplied, and the standard distribution map of sign attribution and the surrounding distribution map are compared, and the same mesh Multiply by utility pole attribution. Finally, the sum of the multiplication values of all meshes (electric poles, street lamps, signs) calculates the context feature quantity of the electric pole.

  In the same procedure, prepare the standard distribution map of the streetlight, assuming that the columnar object segment etc. to be focused on is “streetlight”, and compare it with the surrounding distribution map where the streetlight attribution is scored. Calculate the amount. Further, the context feature amount of the sign is calculated in the same procedure. As a result, the type of the columnar object segment or the like can be determined based on the context feature value showing the largest value. For example, when the context feature value of the utility pole shows the largest value, the columnar object segment etc. is evaluated as “electric pole”, and when the context feature value of the streetlight shows the largest value, the columnar object segment etc. And the columnar object segment and the like are evaluated as “signs”.

  FIG. 23 shows a specific example in which the context feature amount is calculated. In this figure, an area of horizontal 9 mesh × vertical 1 mesh is set, and the peripheral distribution map, standard distribution map, and context feature amount are shown in this order from the top. In the perimeter distribution map, the utility poles are assigned around the columnar object segment, etc. (lane mesh) of interest. The second line is the street lamp attribute, and the third line is the sign attribute. doing.

  The standard distribution map is shown in the order of electric poles, street lamps, and signs. The first line is assigned the degree of ownership of the electric pole, the second line is assigned the degree of attribution of the street lamp, and the third line is the sign. The degree of attribution is assigned.

  The context feature value is the context feature value calculated by comparing the peripheral distribution map and the standard distribution map of the power pole, the context feature value calculated by checking the peripheral distribution map and the standard distribution map of the streetlight, the standard of the peripheral distribution map and the sign The context feature values calculated by comparing the distribution map are shown in this order. The first line is the value obtained by multiplying the degree of attribution of the utility pole corresponding to each mesh, and the second line is the degree of attribution of the streetlight for each mesh. The third row is a value obtained by multiplying the degree of attribution of the sign corresponding to each mesh. As a result, since the context feature value (3.57) of the utility pole is the largest value, the columnar object segment and the like can be evaluated as “electric pole”.

(Method, program, device)
By performing the procedure described above, the columnar object extraction method, the columnar object extraction program, and the columnar object extraction apparatus according to the present invention can be implemented.

  The columnar object extraction method, the columnar object extraction program, and the columnar object extraction apparatus of the present invention are inventions that extract columnar objects from standing objects. By applying this invention, a board such as a signboard or a wall surface is used. It is possible to extract an object that is classified as a columnar object or a plate-like object. It is also possible to extract a combination of a columnar object and a plate-like object such as a sign attached to a pole. As a result, it can be used for various standing structure management.

i Arbitrary measurement point j Neighbor measurement point q Other neighboring measurement points r (Limit distance) radius I ^* Neighboring shape composing point set PI Coordinate value of _I shape composing point _PIO I ^* Gravity center coordinate value P _J Coordinate value θ included angle Ss component point set Cs columnar point set Ds vertical point set

Claims (17)

