Feature type estimation system

Abstract

To provide a feature type estimation system capable of estimating a type of feature formed by measurement points by using a plurality of constituent points having three-dimensional coordinates.SOLUTION: A feature type estimation system of the present invention, which is a system for estimating a type of a feature on the basis of a plurality of constituent points having three-dimensional coordinates, includes 3D terrain model storage means, learning means, and feature type output means. An input data set includes height information of constituent points included in an input region of interest. The learning means generates a feature estimation model by machine-learning a type of feature and height information of the constituent points contained in a learning data set. The feature type output means outputs a type of feature using the height information of the constituent points included in the input data set as an input value.SELECTED DRAWING: Figure 8

Description

[Technical Field]

[0001] The present invention relates to a technology for estimating the type of a feature by using a set (hereinafter referred to as a "component point group") of points (hereinafter referred to as "component points") that constitute the feature and have three-dimensional coordinates, and more specifically to a feature type estimation system that estimates the type of feature based on elevation information of component points contained within a specified area. [Background technology]

[0002] In recent years, with the advancement of measurement technology, the demand for topographical information (spatial information) has increased, and an increasing number of managers are requesting spatial information on facilities, such as their shape and location, in order to more effectively manage facilities installed on and along roads. Furthermore, the realization of Society 5.0, which is currently being promoted jointly by the public and private sectors, also sees advanced maintenance and management of social infrastructure (hereinafter simply referred to as "social infrastructure") as a key issue, and for this reason spatial information on facilities, etc., has become extremely important and indispensable information.

[0003] Previously, to obtain spatial information on road facilities, direct surveying of the target feature was performed using Total Station (TS) or similar, but recently measurement methods that obtain comprehensive spatial information without targeting specific features have often been adopted. For example, aerial photogrammetry and Mobile Mapping System (MMS) are typical measurement methods that obtain comprehensive spatial information. Among these, MMS is a method that uses mobile vehicles equipped with sensors such as laser scanners, cameras, a satellite positioning system (GNSS: Global Navigation Satellite System) to obtain their own position, an Inertial Measurement Unit (IMU), and odometry. In other words, it is a measurement technology that can obtain a large number of measurement points using laser scanners while moving along the roadway.

[0004] Measurement using a laser scanner (hereafter simply referred to as "laser measurement") is a technique in which a laser pulse is emitted onto the target to be measured and the reflected signal is received to obtain the measurement. More specifically, the distance from the emission position to the measurement point (the point where the laser pulse is reflected) is calculated by measuring the time difference between the emission and reception times. Furthermore, the emission position (x, y, z) of the laser pulse is acquired using a positioning method such as GNSS, and the emission attitude (ω, φ, κ) is acquired using an inertial measurement method such as an IMU, thereby obtaining the 3D coordinates of the measurement point. In laser measurement using an MMS, the laser scanner's laser emitter rotates internally while emitting the laser. As a result, the laser is emitted in a direction oblique to the roadway axis as the vehicle travels, obtaining a large number of measurement points.

[0005] However, each measurement point obtained by aerial photogrammetry or MMS basically only contains coordinate information, making it difficult to understand the type of feature that the measurement point constitutes. For example, in the past, operators would identify the type of feature that a measurement point constitutes by comparing photographs or video of the corresponding area with existing floor plans, or by having workers carry out on-site surveys using existing floor plans. However, these methods increase the cost of identifying the type of feature and inevitably lead to operator misjudgment (human error).

[0006] If it were possible to automatically identify the type of feature from a large amount of laser measurement data, the workload of the operator would be significantly reduced and human error could be eliminated. Therefore, Patent Document 1 discloses an invention that can automatically extract pole-shaped features such as streetlights, utility poles, and signs from a large amount of measurement data obtained by laser measurement. [Prior art documents] [Patent documents]

[0007] [Patent Document 1] Japanese Patent Application Laid-Open No. 2017-9616 Summary of the Invention [Problem to be solved by the invention]

[0008] The invention disclosed in Patent Document 1 can automatically extract columnar features from a measurement point cloud obtained by an MMS, and can also select planar objects (such as signs and billboards) and three-dimensional objects that columnar features have. This significantly reduces the operator's workload, shortens the work time, and avoids variations and errors associated with individual operators. However, the invention of Patent Document 1 only allows for columnar features to be automatically extracted. In other words, while it can estimate the type of columnar features, it is difficult to estimate the type of each of the diverse types of features, including columnar features.

