JP7813638B2 - Correction Area Extraction System - Google Patents

Description

This invention relates to technology for generating ground surfaces from measurement results obtained by a laser measuring instrument, and more specifically to a correction area extraction system that can automatically extract areas that require manual filtering after automatic filtering processing.

When measuring "ground" or "land features" (hereinafter collectively referred to as "ground, etc.") over a wide area, aerial photogrammetry has traditionally been the mainstream method. However, in recent years, various measurement methods have emerged, such as airborne laser measurement, measurement using satellite photos, and measurement using synthetic aperture radar, making it possible to select the most appropriate method depending on the situation. Note that "land features" here refers to "objects" excluding the ground.

Among these, airborne laser measurement is a technique in which an aircraft flies over the target area to be measured and measures the reflected waves of laser pulses emitted from the ground and other objects within the target area. Aircraft are typically equipped with a positioning device such as a Global Navigation Satellite System (GNSS) and an inertial measurement device such as an IMU (Inertial Measurement Unit), and the GNSS and IMU can be used to record the irradiation position (x, y, z) and irradiation attitude (ω, φ, κ) when the laser pulse is emitted.

When a laser pulse is emitted from an aircraft, the time of emission is recorded. Furthermore, when the laser pulse is reflected from the ground or other surface, it is received by a sensor mounted on the aircraft, and the time of reception is also recorded. Therefore, the distance to the measurement point (the point where the laser pulse is reflected) can be obtained from the time difference between the emission time and the reception time. Since the irradiation position (x, y, z) and irradiation attitude (ω, φ, κ) at the time of laser pulse emission are also recorded, three-dimensional coordinate data (hereinafter referred to as "measurement point data") of the laser pulse irradiation point (i.e., the measurement point) can be obtained. Furthermore, when the sensor mounted on the aircraft receives a laser pulse, it records the intensity of the reflected wave (hereinafter referred to as "reflection intensity"). This reflection intensity is, in effect, the magnitude of the energy of the received reflected wave (the amplitude of the laser pulse) and is measured directly as a voltage. The energy magnitude can be obtained by converting this voltage.

As explained so far, airborne laser measurement is a method of obtaining measurement point data by irradiating the ground, etc. with laser pulses from an aircraft in flight. These laser pulses are emitted 100,000 to 2,000,000 times per second, so a huge number of measurement point data are obtained in one measurement (flight). Furthermore, when measuring forests, etc., it is naturally impossible to irradiate the laser pulses only on the ground by weaving through the trees (in other words, to irradiate the laser pulses directly on the ground), so in addition to measurement point data reflected on the ground, measurement point data reflected on tree leaves and trunks is also obtained.

In many cases, airborne laser measurement is performed to determine the ground level within the target area, and measurement point data reflected by tree leaves and trunks (non-ground data) is essentially unnecessary data. For this reason, so-called filtering is carried out to remove non-ground data, and various improved filtering techniques have been proposed. For example, Patent Document 1 proposes a technique that compares a terrain model (so-called surface model) obtained through photometry with measurement point data obtained through airborne laser measurement, and removes non-ground data by determining that measurement point data that approximates the terrain model obtained through photometry has been reflected by tree leaves and trunks (equivalent to a surface).

JP 2011-158278 A

As mentioned above, filtering is performed to remove non-ground data in order to determine ground height. However, it is not practical for a human (operator) to manually remove the vast amount of measurement point data obtained from a single measurement. Therefore, "automatic filtering" is generally used to automatically remove non-ground data under certain conditions. Typical examples of automatic filtering include an automatic filtering process that creates a TIN (Triangulated Irregular Network) model from points extracted as initial ground points (the lowest points within a specific range) and detects (removes) non-ground data by searching for ground points near the surface within the TIN model, or an automatic filtering process that removes measurement point data where the reflection intensity (amplitude of the laser pulse) of the received laser pulse falls below a predetermined threshold as non-ground data. It is known that the reflection intensity of a laser pulse reflected by leaves, etc. is weaker than that reflected by the ground. Therefore, measurement point data related to laser pulses with low (weak) reflection intensity is unconditionally treated as non-ground data.

This automatic filtering process automatically extracts non-ground data, significantly reducing the amount of work required by the operator, but of course it is not possible to extract all non-ground data using automatic filtering. Therefore, the measurement point data (primary processed data) from which non-ground data has been removed by automatic filtering is usually subjected to further "manual filtering" by the operator. In other words, automatic filtering is performed on the measurement point data (original data) obtained by measurement to generate primary processed data, and manual filtering is then performed on this primary processed data to generate the final measurement point data (ground data).

Manual filtering is a process in which an operator visually examines the primary processed data to extract areas that need to be corrected, and then makes the corrections. This is an extremely nerve-wracking operation, and it is impossible to completely eliminate incorrect operations (so-called human error). Therefore, the inventors of the present invention decided to automatically extract candidates for areas that need to be corrected in the primary processed data, and further attempted to achieve automatic extraction using artificial intelligence (AI). By presenting the areas that need to be corrected, the operator only needs to target those areas, significantly reducing the work required for manual filtering.

However, when the areas to be corrected were extracted using a trained model generated through machine learning (e.g., deep learning), the extraction was not very accurate. The inventors then discovered that the cause of this was the difference (in other words, the difference in characteristics) between "excessive filtering" and "omissions in filtering" in automatic filtering processes. Here, "excessive filtering" refers to the process of mistakenly removing ground data, while "omissions in filtering" refers to the process of not removing (i.e., maintaining) non-ground data. In other words, there was a problem with treating non-ground data that had been "over-extracted" due to excessive filtering and data that had been "omissions in extraction" due to omissions in filtering equally.

The objective of the present invention is to solve the problems of the past, namely, to provide a correction area extraction system that can automatically extract candidates for areas to be corrected from primary processed data, taking into account the difference in characteristics between excessive filtering and omissions due to automatic filtering processing.