In a method of extracting a columnar object having a columnar shape from the measurement point group that is a set of measurement points in a predetermined space including a plurality of standing objects standing on the ground,
For the measurement point group , a thinning step for obtaining standing object shape data by performing a thinning process to clarify the shape characteristics of the standing object; and
A shape classification step for obtaining shape eigenvalues and eigenvectors by performing principal component analysis on the shape composing points constituting the standing object shape data, and giving a shape classification based on a feature amount calculated from the eigenvalues;
A columnar object determination step of performing columnar determination of a standing object based on the shape classification of the shape composing points, and
The thinning process in the thinning step is to move the position of an arbitrary measurement point, and an angle between an arbitrary measurement point, a nearby measurement point close to this point, and two nearby measurement points at the arbitrary measurement point The position of the arbitrary measurement point after movement is calculated so that the influence of the arbitrary measurement point remains stronger and the influence of the nearby measurement point is suppressed as the thinning intensity is smaller. ,
The columnar object extraction method characterized in that the thinning intensity indicates a smaller value as the included angle between two adjacent measurement points at an arbitrary measurement point is smaller. The eigenvalues obtained by principal component analysis are first eigenvalue, second eigenvalue, and third eigenvalue in descending order of values,
The eigenvector obtained by principal component analysis is the first eigenvector corresponding to the first eigenvalue, the second eigenvector corresponding to the second eigenvalue, and the third eigenvector corresponding to the third eigenvalue,
The feature amount includes a first feature amount calculated based on the first eigenvalue and a second eigenvalue, a second feature amount calculated based on the second eigenvalue and a third eigenvalue, and the third eigenvalue. A third feature amount calculated on the basis of,
The shape classification is composed of two or more classifications including “points on a columnar object”, and the shape classification point having the maximum value of the first feature amount among the feature amounts is defined as “on the columnar object”. The columnar object extraction method according to claim 1, wherein the point is given. The columnar object determination step includes an adjacent connection graph generation step, a basic component formation step, a basic component classification step, and a columnar object segment creation step,
The adjacent connection graph generation step obtains an adjacent point adjacent to each shape constituent point based on the eigenvector of each shape constituent point, and generates an edge between the adjacent point and the shape constituent point. , One continuous component obtained as a result is obtained as one adjacent connection graph,
In the basic component forming step, for each shape composing point constituting the adjacent connection graph, the adjacent point assigned the same shape classification as each shape composing point is obtained, and the adjacent point and the shape composing point are determined. Are generated, and one continuous component obtained as a result is obtained as one basic component.
The basic component classification step classifies components corresponding to the shape classification with respect to the basic component, and a basic component whose shape classification is “point on a columnar object” is a “post component”. Classify and
The columnar object segment creating step extracts the adjacent connection graph including the column component and combines one or more basic components included in the adjacent connection graph to create a columnar object segment. ,
As the adjacent point, a shape composing point closest to the shape composing point is selected within a spatial range that extends from the shape composing point toward the eigenvector direction,
The columnar object extraction method according to claim 2, wherein the columnar determination of the standing object is performed based on the columnar object segment.   In the basic component classification step, when two or more basic components whose shape classification is “points on columnar objects” are included in one adjacent connection graph, the basic classification is “points on columnar objects”. 4. The columnar object extraction method according to claim 3, wherein the component having the minimum elevation value is classified as a “post component”. The shape component point at which the second feature amount has the maximum value is given a “point on the plane” as the shape classification,
In the basic component classification step, a basic component whose shape classification is “point on plane” is classified as “plane component”, and based on the normal direction of the surface formed by the shape component points constituting the plane component , Extract the almost horizontal plane component as “ground component”,
5. The columnar object segment creating step is configured to create the columnar object segment by combining one or more of the basic components after removing the ground component. Columnar object extraction method. In the basic component classification step, a basic component whose shape classification is “a point on a columnar object”, and a basic component in which the first eigenvector of a shape component point constituting the basic component is substantially horizontal is referred to as a “connected component”. Extract as
6. The columnar object segment creating step of creating a columnar object segment by combining one or more basic components after excluding the connected component. The columnar object extraction method according to any one of the above. Among the basic components constituting the columnar object segment, the basic components other than the prop components, and a combination of the basic components connected to each other is extracted as a part,
The parts are classified into part types based on the shape classification,
Based on the evaluation of the height of the columnar object segment, the number of parts included in the columnar object segment, and the part type, the degree of attribution of the columnar object segment is determined.
The columnar object extraction method according to any one of claims 3 to 6, wherein a type of the columnar object is selected based on the degree of belonging. A plurality of meshes are configured by a predetermined grid, and the attribution is assigned to the closest mesh,
Set an area composed of a predetermined number of meshes,
Based on the area, create a standard distribution map for each type of columnar object,
Based on the area, create a peripheral distribution map for each columnar object segment or each standing object shape data,
Further, based on the standard distribution map and the peripheral distribution map, the context feature amount is calculated for the columnar object segment or the standing object shape data,
The columnar object extraction method according to claim 7, wherein the type of the columnar object is selected based on the context feature amount. The computer executes a function of extracting a columnar object having a columnar shape from the measurement point group, which is a set of measurement points in a predetermined space including a plurality of standing objects standing on the ground. In the columnar object extraction program to let
For the measurement point group , a thinning function that creates a standing object shape data by performing a thinning process to clarify the shape characteristics of the standing object; and
A principal component analysis function for calculating eigenvalues and eigenvectors by performing arithmetic processing for principal component analysis on the shape constituent points constituting the standing object shape data,
Furthermore, a shape classification function for assigning a shape classification according to the feature amount calculated based on the eigenvalue to each of the shape constituent points;
A columnar object determination function for performing columnar determination processing of a standing object based on the feature amount of the shape composing point, and
The thinning process performed by the thinning function moves the position of an arbitrary measurement point, and includes an arbitrary measurement point, a nearby measurement point close to this point, and two nearby measurement points at the arbitrary measurement point. Based on the thinning strength obtained from the included angle, the position of the arbitrary measurement point after movement is calculated so that the influence of the arbitrary measurement point remains stronger and the influence of the nearby measurement point is suppressed as the thinning strength decreases. Is,
