KR102621091B1 - A Method of Estimating a River Cross-Sectional Area using a Point Cloud obtained from a UAV - Google Patents

Abstract

The present invention relates to a method for estimating river cross-section using a 3D point cloud obtained from an unmanned aerial vehicle (UAV). More specifically, 3D points scanned by a server from a photogrammetry camera provided on an unmanned aerial vehicle (UAV) By the point cloud acquisition step in which a set point cloud is acquired and the server, the point cloud is added to the K-Nearest Neighbor Algorithm (KNN)-based Local Linear Regression model (KLR) This relates to a river cross-section estimation method using a 3D point cloud obtained from an unmanned aerial vehicle (UAV), which includes a river cross-section estimation step in which the river cross-section is estimated using .

Description

{A Method of Estimating a River Cross-Sectional Area using a Point Cloud obtained from a UAV}

The present invention relates to a method for estimating river cross-section using a point cloud obtained from an unmanned aerial vehicle (UAV), and more specifically, to a local linear regression model (KLR) based on the K-nearest neighbor algorithm and a photo provided on an unmanned aerial vehicle (UAV). This relates to a river cross-section estimation method using a 3D point cloud obtained from an unmanned aerial vehicle (UAV), in which the river cross-section is estimated using a point cloud obtained from a survey camera.

Previously, for river channel surveying, a surveyor boarded a boat and measured using a single beam. Additionally, the survey method for river floodplains should also be conducted using Network-RTK while moving on foot at 20m intervals, excluding physically inaccessible areas such as wetlands and steep riverbanks. Therefore, the commonly used survey method is to acquire river bed elevation through RTK-GPS, limited to areas accessible to humans.

Topographic data on areas around rivers are essential for establishing comprehensive plans for various rivers such as national rivers, local rivers, and small rivers, as well as the environment of target rivers, restrictions on indiscriminate development activities, and disaster management such as floods.

However, the previously mentioned method has limitations in time and cost to observe all wide river areas, and cannot model the continuously changing shape of rivers in three dimensions. In particular, there are many topographical features such as vegetation and trees in irregular rivers. When distributed, there are technical limitations that make it difficult to reflect the exact elevation value of the ground.

Related document 1 is about a river surveying system using an unmanned aerial vehicle (UAV), which receives and analyzes image information and GPS coordinate information from an unmanned aerial vehicle (UAV) and compares them with the coordinates on a digital map to identify discrepant areas. , Since the shape of the ground within the river cannot be analyzed, the measurement inaccuracy problem caused by conventional direct river channel surveys cannot be solved.

KR 10-2193108

The present invention is intended to solve the above problems by using a local linear regression model (KLR) based on the K-nearest neighbor algorithm and an unmanned aerial vehicle (UAV) to easily check the cross section of a river that can only be known through conventional direct surveying. The purpose is to obtain a river cross-section estimation method using a 3D point cloud acquired from an unmanned aerial vehicle (UAV) by using a point cloud obtained from an equipped photogrammetry camera to estimate the river cross-section.

In order to achieve the above purpose, the river cross-section estimation method using a point cloud obtained from an unmanned aerial vehicle (UAV) of the present invention is a set of three-dimensional points scanned by a server from a photogrammetry camera provided on the unmanned aerial vehicle (UAV). A point cloud acquisition step in which an in-point cloud is acquired; and a river cross-section estimation step in which the river cross-section is estimated by the server by using the point cloud in a local linear regression model (KLR) based on the K-nearest neighbor algorithm.

As described above, according to the present invention, the point cloud obtained from a photogrammetry camera equipped on an unmanned aerial vehicle (UAV) is used to estimate the river cross section in a local linear regression model (KLR) based on the K-nearest neighbor algorithm, The cross section of a river, which can only be known through direct surveying, can be easily confirmed. In addition, the present invention can estimate a river cross section that is substantially the same as the actual river cross section compared to the existing approximate method of measuring and connecting several points.