The present invention focuses on the fact that the process for extracting over-extracted data and the process for extracting under-extracted data are performed separately, i.e., a trained model for over-filtering and a trained model for under-filtering are generated separately, and is an invention based on a previously unconventional idea.

The correction area extraction system of the present invention is a system that automatically extracts correction areas to be subjected to manual filtering when generating ground data based on a terrain model (a model composed of multiple meshes, each mesh assigned an elevation), and is equipped with an over-extracted mesh detection means and an unextracted mesh detection means. The over-extracted mesh detection means assigns ground or non-ground labels to the removed meshes by inputting the removed meshes and elevations that were excluded by the automatic filtering of the terrain model into an over-extraction trained model. The unextracted mesh detection means assigns ground or non-ground labels to the retained meshes by inputting the retained meshes and elevations that were not excluded by the automatic filtering of the terrain model into an unextraction trained model. The over-extraction trained model is generated by machine learning using the "ground label or non-ground label (teacher data)" and "elevation (input data)" attached to the removed mesh as training data, while the unextraction trained model is generated by machine learning using the "ground label or non-ground label (teacher data)" and "elevation (input data)" attached to the maintained mesh as training data. The removed mesh to which a ground label has been assigned by the over-extraction mesh detection means and the maintained mesh to which a non-ground label has been assigned by the unextraction mesh detection means are defined as the correction area.

The correction area extraction system of the present invention may further include an elevation difference calculation means. This elevation difference calculation means calculates the elevation difference between the elevation of the removed mesh and the elevation of the ground data generated after the removed mesh is excluded. In this case, the over-extraction mesh detection means assigns a ground label or a non-ground label to the removed mesh by inputting the removed mesh and the elevation difference into an over-extraction trained model. In this case, the over-extraction trained model is generated by machine learning using the "ground label or non-ground label (teacher data)" and "elevation difference (input data)" assigned to the removed mesh as training data.

The correction area extraction system of the present invention can also automatically extract correction areas based on the grayscale (or color information) assigned to the mesh in addition to the elevation and elevation difference. In this case, the over-extracted mesh detection means assigns a ground label or non-ground label to the removed mesh by inputting the grayscale (or color information) and elevation (or elevation difference) associated with the removed mesh into the over-extraction trained model, and the unextracted mesh detection means assigns a ground label or non-ground label to the maintained mesh by inputting the grayscale (or color information) and elevation (or elevation difference) associated with the maintained mesh into the unextraction trained model. In this case, the over-extraction trained model is generated by machine learning using the "ground label or non-ground label (trainer data)" attached to the removed mesh and the "grayscale or color information, and elevation or elevation difference (input data)" related to the removed mesh as training data, while the under-extraction trained model is generated by machine learning using the "ground label or non-ground label (trainer data)" attached to the maintained mesh and the "grayscale or color information, and elevation or elevation difference (input data)" related to the maintained mesh as training data.

The correction area extraction system of the present invention can also automatically extract correction areas based on the grayscale (or color information) assigned to the mesh, without including it in the elevation or elevation difference. In this case, the over-extraction mesh detection means assigns a ground label or non-ground label to the removed mesh by inputting the grayscale (or color information) associated with the removed mesh into the over-extraction trained model, and the unextracted mesh detection means assigns a ground label or non-ground label to the maintained mesh by inputting the grayscale (or color information) associated with the maintained mesh into the unextraction trained model. In this case, the over-extraction trained model is generated by machine learning using the "ground label or non-ground label (trainer data)" assigned to the removed mesh and the "grayscale and color information removed (input data)" associated with the removed mesh as training data, and the unextraction trained model is generated by machine learning using the "ground label or non-ground label (trainer data)" assigned to the maintained mesh and the "grayscale and color information (input data)" associated with the maintained mesh as training data.

The correction region extraction system of the present invention can also include a candidate mesh detection means and a correction region setting means. The candidate mesh detection means is a means for selecting, as correction candidate meshes, removed meshes that have been assigned a ground label by the over-extracted mesh detection means and retained meshes that have been assigned a non-ground label by the unextracted mesh detection means. The correction region setting means is a means for setting a predetermined judgment region as a correction region when the proportion of candidate meshes in the judgment region exceeds a predetermined proportion threshold.

The correction region extraction system of the present invention has the following advantages.
(1) The process of extracting over-extracted data and the process of extracting under-extracted data are performed separately, so that the region to be subjected to manual filtering can be extracted with relatively high accuracy.
(2) Since the region to be subjected to the manual filtering process is automatically extracted, the operator's work required for the manual filtering process is significantly reduced.
(3) Furthermore, since the area to be subjected to manual filtering processing is automatically extracted, human errors that may occur during operator work can be easily avoided.

A model diagram showing a terrain model composed of multiple meshes. A cross-sectional view showing the "ground data" in schematic form. A model diagram explaining "excessive filtering" and "missed filtering" caused by automatic filtering processing. FIG. 1 is a block diagram showing the main configuration of a correction region extraction system according to the present invention. FIG. 2 is a flowchart showing an example of the main processing flow of the correction region extraction system 100 according to the first embodiment. FIG. 10 is a flowchart showing an example of the main processing flow of the correction area extraction system 100 according to the second embodiment. FIG. 11 is a flowchart showing an example of the main processing flow of the correction region extraction system 100 according to the third embodiment.

An example of an embodiment of the correction area extraction system of the present invention is explained below with reference to the accompanying drawings.

1. Definitions Before describing the embodiments of the present invention, the definitions of terms used herein are first provided.