[0009] The object of the present invention is to solve the problems associated with the prior art, namely, to provide a feature type estimation system that can estimate the type of feature constituted by multiple measurement points having three-dimensional coordinates. [Means for solving the problem]

[0010] The present invention is based on the unprecedented idea of ​​noting that each feature has a different pattern of elevation information for its constituent points, and estimating the type of feature based on a combination of elevation information for multiple constituent points within a specified area based on the feature.

[0011] The feature type estimation system of the present invention is a system for estimating the type of a feature based on a plurality of constituent points having three-dimensional coordinates, and includes a three-dimensional terrain model storage means, a learning means, and a feature type output means. The three-dimensional terrain model storage means stores a three-dimensional terrain model based on the plurality of constituent points. The learning means generates a feature estimation model for outputting the type of the feature by machine learning training data sets prepared for each type of feature. The feature type output means outputs the type of the feature associated with the input dataset based on the input input dataset and the feature estimation model. The input dataset includes elevation information for a plurality of constituent points included in an input region of interest. The training region is a two-dimensional region having a predetermined size and shape and based on a training reference point selected from the feature. The input region is a two-dimensional region having the same size and shape as the training region and based on an input reference point specified by an operator. The learning means generates a feature estimation model by machine learning the type of feature and the height information of the multiple constituent points included in the training dataset, and the feature type output means outputs the type of feature using the height information of the multiple constituent points included in the input dataset as an input value.

[0012] In the feature type estimation system of the present invention, the training dataset may include elevation information of the component points corresponding to a plurality of predetermined selection statistical values ​​among the plurality of component points, and the input dataset may include elevation information of the component points corresponding to the selection statistical values ​​among the plurality of component points. Note that the selection statistical values ​​may be percentile values ​​calculated based on the elevation information of the plurality of component points.

[0013] The feature type estimation system of the present invention can also set two or more training areas with different shapes for one input reference point. In this case, two or more input data sets can be prepared for one training reference point. Furthermore, by generating two or more input areas with different shapes for one input reference point, it is also possible to input two or more input data sets for one input reference point.

[0014] The feature type estimation system of the present invention can set the training area and input area as circles. In this case, by setting two or more training areas with different radii for one training reference point, it is possible to prepare two or more input data sets for one training reference point. Furthermore, by generating two or more input areas with different radii for one input reference point, it is also possible to input two or more input data sets for one input reference point. [Effects of the Invention]

[0015] The feature type estimation system of the present invention has the following advantages. (1) There is no need to identify the type of feature by comparing photographs with existing floor plans, or by carrying a floor plan and conducting a field survey to identify the type of feature. In other words, since the type of feature is estimated without human judgment, human error is eliminated and results can be obtained in a short time. (2) By setting two or more types of training areas for one training reference point, it is possible to generate a feature estimation model trained on even more training data sets, thereby enabling more accurate estimation of feature types. [Brief explanation of the drawings]

[0016] [Figure 1] 1 is a block diagram showing the main configuration of an object type estimation system according to the present invention. [Figure 2] A model diagram that schematically shows a training dataset, training reference points, training areas, and training construction points. [Figure 3]A model diagram showing a learning reference point set on a road edge, a type of feature, and a circular learning area set based on that learning reference point. [Figure 4] 1 is a model diagram showing a circular learning area set for the learning reference point 1, a right semicircular learning area, and a left semicircular learning area. [Figure 5] A model diagram showing a schematic diagram of the large-circle learning area, the medium-circle learning area, and the small-circle learning area set for the learning reference point 1. [Figure 6] 1 is a model diagram showing a schematic diagram of an input dataset, input reference points, input area, and input configuration points. [Figure 7] FIG. 3 is a step diagram schematically showing a procedure for generating an input dataset by an input dataset generating means. [Figure 8] FIG. 2 is a flowchart showing an example of the main processing flow of the feature type estimation system. [Figure 9] A model diagram showing the elevation pattern of the road edge on the viaduct. [Figure 10] Model diagram showing the elevation pattern of road edges on an elevated bridge with nearby buildings. [Figure 11] A model diagram showing the elevation pattern of road edges (directly below road trees) with roadside trees nearby. [Figure 12] A model diagram showing the elevation pattern of road edges (gaps between road trees) with nearby road trees. [Figure 13] A model diagram showing the elevation pattern of a rough road edge. [Figure 14] A model diagram showing the distribution of patterned features. DETAILED DESCRIPTION OF THE INVENTION

[0017] An example of an embodiment of the feature type estimation system of the present invention will be described with reference to the drawings.