The columnar object extraction program characterized in that the thinning intensity indicates a smaller value as the included angle between two adjacent measurement points at an arbitrary measurement point is smaller. The eigenvalues obtained by the principal component analysis function are, in order from the largest value, the first eigenvalue, the second eigenvalue, and the third eigenvalue,
The eigenvector obtained by the principal component analysis function is a first eigenvector corresponding to the first eigenvalue, a second eigenvector corresponding to the second eigenvalue, and a third eigenvector corresponding to the third eigenvalue,
The feature amount includes a first feature amount calculated based on the first eigenvalue and a second eigenvalue, a second feature amount calculated based on the second eigenvalue and a third eigenvalue, and the third eigenvalue. A third feature amount calculated on the basis of,
The shape classification is composed of two or more classifications including “points on a columnar object”, and the shape classification point having the maximum value of the first feature amount among the feature amounts is defined as “on the columnar object”. The columnar object extraction program according to claim 9, further comprising: The columnar object determination function includes an adjacent connection graph generation function, a basic component formation function, a basic component classification function, and a columnar object segment creation function,
The adjacent connection graph generation function selects an adjacent point adjacent to each shape component point based on the eigenvector of each shape component point, and generates an edge between the adjacent point and the shape component point. Then, one continuous component obtained as a result is used as one adjacent connection graph,
The basic component forming function selects, for each shape composing point constituting the adjacent connection graph, the adjacent point assigned the same shape classification as each shape composing point, and the adjacent point and shape composing point. An edge is generated between and a continuous component obtained as a result is used as a basic component,
The basic component classification function assigns a component classification corresponding to the shape classification to the basic component, and the basic component whose shape classification is “point on a columnar object” is a component. We give "post component" as classification,
The columnar object segment creation function extracts the adjacent connection graph including the column component and combines one or more basic components included in the adjacent connection graph to create a columnar object segment. ,
As the adjacent point, a shape composing point closest to the shape composing point is selected within a spatial range that extends from the shape composing point toward the eigenvector direction,
The columnar object extraction program according to claim 10, wherein the columnar determination of a standing object is performed based on the columnar object segment.   When the basic component classification function includes two or more basic components whose shape classification is “point on a columnar object” in one adjacent connection graph, the basic classification is “point on columnar object”. The columnar object extraction program according to claim 11, wherein a “post component” is assigned as a component classification to a component having a minimum elevation value among components. To the shape component point at which the second feature amount has the maximum value, a “point on the plane” is given as the shape classification,
The basic component classifying function assigns a “plane component” as a component classification to a basic component whose shape classification is “point on a plane”, and a surface formed by shape composing points constituting the plane component. Based on the normal direction of, we extract a substantially horizontal planar component as a “ground component”
13. The columnar object segment creation function creates the columnar object segment by combining one or more basic components after removing the ground component. Columnar object extraction program. The basic component classifying function is a basic component whose shape classification is “a point on a columnar object”, and a basic component in which the first eigenvector of the shape composing point constituting the basic component is substantially horizontal is “connected component”. Extract as
14. The columnar object segment creation function is configured to create a columnar object segment by combining one or more basic components after excluding the connected component. The columnar object extraction program according to any one of the above. Among the basic components constituting the columnar object segment, a basic component other than the pillar component, and a part extraction function for extracting a combination of the basic components connected to each other as a part;
A parts type classification function for classifying the parts into part types based on the shape classification;
Based on the height of the columnar object segment, the number of parts included in the columnar object segment, and the evaluation related to the part type, an attribute calculation function for calculating the attribute of the columnar object segment,
The columnar object extraction program according to any one of claims 11 to 14, further comprising: a columnar object type recognition function for selecting a columnar object type based on the degree of belonging. A plurality of meshes are configured by a predetermined grid, and the attribution score function for scoring the attribution score to the nearest mesh,
An area setting function for setting an area composed of a predetermined number of meshes;
Based on the area, a standard distribution map creation function for creating a standard distribution map for each type of columnar object,
Based on the area, a peripheral distribution map creating function for creating a peripheral distribution map for each columnar object segment or each standing object shape data,
Further, based on the standard distribution map and the peripheral distribution map, a context feature amount calculation function for calculating a context feature amount for the columnar object segment or the standing object shape data;
The columnar object extraction program according to claim 15, further comprising: a columnar object type recognition function for selecting a columnar object type based on the context feature amount. In a columnar object extraction device that extracts a columnar object having a columnar shape from among measurement points that are a set of measurement points in a predetermined space including a plurality of standing objects standing on the ground. ,
Measuring means for measuring the predetermined space and acquiring the measurement point group;
Storage means for storing the measurement point group acquired by the measurement means;
Computation processing means for extracting the columnar object from the standing object based on the measurement point group,
Further, the arithmetic processing means includes a thinning function, a principal component analysis function, a shape classification function, and a columnar object determination function,
The thinning function is to create standing structure shape data by performing thinning processing to clarify the shape characteristics of the standing object for the measurement point group ,
The principal component analysis function is to obtain eigenvalues and eigenvectors by performing arithmetic processing for principal component analysis for each of the shape component points constituting the standing object shape data,
The shape classification function is to give a shape classification corresponding to a feature amount calculated based on the eigenvalue to each of the shape constituent points,
The columnar object determination function performs a columnar determination process of a standing object based on the feature amount of the shape composing point,
The thinning process performed by the thinning function moves the position of an arbitrary measurement point, and includes an arbitrary measurement point, a nearby measurement point close to this point, and two nearby measurement points at the arbitrary measurement point. Based on the thinning strength obtained from the included angle, the position of the arbitrary measurement point after movement is calculated so that the influence of the arbitrary measurement point remains stronger and the influence of the nearby measurement point is suppressed as the thinning strength decreases. Is,
The columnar object extraction apparatus according to claim 1, wherein the thinning intensity indicates a smaller value as the included angle between two adjacent measurement points at an arbitrary measurement point is smaller.

Images