(terrain model)
When measuring the ground, etc., using airborne laser measurement, a large number of measurement point data having three-dimensional coordinates (hereinafter referred to as a "3D point cloud") are obtained. Here, a "terrain model" is a model generated based on this three-dimensional point cloud, and examples of such a model include a digital surface model (DSM) and a digital elevation model (DEM). Typically, a terrain model is composed of small regions (hereinafter referred to as "meshes MS") obtained by dividing the planar range of the measurement target into multiple subregions (180 subregions in the figure), as shown in FIG. 1 . These meshes MS are divided into, for example, orthogonal grids and are also called pixels. Each mesh MS has a representative point assigned with an elevation. Because the three-dimensional point cloud obtained by airborne laser measurement is composed of random data (data irregularly arranged on a plane), geometric calculations are often used to assign elevations to the representative points of the subregions. Examples of calculation methods include a TIN method that finds height using an irregular triangulation network formed from random data, a nearest neighbor method that uses the nearest laser measurement point, inverse distance weighting (IDW), the Kriging method, and the averaging method.

(Ground data)
A terrain model generated based on airborne laser measurement typically represents the surface of the ground (e.g., a DSM). Here, the surface of an object refers to the top surface of the ground or object, such as green cover (e.g., farmland or forest), or artificial structures (e.g., buildings) constructed on the ground. In contrast, "ground data" is data representing a state without any features, as shown in Figure 2 , i.e., data obtained after removing features from the initial terrain model generated based on airborne laser measurement (hereinafter, specifically referred to as the "original terrain model"). The ground data shown in Figure 2 is a ground model (e.g., a DEM) obtained by removing non-ground data from the original terrain model using automatic filtering, and then performing interpolation on the resulting mesh.

(Color information and grayscale)
Originally, color is perceived visually by humans and is subject to individual differences. In recent years, color has been modeled for use on computers (electronic calculators). Various color modeling methods are known, including RGB, which uses red, green, and blue as its three basic colors; CMYK, which uses cyan, magenta, yellow, and black (key color), and the NCS and Ostwald color systems, which use yellow, red, blue, green, black, and white as their six basic colors. Color has three attributes: hue, saturation, and lightness. For example, RGB expresses various hues, saturations, and lightness by mixing the three primary colors of red, green, and blue (additive color mixing). Here, the values that specify colors, as defined by color modeling such as RGB, CMYK, NCS, and the Ostwald color system, are referred to as “color information.” For example, when RGB is used, the color information for pure red is (255,0,0), the color information for pure green is (0,255,0), and the color information for pure blue is (0,0,255).

"Grayscale" is a value that specifies the degree of darkness in a model that divides colors into multiple levels from white to black. For example, if there are 256 levels from white to black, white can be represented by a grayscale of 255, and black can be represented by a grayscale of 0.

As mentioned above, a terrain model is made up of multiple meshes MS, each of which is assigned an elevation. However, color information or a grayscale may also be set according to the elevation. In other words, in this case, each mesh MS (pixel) that makes up the terrain model can also be given color information or a grayscale in addition to the elevation. For convenience, a terrain model in which color information or a grayscale is assigned to the meshes MS will be referred to as a "colored terrain model."

(Removal mesh and maintenance mesh)
As mentioned above, when generating ground data based on three-dimensional coordinates, an automatic filtering process is performed on the original terrain model, and then a manual filtering process is performed. For convenience, the data obtained as a result of the automatic filtering process on the original terrain model will be referred to as the "intermediate terrain model." In other words, an "intermediate terrain model" is obtained by automatically filtering the "original terrain model," and "ground data" is obtained by manually filtering the "intermediate terrain model."

In the automatic filtering process, each mesh MS that makes up the original terrain model is classified as either a mesh MS that represents ground or a mesh MS that is not ground (non-ground data). Mesh MS determined to represent ground are left as they are, while mesh MS determined to represent non-ground data are removed. For convenience, mesh MS determined to represent ground by the automatic filtering process will be referred to as "maintained mesh MSk" and mesh MS determined to represent non-ground data will be referred to as "removed mesh MSe." For example, in Figure 1, of the 180 mesh MS, 159 mesh MS are classified as maintained mesh MSk, and the 21 mesh MS (shaded) are classified as removed mesh MSe. Note that in a typical automatic filtering process, a new elevation (elevation representing the original ground) is assigned to the removed mesh MSe. In other words, elevations are basically assigned to all mesh MS (maintained mesh MSk and removed mesh MSe) before ground data is generated.

(Pre-trained model)
As described above, automatic filtering can result in "excessive filtering," whereby ground is mistakenly removed as a removed mesh MSe, or "filtering omission," whereby non-ground is retained as a maintained mesh MSk. For example, in the case of Figure 3, non-ground data is removed from the original terrain model by automatic filtering, and interpolation processing is performed on the removed meshes (e.g., under the levee and under the building on the left). This results in "excessive filtering," whereby the "levee," which is actually ground, is removed but should be retained as a maintained mesh MSk. On the other hand, this results in "filtering omission," whereby the "building," which is a feature, is retained but should be removed as a removed mesh MSe. Therefore, the present invention utilizes machine learning techniques to detect "excessive filtering" and "filtering omission" caused by automatic filtering. In other words, a model generated by machine learning a large amount of training data (teacher data and input data) is used to detect erroneous determinations made in the automatic filtering process. For convenience, the model generated by machine learning will be referred to as a "trained model." In the present invention, various techniques including deep learning can be adopted as machine learning techniques for generating a trained model.

Another technical feature of the present invention is the use of a trained model that processes the removed mesh MSe (hereinafter, specifically referred to as the "over-extracted trained model") and a trained model that processes the maintained mesh MSk (hereinafter, specifically referred to as the "unextracted trained model"). In other words, "excessive filtering" caused by the automatic filtering process is detected by the over-extracted trained model, and "missed filtering" caused by the automatic filtering process is detected by the unextracted trained model. More specifically, the over-extracted trained model assigns a label indicating ground (hereinafter, simply referred to as the "ground label") or a label indicating non-ground (hereinafter, simply referred to as the "non-ground label") to the removed mesh MSe, while the unextracted trained model assigns a ground label or non-ground label to the maintained mesh MSk.