[0018] 1 is a block diagram showing the main components of a feature type estimation system 100 according to the present invention. As shown in the figure, the feature type estimation system 100 according to the present invention comprises learning means 101, feature type output means 102, and 3D topographic model storage means 105, and can also comprise control point input means 103, input dataset generation means 104, learning dataset storage means 106, and feature estimation model storage means 107.

[0019] The learning means 101, feature type output means 102, control point input means 103, and input dataset generation means 104 that constitute the feature type estimation system 100 of the present invention can be manufactured as dedicated units, or a general-purpose computer device can be used. This computer device includes a processor such as a CPU, memories such as ROM and RAM, and may further include input means such as a mouse and keyboard, and display means such as a display, and can be configured, for example, as a personal computer (PC) or server.

[0020] The 3D topographical model storage means 105, the learning dataset storage means 106, and the feature estimation model storage means 107 can be stored in a storage device of a general-purpose computer (e.g., a PC), or can be built on a database server. When built on a database server, they can be placed on a local network (LAN: Local Area Network), or can be a cloud server that stores data via the Internet.

[0021] Below, each of the main elements that make up the feature type estimation system 100 of the present invention in the first embodiment will be described in detail.

[0022] (Various storage means) The three-dimensional terrain model storage means 105 is a means for storing a three-dimensional terrain model (hereinafter referred to as a "3D terrain model"). Here, a 3D terrain model is a representation of terrain, including features, using three-dimensional coordinates (i.e., constituent points), and is a terrain model typified by a DSM (Digital Surface Model) or a DEM (Digital Elevation Model). A 3D terrain model is typically composed of small areas obtained by dividing a target planar range. These small areas are also called meshes, and are formed by dividing the range into, for example, orthogonal grids, and each small area has a representative point. Because a three-dimensional point cloud (i.e., constituent point cloud) obtained by measurement is often random data (data that is irregularly arranged on a plane), geometric calculations are often performed to assign heights to the representative points of the small areas. This calculation method includes the TIN (Triangulated Irregular Network) method, which finds height using an irregular triangular network formed from random data, the Nearest Neighbor method, which uses the nearest laser measurement point, as well as the Inverse Distance Weighting (IDW) method, the Kriging method, and the averaging method.

[0023] The training dataset storage means 106 is a means for storing a "training dataset," and the feature estimation model storage means 107 is a means for storing a "feature estimation model." This training dataset is so-called training data used by the training means 101 for machine learning, as will be described later, and includes at least the type of feature and elevation information for multiple constituent points belonging to (contained in) an area related to that feature (hereinafter referred to as the "training area"). Naturally, many training datasets are prepared for multiple feature types. Note that "belonging to the training area" here means that when a constituent point is planarly projected onto a region of interest, the constituent point is included within the training area. On the other hand, the feature estimation model is a model obtained as a result of machine learning performed by the training means 101, as will be described later, and outputs (returns) the type of feature when an "input dataset" is input to this feature estimation model. Here, the input dataset includes elevation information for multiple constituent points belonging to (contained in) a specified area (hereinafter referred to as the "input area"). The learning area is a predetermined area set based on a point related to a feature (hereinafter referred to as the "learning reference point"), and the input area is a predetermined area set based on a point designated by an operator (hereinafter referred to as the "input reference point"), for example, and both the learning area and the input area are two-dimensional areas with a predetermined size (dimensions) and shape. Note that height information refers to various concepts that represent the relative height difference from a certain reference height, but for convenience, the height information will be described here as "elevation."