Figure 1 is a diagram showing an image of a river captured by an unmanned aerial vehicle (UAV) according to an embodiment of the present invention.
Figure 2 is a diagram showing the KLR for a 3D point twice the standard by the number of neighbors according to an embodiment of the present invention.
Figure 3 is a diagram showing the KLR for a 3D point 10 times the standard by the number of neighbors according to an embodiment of the present invention.
Figure 4 shows the root mean square deviation (A) for each multiple (a) of the number of neighbors of the KLR for a 3-dimensional point twice the reference and the neighbors of the KLR for a 3-dimensional point 10 times the reference according to an embodiment of the present invention. This is a diagram showing the root mean square deviation (B) for each number multiple (a).
Figure 5 shows the root mean square deviation (A) for each multiple of the number of neighbors (a) of KLR for a three-dimensional point M times the reference and the optimal number of neighbors for each multiple of the number of points (M) according to an embodiment of the present invention. This is a drawing showing multiples (B).

The terms used in this specification are general terms that are currently widely used as much as possible while considering the function in the present invention, but this may vary depending on the intention or precedent of a person skilled in the art, the emergence of new technology, etc. In addition, in certain cases, there are terms arbitrarily selected by the applicant, and in this case, the meaning will be described in detail in the description of the relevant invention. Therefore, the terms used in the present invention should be defined based on the meaning of the term and the overall content of the present invention, rather than simply the name of the term.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as generally understood by a person of ordinary skill in the technical field to which the present invention pertains. Terms defined in commonly used dictionaries should be interpreted as having a meaning consistent with the meaning in the context of the related technology, and unless clearly defined in the present application, should not be interpreted in an ideal or excessively formal sense. No.

Hereinafter, embodiments according to the present invention will be described in detail with reference to the accompanying drawings. However, the described embodiments of the present invention may be modified into various other forms, and the scope of the present invention is not limited to the embodiments described below. The shapes and sizes of elements shown in the drawings may be exaggerated for clearer explanation.

First, the river cross-section estimation method using a 3D point cloud obtained from an unmanned air vehicle (UAV) of the present invention uses a photogrammetry camera ( A point cloud acquisition step (S100) in which a point cloud, which is a set of 3D points scanned from 110), is obtained, and the server 200 performs local linear regression based on the K-Nearest Neighbor Algorithm. It includes a river section estimation step (S200) in which the point cloud is used in a (Local Linear Regression) model (KLR) to estimate the river section.

The unmanned aerial vehicle (UAV) 100 is an airplane or helicopter-shaped aircraft that can be flown and controlled by the guidance of radio waves without a pilot, and is most preferably a drone. And the unmanned aerial vehicle (UAV) 100 includes a photogrammetry camera 110 capable of photogrammetry through 3D scanning.

Generally, when light is emitted for surveying, Lidar is used, which measures the distance from the time the light hits the object to be measured, is reflected, and returns. However, there are limitations in easily implementing it because it is an expensive equipment. . Additionally, in terms of performance, there is a difficulty in identifying densely packed objects such as a dense forest of trees.

Therefore, in the point cloud acquisition step (S100), a point cloud composed of numerous 3D points can be acquired from an image or photo taken from the photogrammetry camera 110 through software in the server 100. Therefore, it does not require separate expensive equipment and is more suitable for locally measuring the cross section of a river by clearly identifying dense grass and rocks clustered around the river.

Next, conventional linear regression is a regression analysis technique that models the linear correlation between a dependent variable and one or more independent variables, and when it is based on two or more independent variables, it is called multiple linear regression. Then, a linear prediction function is used to model the regression equation, and the unknown parameters in the regression equation can be estimated from a plurality of data.

In this regard, the river cross-section estimation step (S200) uses a Local Linear Regression (KLR) model based on the K-Nearest Neighbor Algorithm. The KLR model sets the difference between the local data cluster and the target data as the distance, selects k data, and applies only the selected k data to the Weighted Linear Regression model.

More specifically, the river cross-section estimation step (S200) includes a distance (D _j ) calculation step (S210), an elevation index storage step (S220), a fitting step (S230), and a dependent variable (S230) so that the river cross-section can be estimated. y _t ) may include an estimation step (S240).

First, in the distance (D _j ) calculation step ₍ S210), the distance (D j) between the observation point (x _j ) and the prediction point (x _t ) for n 3D points selected in the point cloud is calculated. It can be. It is most desirable to use the Euclidean distance measurement method to determine the distance (D _j ) between the observation point (x _j ) and the prediction point (x _t ).

Next, in the altitude index storage step (S220), the prediction point (x _t ) is used as a standard to determine the altitude index for the k distances (D _j ) with the lowest value among the distances (D _j ). can be stored respectively.