(Correction area and judgment area)
The present invention extracts the area to be subjected to manual filtering by detecting erroneous judgments made by the automatic filtering process. In other words, the target area is extracted by detecting meshes MS that actually represent the ground (hereinafter referred to as "ground meshes MS") from among the removed meshes MSe classified by the automatic filtering process, and detecting meshes MS that actually do not represent the ground (hereinafter referred to as "non-ground meshes MS") from among the maintained meshes MSk classified by the automatic filtering process. For convenience, the area to be subjected to manual filtering will be referred to as the "correction area."

As mentioned above, because ground mesh MS and non-ground mesh MS are detected for each mesh MS, correction areas can be extracted on a mesh MS basis. Alternatively, correction areas can be extracted for each area that groups together a certain number of mesh MS. For convenience, the area used to extract correction areas in this way, that is, a collection of mesh MS with a predetermined number and arrangement, will be referred to here as a "determination area." For example, in Figure 1, the determination area is set using mesh MS that are 4 horizontally and 5 vertically, resulting in a total of 9 determination areas.

2. Correction Area Extraction System Next, the correction area extraction system of the present invention will be described in detail with reference to the drawings. Figure 4 is a block diagram showing the main components of the correction area extraction system 100 of the present invention. As shown in this figure, the correction area extraction system 100 is configured to include an over-extracted mesh detection means 101 and an unextracted mesh detection means 102, and can also be configured to include an elevation difference calculation means 103, a candidate mesh detection means 104, a correction area setting means 105, an automatic filtering processing means 106, an over-extracted trained model generation means 107, an unextracted trained model generation means 108, a topographical model storage means 109, an over-extracted trained model storage means 110, an unextracted trained model storage means 111, etc.

The over-extracted mesh detection means 101, un-extracted mesh detection means 102, elevation difference calculation means 103, candidate mesh detection means 104, correction area setting means 105, automatic filtering processing means 106, over-extracted trained model generation means 107, and un-extracted trained model generation means 108 that make up the correction area extraction system 100 can be manufactured as dedicated units, or general-purpose computer devices can be used. In other words, the computer device executes calculations using a specified program to perform processing specific to each means. This computer device is equipped with a processor such as a CPU, memory such as ROM and RAM, and some also include input means such as a mouse and keyboard, and a display, and can be configured, for example, as a personal computer (PC) or server.

Furthermore, the terrain model storage means 109, over-extracted trained model storage means 110, and under-extracted trained model storage means 111 can use the storage device of a general-purpose computer (e.g., a personal computer), or can be built on a database server. If built on a database server, it can be placed on a local network (LAN: Local Area Network), or it can be a cloud server that stores data via the Internet.

Below, we will explain in detail each of the main elements that make up the correction area extraction system 100 of the present invention.

(Automatic filtering processing means)
The automatic filtering processing means 106 is a means for generating an "intermediate terrain model" by automatically filtering the "original terrain model" read from the terrain model storage means 109. For example, the automatic filtering processing means 106 may be a means for performing automatic filtering processing by having a computer device execute calculation processing using a predetermined program. Various conventional (commercially available) programs can be used as this program.

When the automatic filtering processing means 106 generates an intermediate terrain model, removed meshes MSe are extracted from the meshes MS that make up the original terrain model. As described above, the automatic filtering process assigns a new elevation (an elevation representing the original ground surface) to the removed mesh MSe. In other words, for the removed mesh MSe extracted by the automatic filtering processing means 106, the elevation (in the original terrain model) before filtering (hereinafter referred to as the "original elevation") and the elevation (in the intermediate terrain model) after filtering (hereinafter referred to as the "new elevation") are set.

(Elevation difference calculation means)
The elevation difference calculation means 103 is a means for calculating the difference between the original elevation and the new elevation of the removed mesh MSe (hereinafter referred to as the "elevation difference"). This elevation difference can be calculated by subtracting the new elevation from the original elevation, or by subtracting the original elevation from the new elevation, or by calculating the absolute value of the difference between the original elevation and the new elevation. The elevation difference calculation means 103 calculates the elevation difference for all removed meshes MSe extracted by the automatic filtering processing means 106.

(Over-extraction trained model generation means)
The over-extraction trained model generation means 107 is a means for generating an "over-extraction trained model" by machine learning a combination of a removed mesh MSe and its elevation. Specifically, the means selects a removed mesh MSe to which a "ground label" has been assigned by manual filtering (i.e., a removed mesh MSe when the judgment of the automatic filtering process is incorrect), and prepares "positive training data" consisting of the "ground label (teacher data)" and its "elevation (input data)." The means also selects a removed mesh MSe to which a "non-ground label" has been assigned by manual filtering (i.e., a removed mesh MSe when the judgment of the automatic filtering process is correct), and prepares "negative training data" consisting of the "non-ground label (teacher data)" and its "elevation (input data)." The means then performs machine learning on a large amount of training data to generate an "over-extraction trained model."

The over-extraction trained model generation means 107 can also be used to generate an over-extraction trained model using "elevation difference" instead of the elevation of the removed mesh MSe. In other words, the combination of the "ground label (training data)" and "elevation difference (input data)" attached to the removed mesh MSe is treated as "positive training data," and the combination of the "non-ground label (training data)" and "elevation difference (input data)" attached to the removed mesh MSe is treated as "negative training data," and then machine learning is performed.

The over-extraction trained model generation means 107 can also be a means for generating an over-extraction trained model using "grayscale" or "color information" in addition to the removed mesh MSe and elevation (or elevation difference). In other words, the combination of the "ground label (training data)" attached to the removed mesh MSe and the "elevation or elevation difference, and grayscale or color information (input data)" is considered to be "positive training data," and the combination of the "non-ground label (training data)" attached to the removed mesh MSe and the "elevation or elevation difference, and grayscale or color information (input data)" is considered to be "negative training data," and then machine learning is performed.