[0024] (Learning tools) The learning means 101 is a means for generating a feature estimation model by performing machine learning (e.g., deep learning) using a training dataset as training data. The inventors discovered that the combinations (patterns) of elevations of multiple constituent points constituting a feature, or of a feature and multiple constituent points located around it, tend to differ depending on the type of feature. For example, even for the same road edge, a flat road edge, a road edge with an overpass nearby, or a road edge with roadside trees nearby will have different elevation patterns of constituent points, while the same flat road edge will have a certain tendency in its elevation pattern. Therefore, in the present invention, a feature estimation model is generated by having the learning means 101 perform machine learning on the training dataset (type of feature and elevations of constituent points), and the type of feature can be estimated by inputting an input dataset (elevations of constituent points included in an input area) into the feature estimation model.

[0025] As shown in Figure 2, the training dataset includes the type of feature and the elevation of the component point CP. However, as shown in this figure, the component points CP targeted by the training dataset are the component points CP included in the training area LRG (hereinafter, specifically referred to as "training component points LCP"). The training area LRG is set to a predetermined size and shape based on the training reference point LSP, which is a characteristic or representative point related to the feature. For example, in Figure 3, a point located on the road edge EL (feature) is set as the training reference point LSP, and the training area LRG is set to be a circle with a predetermined radius (a predetermined radius) centered on the training reference point LSP. Of course, the training area LRG is not limited to a circle; it can be set to a polygon, ellipse, or any other shape. Furthermore, the training reference point LSP can be specified by an operator (i.e., by manual input) or can be set automatically (mechanically) using image recognition technology or the like.

[0026] Furthermore, training datasets can also be generated by setting multiple types of training areas LRG for one training reference point LSP. For example, in FIG. 4, a "circular training area LRG11" is set with the training reference point LSP on the road edge EL at its center, and "right semicircular training areas LRG12" and "left semicircular training areas LRG13" are set by dividing this training area LRG1 by the road edge EL. In this case, three types of training datasets are generated for one training reference point LSP, and the feature type for all of them is "road edge." However, the training points LCP included in the circular training area LRG11, the training points LCP included in the right semicircular training area LRG12, and the training points LCP included in the left semicircular training area LRG13 are naturally all different. In this way, by setting two or more training areas LRG with different shapes for one training reference point LSP, two or more training datasets can be prepared for one training reference point LSP.

[0027] Alternatively, by changing the size (dimensions) of the training areas LRG, multiple types of training areas LRG can be set for one training reference point LSP, even though they have the same shape. For example, in FIG. 5, a circular training area LRG is set with the training reference point LSP on the road edge EL as its center, but by changing the radius, a "large-circle training area LRG21," a "medium-circle training area LRG22," and a "small-circle training area LRG23" are set. In this case, three types of training data sets are generated for one training reference point LSP, and the feature type for all of them is "road edge." However, the training points LCP included in the large-circle training area LRG21, the medium-circle training area LRG22, and the small-circle training area LRG23 are naturally all different. In this way, by setting two or more training areas with different dimensions for one training reference point LSP, two or more training data sets can be prepared for one training reference point.

[0028] Of course, two or more training regions LRG with different shapes can be set for one training reference point LSP, and two or more training regions with different dimensions can be set for one training reference point LSP, so that two or more training data sets can be prepared for one training reference point LSP. For example, in the cases of Figures 4 and 5, nine types (three shapes x three radii) are prepared for one training reference point LSP. In any case, the learning means 101 generates a feature estimation model by performing machine learning on the multiple training data sets prepared.

[0029] (Means for outputting feature types) The feature type output means 102 is a means for outputting the type of feature related to the input dataset based on the input dataset and the feature estimation model generated by the learning means 101. The input dataset includes the elevations of the configuration points CP as shown in FIG. 6. However, as shown in this figure, the configuration points CP targeted by the input dataset are configuration points CP included in the input region NRG (hereinafter, particularly referred to as "input configuration points NCP"). The input region NRG is set based on the input reference point NSP specified by the operator (or set in advance), and is set with the same region specifications as the size and shape (hereinafter, referred to as "region specifications") for setting the learning region LRG. For example, if the learning region LRG is set as a circle as shown in FIG. 3, the input region NRG is also set as a circle centered on the input reference point NSP and with the same diameter as the learning region LRG.

[0030] Fig. 7 is a step diagram showing a schematic procedure for generating an input dataset by the input dataset generating means 104. To generate an input dataset, first, an input reference point NSP is specified, as shown in Fig. 7(a). At this time, the operator can specify the desired input reference point NSP using a reference point input means 103 such as a pointing device (mouse, touch panel, pen tablet, touchpad, trackpad, trackball, etc.) or a keyboard while visually viewing the 3D terrain model displayed on a display means (such as a display), or the input reference point NSP can be set by the input dataset generating means 104 sequentially selecting a plurality of points set in advance.