Here, the k distances (D _j ) with the lowest value are the k nearest distances (D _j ) among the n distances (D _j ) calculated from the distance (D j ) calculation step (S210 ₎ . . Here, k is the number of neighbors (K) in the K-Nearest Neighbor Algorithm.

Next, the fitting step (S230) is performed by local linear regression using the elevation index for the k distances (D _j ) as a dataset ([x(p), y(p)]). ) model (KLR) can be fitted. In the above dataset, x(p) is the distance (D _j ), and y(p) is the elevation index. Here, j is from 1 to n, p is from 1 to k, n is a positive integer, and k is a positive rational number.

Here, in the fitting step (S230), a weight may be considered in the distance (D _j ) in order to improve fitting accuracy, and a parameter reflecting the weight may be estimated. First, the fitting step (S230) may include a weight matrix generation step (S231) in which a weight matrix is generated using Equation 1 below.

Here, in the weight matrix generation step (S231), a matrix with each element (δ) on the main diagonal can be derived through the diag function. The matrix created like this This is the weight matrix.

Next, in the fitting step (S230), a design matrix is generated using the k distances (D _j ) in the dataset ([x(p), y(p)]) as shown in Equation 2 below. A design matrix generation step (S232) may be further included.

here, is a design matrix, x _t is a prediction point, x _(k) is the kth closest point to the prediction point (x _t ) in the dataset, and k is the number of neighbors (K), which is a positive rational number.

For example, if the number of neighbors (K) is 3, the design matrix ( ), the second row is the first distance (D ₁ ) _, which is the difference between the prediction point (x _t ) and the first closest point, and the prediction point (x _t ) and the prediction point (x _a second distance (D _{2 )} , which is _the difference between _the predicted point ( ) may include.

Next, the fitting step (S230) is the weight matrix ( ) and design matrix ( ) may be substituted into [Equation 3] below to include the parameter estimation step (S233) in which the parameters of the Local Linear Regression model (KLR) are estimated.

here, is a parameter, is the transpose matrix of the design matrix, is the weight matrix generated using [Equation 1] above. In general, a transposed matrix refers to a matrix whose rows and columns have been switched from an existing matrix.

Next, in the dependent variable (y _t ) estimation step (S240), the dependent variable (y _t ) of the Local Linear Regression model (KLR) can be estimated. The dependent variable (y _t ) estimation step (S240) is a parameter estimated from the parameter estimation step (S233) ( ) can be derived by reflecting [Equation 4] below. And the dependent variable (y _t ) can be estimated using [Equation 4].

Here, y _t is the dependent variable, is a parameter.

Therefore, in the river cross-section estimation step (S200), the shape of the graph created with the prediction point (x _t ) and the dependent variable (y _t ) estimated from the dependent variable (y _t ) estimation step (S240) is changed to the cross section of the river. can be estimated.

Next, the K-Nearest Neighbor Algorithm is one of the unsupervised learning techniques in machine learning, enabling classification and prediction of new observation data. First, it is possible to classify new observation data by finding the k pieces of previous data that are closest to the previous data among various labels and determining the label for the new observation data. And rather than presenting a specific type of model from prior data like a linear model, new data can be predicted based only on adjacent data among the prior learning data.

However, in the conventional K-Nearest Neighbor Algorithm, the number of k neighbors (K) must be arbitrarily set by the operator, so the number of possible neighbors (K) is determined until the most suitable number of neighbors (K) is confirmed. There is the inconvenience of having to substitute them all.

Additionally, depending on the number of neighbors (K), the shape of the graph created with the predicted point (x _t ) and the dependent variable (y _t ) may differ from the actual river cross section. For example, as the number of neighbors (K) becomes smaller, overfitting occurs and the river cross section can be estimated more precisely than expected. Conversely, as the number of neighbors (K) increases, underfitting occurs and the actual river cross section can be estimated. There may be a large error and distortion may occur.

Therefore, so that an appropriate number of neighbors (K) can be selected, the river cross-section estimation step (S200) of the present invention uses a heuristic analysis method to calculate the K-Nearest Neighbor Algorithm using the following [Equation 5] ), the number of k neighbors (K) can be selected.

Here, k is the number of neighbors (K), a is a multiple of the number of neighbors, and n is the number of 3D points selected in the point cloud. k, a are positive rational numbers, and n is a positive integer.

Most preferably, k is characterized in that it is within the range of 1.5 to 2.5.