Furthermore, the over-extraction trained model generation means 107 can also be used to generate an over-extraction trained model using "grayscale" or "color information" instead of the elevation (or elevation difference) of the removed mesh MSe. In other words, the combination of the "ground label (trainer data)" attached to the removed mesh MSe and the "grayscale or color information (input data)" is used as "positive training data," and the combination of the "non-ground label (trainer data)" attached to the removed mesh MSe and the "grayscale or color information (input data)" is used as "negative training data," and then machine learning is performed. The over-extraction trained model generated by the over-extraction trained model generation means 107 is stored in the over-extraction trained model storage means 110 (Figure 4).

(Unextracted trained model generation means)
The unextracted trained model generation means 108 is a means for generating an "unextracted trained model" by machine learning a combination of a maintained mesh MSk and its elevation. Specifically, a maintained mesh MSk with a "non-ground label" attached (i.e., a maintained mesh MSk when the automatic filtering process makes an incorrect judgment) is selected, and "positive training data" consisting of the "non-ground label (teacher data)" and its "elevation (input data)" is prepared, and a maintained mesh MSk with a "ground label" attached (i.e., a maintained mesh MSk when the automatic filtering process makes a correct judgment) is selected, and "negative training data" consisting of the "non-ground label (teacher data)" and its "elevation (input data)" is prepared, and an "unextracted trained model" is generated by machine learning a large amount of training data.

The unextracted trained model generation means 108 can also be a means for generating an unextracted trained model using "color information" and "grayscale" in addition to the maintained mesh MSk and elevation. In other words, the combination of the "non-ground label (trainer data)" attached to the maintained mesh MSk and "elevation, and grayscale or color information (input data)" is considered to be "positive training data," and the combination of the "ground label (trainer data)" attached to the maintained mesh MSk and "elevation, and grayscale or color information (input data)" is considered to be "negative training data," and then machine learning is performed.

The unextracted trained model generation means 108 can also be used to generate an unextracted trained model using "color information" or "grayscale" instead of the elevation of the maintained mesh MSk. In other words, the combination of the "non-ground label (trainer data)" attached to the maintained mesh MSk and "grayscale or color information (input data)" is considered to be "positive training data," and the combination of the "ground label (trainer data)" attached to the maintained mesh MSk and "elevation, and grayscale or color information (input data)" is considered to be "negative training data," and then machine learning is performed. The unextracted trained model generated by the unextracted trained model generation means 108 is stored in the unextracted trained model storage means 111 (Figure 4).

(Means for detecting over-extracted meshes)
The over-extraction mesh detection means 101 is a means for assigning either a ground label or a non-ground label to the removed mesh MSe by inputting a combination of the removed mesh MSe extracted by the automatic filtering processing means 106, its elevation, elevation difference, and grayscale (color information) into an over-extraction trained model. Then, the over-extraction mesh detection means 101 extracts the removed mesh MSe to which the ground label has been assigned as a "ground mesh MS."

(Unextracted mesh detection means)
The unextracted mesh detection means 102 is a means for assigning either a non-ground label or a ground label to the maintained mesh MSk by inputting a combination of the maintained mesh MSk extracted by the automatic filtering processing means 106 and its elevation and grayscale (color information) into an unextracted trained model. The unextracted mesh detection means 102 then extracts the maintained mesh MSk to which the non-ground label has been assigned as a "non-ground mesh MS."

(Candidate mesh detection means)
The candidate mesh detection means 104 is a means for detecting meshes MS (hereinafter referred to as "candidate meshes") that make up the correction area. Specifically, it detects as candidate meshes the "ground mesh MS (removed mesh MSe with a ground label)" extracted by the over-extracted mesh detection means 101 and the "non-ground mesh MS (maintained mesh MSk with a non-ground label)" extracted by the unextracted mesh detection means 102.

(Correction area setting means)
The correction region setting means 105 is a means for setting a "correction region" based on the candidate meshes detected by the candidate mesh detection means 104. The correction region setting means 105 can be configured to set a correction region on a mesh MS basis, i.e., to set all candidate meshes as correction regions. Alternatively, the correction region can be configured to set a correction region for each "determination region." In this case, when the proportion of candidate meshes in the determination region exceeds (or is equal to or greater than) a predetermined threshold (hereinafter referred to as the "proportion threshold"), the determination region can be set as a correction region. For example, in FIG. 1, the determination region is composed of 20 mesh MS. If the proportion threshold is set to 25%, the correction region setting means 105 in this case would set a determination region containing more than five (or five or more) candidate meshes as a correction region.

3. Examples Next, examples of the correction area extraction system 100 of the present invention will be described in detail with reference to the drawings. The correction area extraction system 100 of the present invention can be broadly divided into an example in which a trained model is generated using the elevation associated with the removed mesh MSe (hereinafter referred to as a "first example"), an example in which a trained model is generated using the elevation difference associated with the removed mesh MSe (hereinafter referred to as a "second example"), and an example in which a trained model is generated using color information (or grayscale) associated with the removed mesh MSe and the maintained mesh MSk (hereinafter referred to as a "third example"). Each example will be described in turn below.

(First Example)
FIG. 5 is a flow diagram showing an example of the main processing flow of the correction area extraction system 100 in the first embodiment, in which the processing to be performed is shown in the center column, what is necessary for that processing is shown in the left column, and what results from that processing is shown in the right column.

In the first embodiment, the over-extraction trained model generation means 107 generates an "over-extraction trained model" by machine learning the combination of the removed mesh MSe and its elevation. Specifically, "positive training data" consisting of the "ground label (training data)" attached to the removed mesh MSe and its "elevation (input data)" is prepared, and "negative training data" consisting of the "non-ground label (training data)" attached to the removed mesh MSe and its "elevation (input data)" is prepared, and an over-extraction trained model is generated by machine learning a large amount of training data.