[0031] When the input reference point NSP is specified, the input dataset generation means 104 sets an input region NRG based on the input reference point NSP, as shown in Figure 7(b), and further, as shown in Figure 7(c), the input dataset generation means 104 extracts input composition points NCP included in the input region NRG and uses these elevations as the input dataset.

[0032] Incidentally, it has been explained that, for the training data set, two or more training regions LRG (i.e., training data sets) with different shapes can be prepared for one training reference point LSP as shown in Figure 4, and two or more training regions LRG (i.e., training data sets) with different sizes can be prepared for one training reference point LSP as shown in Figure 5. In this way, when multiple types of training regions LRG are set for one training reference point LSP, it is possible to set input regions NRG corresponding to all types of training regions LRG (i.e., with the same region specifications as all types of training regions LRG), and then generate an input data set. For example, when the training dataset is set in the case of Figure 4, the input dataset generation means 104 sets a "circular input region NRG", a "right semicircular input region NRG", and a "left semicircular input region NRG" to correspond to the training region LRG, and then generates an input dataset; when the training dataset is set in the case of Figure 5, the input dataset generation means 104 sets a "large circular input region NRG", a "medium circular input region NRG", and a "small circular input region NRG" to correspond to the training region LRG, and then generates an input dataset; when the training dataset is set in the case of a combination of Figures 4 and 5, the input dataset generation means 104 sets nine types of input regions NRG to correspond to the training region LRG, and then generates an input dataset.

[0033] Alternatively, even when multiple types of training regions LRG are set for one training reference point LSP, it is also possible to set input regions NRG corresponding to several types of training regions LRG and then generate an input dataset. For example, when the training dataset is set in the case of Fig. 4, the input dataset generation means 104 sets only the "right semicircular input region NRG" and the "left semicircular input region NRG" and then generates an input dataset. When the training dataset is set in the case of Fig. 5, the input dataset generation means 104 sets only the "great circle input region NRG" and then generates an input dataset. When the training dataset is set in the case of a combination of Fig. 4 and Fig. 5, the input dataset generation means 104 sets only the "great circle input region NRG," the "right semicircular input region NRG obtained by dividing the great circle," and the "left semicircular input region NRG obtained by dividing the great circle," and then generates an input dataset.

[0034] When an input dataset is input, the feature type output means 102 outputs the type of feature related to the input dataset using a feature estimation model. Specifically, it compares the combination (pattern) of elevations of the input composition points NCP included in the input dataset with the learned elevation pattern, selects an approximate (or matching) elevation pattern, and outputs the type of feature associated with that elevation pattern.

[0035] (Elevation of the dataset) Up to this point, it has been explained that the training dataset contains the elevations of multiple training points LCP, and the input dataset contains the elevations of multiple input points NCP. Of course, it is also possible to include the elevations of all training points LCP in the training dataset, and the elevations of all input points NCP in the input dataset. Alternatively, it is also possible to extract training points LCP and input points NCP based on predetermined requirements and include them in each dataset. More specifically, training points LCP and input points NCP are extracted according to predetermined statistics (hereinafter referred to as "selection statistics"), and only the elevations of the extracted training points LCP are included in the training dataset, and only the elevations of the extracted input points NCP are included in the input dataset. For example, using percentile values ​​when arranging elevations from smallest (lowest), percentile values ​​in 5% increments, such as the 0th percentile (minimum elevation in this case), 5th percentile, 10th percentile, and 100th percentile, are set as selection statistics, and only the elevations of the training points LCPs and input points NCPs that correspond to these selection statistics (0th percentile, 5th percentile, and so on) are included in each dataset from among all the training points LCPs and input points NCPs. Note that it is recommended to set the selection statistics for extracting training points LCPs and the selection statistics for extracting input points NCPs to the same value. In this way, by standardizing the multiple elevations included in the training dataset and input dataset, it becomes easier to match the learned elevation patterns with the elevation patterns in the input dataset.

[0036] (Processing flow) The main processing of the feature type estimation system 100 will be described in detail below with reference to Fig. 8. Fig. 8 is a flow diagram showing an example of the flow of the main processing of the feature type estimation system 100, with the central column showing the processing to be performed, the left column showing what is necessary for that processing, and the right column showing what results from that processing.

[0037] As shown in Fig. 8, the processing by the feature type estimation system 100 is roughly divided into a "learning step" up to generating a feature estimation model, and an "output step" in which the type of feature related to the input dataset is actually estimated. In the learning step, a feature is first specified to create a learning dataset (Step 211 in Fig. 8), and characteristic (or representative) points related to the feature are set as learning reference points LSP (Step 212 in Fig. 8). At this time, the feature and learning reference points LSP can be specified by an operator, or the feature and learning reference points LSP can be set automatically (mechanically) using image recognition technology or the like.

[0038] Once the training reference points LSP are set, the training reference points LSP are used as a reference and generated in accordance with predetermined area specifications (Step 213 in FIG. 8). Next, training datasets are created based on the elevations of the training composition points LCP included in the training reference points LSP and the types of features associated with the training reference points LSP (Step 214 in FIG. 8). By repeating the series of processes up to this point (Steps 211 to 214 in FIG. 8), numerous training datasets are prepared for each of various types of features. Then, these training datasets are trained by the training means 101 to generate a feature estimation model (Step 215 in FIG. 8).

[0039] In the output step, first, input control points NSP are designated to generate an input data set (Step 221 in Fig. 8). At this time, the operator can designate the desired input control points NSP using the control point input means 103 while visually checking the 3D topographic model displayed on the display means, or the input control points NSP can be set by the input data set generation means 104 sequentially selecting a plurality of points set in advance.

[0040] Once the input reference point NSP is set, an input region of interest NRG is generated in accordance with predetermined region specifications using this input reference point NSP as a reference (Step 222 in FIG. 8). Next, an input dataset is generated based on the elevations of the learning composition points LCP included in this input region of interest NRG (Step 223 in FIG. 8). Then, the feature type output means 102 outputs the type of feature related to the input dataset based on the input input dataset and feature estimation model (Step 224 in FIG. 8).

[0041] (Verification results) As mentioned above, the inventors have discovered that the elevation patterns made up of the multiple constituent points CP that make up a feature (or multiple constituent points CP arranged around a feature) have a specific tendency for each type of feature. Figures 9 to 13 are model diagrams showing specific elevation patterns for each type of feature. Of these, Figure 9 shows an example of a feature being a "road edge on an viaduct," and shows 12 types of elevation patterns (graph on the right). In this example, percentile values ​​in 5% increments (0 percentile, 5 percentile, 10 percentile, ... 100 percentile) are set as statistical values ​​for selection, and therefore the elevation patterns are represented by percentile values ​​on the horizontal axis and elevations on the vertical axis. As in the example in Figure 4, areas are set in the shape of a circle, a right semicircle, and a left semicircle relative to reference point 1, and with four types of radii (size0.5, size1, size2, and size4 in the figure), with the order of circle, right semicircle, and left semicircle from top to bottom, and size0.5, size1, size2, and size4 from left to right, showing 12 types of elevation patterns. Also, Figure 10 is an example of a feature ``road edge on an elevated bridge with buildings nearby,'' Figure 11 is an example of a feature ``road edge with road trees nearby (but directly below the trees),'' Figure 12 is an example of a feature ``road edge with road trees nearby (but in gaps between the trees),'' and Figure 13 is an example of a feature ``road edge with large undulations,'' all of which show 12 types of elevation patterns in the same manner as Figure 9. As can be seen from Figures 9 to 13, the elevation patterns made up of the multiple component points CP that make up a feature (or a feature and the multiple component points CPs arranged around it) each have a specific tendency for each type of feature. Also, as can be seen by comparing Figures 11 and 12, by selecting an appropriate area size (in this case, by selecting size 4), the similarity of the elevation patterns is further improved, meaning that the trends for each type of feature are emphasized.