In this case, the unextracted trained model generation means 108 generates an "unextracted trained model" by machine learning the combination of the maintained mesh MSk and its elevation. Specifically, it prepares "positive training data" consisting of the "non-ground label (teacher data)" attached to the maintained mesh MSk and its "elevation (input data)", and prepares "negative training data" consisting of the "ground label (teacher data)" attached to the maintained mesh MSk to which the "ground label" is attached and its "elevation (input data)", and then generates an unextracted trained model by machine learning a large amount of training data.

Once the over-extraction trained model and the under-extraction trained model have been generated, an automatic filtering process is first performed, as shown in Figure 5 (Step 211 in Figure 5). Specifically, an "intermediate terrain model" is generated by automatically filtering the "original terrain model" read from the terrain model storage means 109 using the automatic filtering processing means 106. At this time, removed meshes MSe are extracted from the meshes MS that make up the original terrain model, and new elevations (elevations representing the original ground) are assigned to the removed meshes MSe. Meshes MS that are not selected as removed meshes MSe are then selected as maintained meshes MSk. In other words, the meshes MS that make up the original terrain model are classified into removed meshes MSe and maintained meshes MSk by the automatic filtering process.

Once the automatic filtering processing means 106 has classified the removed mesh MSe and the maintained mesh MSk, the over-extracted mesh detection means 101 assigns a ground label or a non-ground label to the removed mesh MSe (Step 212 in Figure 5). Specifically, the combination of the removed mesh MSe and the elevation extracted by the automatic filtering processing means 106 is input into the over-extraction trained model, thereby assigning either a ground label or a non-ground label to the removed mesh MSe. The removed mesh MSe to which the ground label has been assigned is then extracted as a "ground mesh MS."

Next, the unextracted mesh detection means 102 assigns a ground label or a non-ground label to the maintained mesh MSk (Step 213 in Figure 5). Specifically, the combination of the maintained mesh MSk and elevation extracted by the automatic filtering processing means 106 is input into the unextracted trained model, thereby assigning either a non-ground label or a ground label to the maintained mesh MSk. The maintained mesh MSk to which the non-ground label has been assigned is then extracted as a "non-ground mesh MS."

When the over-extracted mesh detection means 101 extracts a ground mesh MS and the unextracted mesh detection means 102 extracts a non-ground mesh MS, the candidate mesh detection means 104 detects these ground mesh MS and non-ground mesh MS as "candidate meshes" (Step 214 in Figure 5). The correction area setting means 105 then sets a "correction area" based on the candidate meshes detected by the candidate mesh detection means 104 (Step 215 in Figure 5). For example, a determination area in which the proportion of candidate meshes exceeds a proportion threshold is extracted and set as the correction area. Once the correction area is set, the operator generates "ground data" by performing manual filtering on the correction area (Step 216 in Figure 5).

In the first embodiment, it is also possible to use an over-extracted trained model or an under-extracted trained model generated using "grayscale" and "color information" in addition to elevation. In this case, the over-extracted trained model generation means 107 defines the combination of the "ground label (trainer data)" attached to the removed mesh MSe and its "elevation, grayscale, and color information (input data)" as "positive training data," and the combination of the "non-ground label (trainer data)" attached to the removed mesh MSe and its "elevation, grayscale, and color information (input data)" as "negative training data," and then performs machine learning to generate an over-extracted trained model. In addition, the unextracted trained model generation means 108 defines the combination of the "non-ground label (teaching data)" attached to the maintenance mesh MSk and its "elevation, and grayscale or color information (input data)" as "positive training data," and the combination of the "ground label (teaching data)" attached to the maintenance mesh MSk and its "elevation, and grayscale or color information (input data)" as "negative training data," and then performs machine learning to generate an unextracted trained model.

Then, by inputting the combination of the removed mesh MSe extracted by the automatic filtering processing means 106, elevation, and grayscale (color information) into an over-extraction trained model, the removed mesh MSe is assigned either a ground label or a non-ground label, and the removed mesh MSe assigned the ground label is extracted as a "ground mesh MS." Furthermore, by inputting the combination of the maintained mesh MSk extracted by the automatic filtering processing means 106, elevation, and color information (grayscale) into an unextraction trained model, the maintained mesh MSk is assigned either a non-ground label or a ground label, and the maintained mesh MSk assigned the non-ground label is extracted as a "non-ground mesh MS."

(Second Example)
FIG. 6 is a flow diagram showing an example of the main processing flow of the correction area extraction system 100 in the second embodiment, in which the processing to be performed is shown in the center column, what is necessary for that processing is shown in the left column, and what results from that processing is shown in the right column.

In the second embodiment, the over-extraction trained model generation means 107 generates an "over-extraction trained model" by machine learning the combination of the removed mesh MSe and its elevation difference. Specifically, "positive training data" consisting of the "non-board label (trainer data)" attached to the removed mesh MSe and its "elevation difference (input data)" is prepared, and "negative training data" consisting of the "board label (trainer data)" attached to the removed mesh MSe and its "elevation difference (input data)" is prepared, and an over-extraction trained model is generated by machine learning a large amount of training data. In addition, in this case, the unextraction trained model generation means 108 generates an "unextraction trained model" by machine learning the combination of the maintained mesh MSk and its elevation, as in the first embodiment.