[0042] The elevation patterns shown in Figures 9 to 13 are a combination of 12 types of regions consisting of three types of shapes and four types of dimensions, and each represents elevations for 21 percentiles, so they can be thought of as 252-dimensional (3 x 4 x 21) data. This was converted into two dimensions using a conventional dimension reduction method and displayed in a graph as shown in Figure 14. From this graph, it can be seen that the patterns can be roughly divided into 6 to 10 groups, meaning that the elevation patterns show specific trends for each type of feature. [Industrial Applicability]

[0043] The feature type estimation system of the present invention can be used to estimate the types of various features, such as estimating different types of roads even for the same road, such as flat roads, roads near viaducts, roads along buildings, etc. Considering that the present invention provides spatial information about features in an ideal manner and ultimately leads to advanced maintenance and management of social infrastructure, it can be said that the invention is not only applicable to industry but is also expected to make a significant contribution to society. [Explanation of symbols]

[0044] 100 Feature type estimation system of the present invention 101 Learning method (for feature type estimation system) 102 (Feature type estimation system) feature type output means 103 (Feature Type Estimation System) Reference Point Input Method 104 (Feature type estimation system) input data set generation means 105 3D terrain model storage means (for feature type estimation system) 106 (Feature type estimation system) learning data set storage means 107 (Feature type estimation system) feature estimation model storage means CP constituent points EL Road Edge LCP Learning Composition Points LRG Learning Area LSP Learning Reference Points NCP input configuration points NRG Input Region of Interest NSP input reference point

Claims

1. A system for estimating a type of a feature based on a plurality of constituent points having three-dimensional coordinates, comprising: a three-dimensional terrain model storage means for storing a three-dimensional terrain model based on the plurality of constituent points; a learning means for generating a feature estimation model for outputting the type of a feature by machine learning a plurality of learning datasets; a feature type output means for outputting the type of feature related to the input data set based on the input input data set and the feature estimation model, the learning dataset includes a type of feature and height information of a plurality of the constituent points included in a learning area related to the feature, the height information corresponding to a predetermined selection statistical value of the constituent points; the input data set includes height information of the constituent points corresponding to the selection statistical value among the plurality of constituent points included in the input region of interest, the training area is a two-dimensional area having a predetermined size and shape, the area being based on a training reference point selected from a feature; the input region of interest is a region based on an input reference point designated by an operator, and is a two-dimensional region having the same size and shape as the learning region; the selection statistical value is a percentile value calculated based on height information of a plurality of component points; A feature type estimation system.

2. A system for estimating a type of a feature based on a plurality of constituent points having three-dimensional coordinates, comprising: a three-dimensional terrain model storage means for storing a three-dimensional terrain model based on the plurality of constituent points; a learning means for generating a feature estimation model for outputting the type of a feature by machine learning a plurality of learning datasets; a feature type output means for outputting the type of feature related to the input data set based on the input input data set and the feature estimation model, the training data set includes a type of feature and height information of a plurality of the constituent points included in a training area related to the feature; the input data set includes height information of the plurality of constituent points included in the input region of interest, the training area is a two-dimensional area having a predetermined size and shape, the area being based on a training reference point selected from a feature; the input region of interest is a region based on an input reference point designated by an operator, and is a two-dimensional region having the same size and shape as the learning region; two or more learning regions having different shapes are set for one learning reference point, thereby preparing two or more learning data sets for one learning reference point; two or more input regions of interest having different shapes are generated for one input reference point, whereby two or more input data sets are input for the one input reference point; A feature type estimation system.

3. A system for estimating a type of a feature based on a plurality of constituent points having three-dimensional coordinates, comprising: a three-dimensional terrain model storage means for storing a three-dimensional terrain model based on the plurality of constituent points; a learning means for generating a feature estimation model for outputting the type of a feature by machine learning a plurality of learning datasets; a feature type output means for outputting the type of feature related to the input data set based on the input input data set and the feature estimation model, the training data set includes a type of feature and height information of a plurality of the constituent points included in a training area related to the feature; the input data set includes height information of the plurality of constituent points included in the input region of interest, the training area is a two-dimensional area having a predetermined size and shape, the area being based on a training reference point selected from a feature; the input region of interest is a region based on an input reference point designated by an operator, and is a two-dimensional region having the same size and shape as the learning region; the learning area and the input area of ​​interest are circular; two or more training regions having different radii are set for one training reference point, thereby preparing two or more training data sets for the one training reference point; two or more input regions of interest having different radii are generated for one input reference point, whereby two or more input data sets are input for the one input reference point; A feature type estimation system.

Images