Once the over-extraction trained model and the un-extraction trained model have been generated, as shown in FIG. 6, automatic filtering processing is first performed, as in the first embodiment (Step 221 in FIG. 6). Once the automatic filtering processing means 106 has classified the removed meshes MSe and the maintained meshes MSk, the elevation difference calculation means 103 calculates the "elevation difference" for the removed mesh MSe (Step 222 in FIG. 6). Then, the over-extraction mesh detection means 101 assigns a ground label or a non-ground label to the removed mesh MSe (Step 223 in FIG. 6). Specifically, the combination of the removed mesh MSe and the elevation difference extracted by the automatic filtering processing means 106 is input into the over-extraction trained model, thereby assigning either a ground label or a non-ground label to the removed mesh MSe. Then, the removed mesh MSe to which the ground label has been assigned is extracted as a "ground mesh MS."

Next, as in the first embodiment, the unextracted mesh detection means 102 assigns a ground label or a non-ground label to the maintained mesh MSk (Step 224 in Figure 6). Once the over-extracted mesh detection means 101 extracts a ground mesh MS and the unextracted mesh detection means 102 extracts a non-ground mesh MS, the candidate mesh detection means 104 detects the ground mesh MS and non-ground mesh MS as "candidate meshes" (Step 225 in Figure 6), and the correction area setting means 105 sets a "correction area" based on the candidate meshes detected by the candidate mesh detection means 104 (Step 226 in Figure 6). Once the correction area is set, the operator generates "ground data" by performing manual filtering on the correction area (Step 227 in Figure 6).

As with the first embodiment, the second embodiment can also use over-extracted trained models and under-extracted trained models generated using "grayscale" and "color information" in addition to elevation and elevation difference. In this case, the over-extracted trained model generation means 107 defines the combination of the "ground label (trainer data)" attached to the removed mesh MSe and the "elevation difference, and grayscale and color information (input data)" as "positive training data," and the combination of the "non-ground label (trainer data)" attached to the removed mesh MSe and the "elevation difference, and grayscale and color information (input data)" as "negative training data," and then performs machine learning to generate an over-extracted trained model. In addition, the unextracted trained model generation means 108 defines the combination of the "non-ground label (teaching data)" attached to the maintenance mesh MSk and the "elevation difference, and grayscale or color information (input data)" as "positive training data," and the combination of the "ground label (teaching data)" attached to the maintenance mesh MSk and the "elevation difference, and grayscale or color information (input data)" as "negative training data," and then performs machine learning to generate an unextracted trained model.

Then, by inputting the combination of the removed mesh MSe extracted by the automatic filtering processing means 106, the elevation difference, and the grayscale (color information) into an over-extraction trained model, the removed mesh MSe is assigned either a ground label or a non-ground label, and the removed mesh MSe assigned the ground label is extracted as a "ground mesh MS." Furthermore, by inputting the combination of the maintained mesh MSk extracted by the automatic filtering processing means 106, the elevation, and the color information (grayscale) into an unextraction trained model, the maintained mesh MSk is assigned either a non-ground label or a ground label, and the maintained mesh MSk assigned the non-ground label is extracted as a "non-ground mesh MS."

(Third Example)
FIG. 7 is a flow chart showing an example of the main processing flow of the correction area extraction system 100 in the third embodiment, in which the central column shows the processing to be performed, the left column shows what is necessary for that processing, and the right column shows what results from that processing.

In the third embodiment, the over-extraction trained model generation means 107 generates an "over-extraction trained model" by machine learning the combination of the removed mesh MSe and color information or grayscale. Specifically, "positive training data" consisting of "ground labels (training data)" and "elevation-based grayscale and color information (input data)" attached to the removed mesh MSe is prepared, and "negative training data" consisting of "non-ground labels (training data)" and "elevation-based grayscale and color information (input data)" attached to the removed mesh MSe is prepared. Then, an over-extraction trained model is generated by machine learning a large amount of training data.

In this case, the unextracted trained model generation means 108 generates an "unextracted trained model" by machine learning the combination of the maintained mesh MSk and color information or grayscale. Specifically, the combination of the "non-ground label (trainer data)" attached to the maintained mesh MSk and the "color information or grayscale based on elevation (input training data)" is considered to be "positive training data," and the combination of the "ground label (trainer data)" attached to the maintained mesh MSk and the "color information or grayscale based on elevation (input training data)" is considered to be "negative training data," and an unextracted trained model is generated by machine learning a large amount of training data.

Once the over-extraction trained model and the un-extraction trained model have been generated, as in the first and second embodiments, as shown in FIG. 7, automatic filtering processing is first performed (Step 231 in FIG. 7). Once the automatic filtering processing means 106 has classified the removed meshes MSe and the maintained meshes MSk, the over-extraction mesh detection means 101 assigns a ground label or a non-ground label to the removed mesh MSe (Step 232 in FIG. 7). Specifically, the combination of the removed mesh MSe extracted by the automatic filtering processing means 106 and grayscale (color information) is input into the over-extraction trained model, thereby assigning either a ground label or a non-ground label to the removed mesh MSe. The removed mesh MSe to which the ground label has been assigned is then extracted as a "ground mesh MS."

Next, the unextracted mesh detection means 102 assigns a ground label or a non-ground label to the maintained mesh MSk (Step 233 in Figure 7). Specifically, the combination of the maintained mesh MSk extracted by the automatic filtering processing means 106 and color information (grayscale) is input into the unextracted trained model, thereby assigning either a non-ground label or a ground label to the maintained mesh MSk. The maintained mesh MSk to which the non-ground label has been assigned is then extracted as a "non-ground mesh MS."

When a ground mesh MS is extracted by the over-extracted mesh detection means 101 and a non-ground mesh MS is extracted by the unextracted mesh detection means 102, the candidate mesh detection means 104 detects the ground mesh MS and non-ground mesh MS as "candidate meshes" (Step 234 in Figure 7), and the correction area setting means 105 sets a "correction area" based on the candidate meshes detected by the candidate mesh detection means 104 (Step 235 in Figure 7). Once the correction area is set, the operator generates "ground data" by performing manual filtering on the correction area (Step 236 in Figure 7).

The modified area extraction system of the present invention can be used to obtain ground elevations in a variety of locations, including mountainous areas, coastal areas, and urban areas, and is particularly suitable for use in locations with forested areas. Because the present invention can obtain ground elevations with high accuracy, it can be effectively used in social infrastructure planning and disaster prevention planning, making it an invention that can be used industrially and is also expected to make a significant contribution to society.

100 Correction area extraction system of the present invention 101 Over-extracted mesh detection means (of correction area extraction system) 102 Un-extracted mesh detection means (of correction area extraction system) 103 Elevation difference calculation means (of correction area extraction system) 104 Candidate mesh detection means (of correction area extraction system) 105 Correction area setting means (of correction area extraction system) 106 Automatic filtering processing means (of correction area extraction system) 107 Over-extracted trained model generation means (of correction area extraction system) 108 Un-extracted trained model generation means (of correction area extraction system) 109 Terrain model storage means (of correction area extraction system) 110 Over-extracted trained model storage means (of correction area extraction system) 111 Un-extracted trained model storage means (of correction area extraction system) MS Mesh MSe Removed mesh MSk Maintained mesh

Claims (5)

A system for automatically extracting a correction area to be subjected to manual filtering processing when generating ground data based on a terrain model, comprising:
The terrain model generated based on the airborne laser measurement is composed of a plurality of meshes, and each of the meshes is assigned an elevation;
an over-extracted mesh detection means for inputting the removed meshes and their elevations, which are removed by automatically filtering the terrain model, into an over-extracted trained model, and thereby assigning a ground label or a non-ground label to the removed meshes;
an unextracted mesh detection means for inputting the maintained meshes and elevations that have not been excluded by automatically filtering the terrain model into an unextracted trained model, thereby assigning the ground label or the non-ground label to the maintained meshes;
The over-extraction trained model is generated by machine learning the removed mesh and elevation to which the ground label or the non-ground label is attached,
The unextracted trained model is generated by machine learning the maintained mesh and elevation to which the ground label or the non-ground label is attached,
The removed meshes to which the ground labels have been assigned by the over-extracted mesh detection means and the maintained meshes to which the non-ground labels have been assigned by the unextracted mesh detection means are defined as the correction region.
A correction region extraction system comprising: A system for automatically extracting a correction area to be subjected to manual filtering processing when generating ground data based on a terrain model, comprising:
The terrain model generated based on the airborne laser measurement is composed of a plurality of meshes, and each of the meshes is assigned an elevation;
an elevation difference calculation means for calculating an elevation difference between the elevation of a removed mesh that has been removed by automatically filtering the terrain model and the elevation of ground data generated after the removed mesh has been removed;
an over-extracted mesh detection means for assigning a ground label or a non-ground label to the removed mesh by inputting the removed mesh and the elevation difference into an over-extracted trained model;
an unextracted mesh detection means for inputting the maintained meshes and elevations that have not been excluded by automatically filtering the terrain model into an unextracted trained model, thereby assigning the ground label or the non-ground label to the maintained meshes;
the over-extraction trained model is generated by machine learning the removed meshes labeled with the ground labels or the non-ground labels and the elevation differences;
The unextracted trained model is generated by machine learning the maintained mesh and elevation to which the ground label or the non-ground label is attached,
The removed meshes to which the ground labels have been assigned by the over-extracted mesh detection means and the maintained meshes to which the non-ground labels have been assigned by the unextracted mesh detection means are defined as the correction region.
A correction region extraction system comprising: The meshes constituting the terrain model are assigned grayscale or color information that is set according to altitude ,
the over-extraction mesh detection means assigns the ground label or the non -ground label to the removed mesh by inputting the grayscale or the color information related to the removed mesh into the over-extraction trained model;
The unextracted mesh detection means assigns the ground label or the non-ground label to the maintenance mesh by inputting the grayscale or the color information related to the maintenance mesh into the unextracted trained model;
the over-extraction trained model is generated by machine learning the grayscale or the color information related to the removed meshes labeled with the ground label or the non-ground label,
The unextracted trained model is generated by machine learning the grayscale or color information related to the maintenance mesh to which the ground label or the non-ground label is attached.
3. The system for extracting a correction area according to claim 1. A system for automatically extracting a correction area to be subjected to manual filtering processing when generating ground data based on a terrain model, comprising:
the terrain model generated based on the airborne laser measurement is composed of a plurality of meshes, and each mesh is assigned grayscale or color information set according to altitude ;
an over -extracted mesh detection means for inputting the removed meshes that have been removed by automatic filtering of the terrain model and the grayscale or color information into an over-extraction trained model, thereby assigning a ground label or a non-ground label to the removed meshes;
an unextracted mesh detection means for inputting the maintained meshes that have not been removed by automatically filtering the terrain model and the grayscale or color information into an unextracted trained model, thereby assigning the ground label or the non-ground label to the maintained meshes;
the over-extraction trained model is generated by machine learning the removed mesh to which the ground label or the non-ground label is attached and the grayscale or the color information;
The unextracted trained model is generated by machine learning the maintained mesh to which the ground label or the non-ground label is attached and the grayscale or the color information,
The removed meshes to which the ground labels have been assigned by the over-extracted mesh detection means and the maintained meshes to which the non-ground labels have been assigned by the unextracted mesh detection means are defined as the correction region.
A correction region extraction system comprising: a candidate mesh detection means for determining, as correction candidate meshes, the removed meshes to which the ground labels have been assigned by the over-extracted mesh detection means and the maintained meshes to which the non-ground labels have been assigned by the unextracted mesh detection means;
and a correction region setting means for setting a predetermined judgment region as the correction region when the proportion of the correction candidate mesh in the predetermined judgment region exceeds a predetermined proportion threshold.
5. The system for extracting a modified region according to claim 1, claim 2, or claim 4.