Meanwhile, in the river cross-section estimation step (S200) of the present invention, the number of three-dimensional points selected in the point cloud is set to the above so that the river cross-section can be estimated to be substantially the same as the shape of the actual river cross-section with minimal errors. It is characterized by being 3 to 12 times the number initially obtained from the server 100. That is, a multiple (M) of the number of 3D points may be 3 to 12.

Therefore, according to the present invention, a point cloud can be obtained from an image or photo taken with a camera installed on an unmanned aerial vehicle (UAV) without a separate expensive device. And in the case of a point cloud, by predicting only adjacent points among numerous points, there is a remarkable effect that a river cross section that is substantially the same as the actual river cross section can be estimated without being influenced by non-adjacent points.

(Example 1)

Experimental conditions and parameters

Figure 1 is a diagram showing an image of a river captured by an unmanned aerial vehicle (UAV) according to an embodiment of the present invention. One embodiment of the present invention targets Rice Stream located in Jinju, Republic of Korea. A video of Rice Valley was filmed using a drone with the model name DJI Phantom 4. The captured image was input to the server of the present invention, and a point cloud was obtained through 3D scanning software installed on the server. The software in question is WebODM, an open source tool for creating map point clouds, terrain, and 3D surface models from aerial imagery. In one embodiment of the present invention, the point cloud was acquired at 0.1m intervals and a total of 161 points were selected.

The ground control point (GCP) of the actual river cross section is a point on the ground whose coordinates are already known through GPS surveying, and 10 specific points were acquired around the Rice River. Generally, the cross section cut in the transverse direction of the river is trapezoidal in shape to facilitate construction and secure maximum outflow. Therefore, the actual river cross-section (measured) was assumed to be 4m on both upper sides, the distance from the upper left side to the upper right side was 6m, and the height from the bottom of the river to the top and the hypotenuse were 6m each.

Figure 2 is a diagram showing the KLR for a 3D point twice the standard by the number of neighbors according to an embodiment of the present invention. Figure 3 is a diagram showing the KLR for a 3D point 10 times the standard by the number of neighbors according to an embodiment of the present invention.

In one embodiment of the present invention, the multiples (M) of the number of points are different. In Figure 2, since 161 points were acquired twice, 322 dots, which is twice the number, are marked with a red circle. Figure 3 is obtained 10 times based on 161 points, so 1,610 points, which is 10 times the number, are indicated with a red circle. It has been done. And in Figures 2 and 3, the actual river cross section (measured) is displayed as a solid line in blue. And the river cross section estimated to be KLR according to an embodiment of the present invention is indicated by a green solid line. Here, in one embodiment of the present invention, the number of neighbors (K) is 1 times (a = 1), 2 times (a = 2), 3 times (a = 3), or 4 times (a = 4) with respect to the square root of n. ) was applied. Here, a is a multiple of the neighboring number.

Experiment result

Looking at the KLR (green solid line) for 322 points in Figure 2, when the number of neighbors (K) is 1 times the square root of n (a = 1), the top and bottom surfaces on both sides are quite irregular, and the number of neighbors (K) is 1 times the square root of n (a = 1). When the square root of n is 4 times (a=4), the top and bottom surfaces of both sides are significantly curved. On the other hand, 2 times (a = 2) and 3 times (a = 3) the square root of n was estimated to be a flat surface quite similar to the actual river cross section (measured) at the top and bottom of both sides.

Looking at the KLR (solid green line) for 1,610 3D points in Figure 3, it was estimated to be a flat surface much more similar to the actual river cross section (Measured) at all neighborhood numbers (K) than the KLR (solid green line) in Figure 2. Therefore, it has been proven that the larger the number of points, that is, the larger the multiple of the number of points (M), the more likely it is to estimate the river cross section similar to the actual one.

Next, looking at the KLR (green solid line) according to the number of neighbors (K) in Figure 3, the number of neighbors (K) is 1 times the square root of n (a = 1), same as in Figure 2, and the number of neighbors (K) is 1 times the square root of n (a = 1) This is quite irregular, and the number of neighbors (K) is 4 times the square root of n (a=4), so the top and bottom surfaces on both sides are estimated to be quite curved. On the other hand, 2 times (a = 2) and 3 times (a = 3) the square root of n was estimated to be a flat surface quite similar to the actual river cross section (measured) at the top and bottom of both sides.

Therefore, it was proven that in order to estimate the cross section of a trapezoidal river similar to the actual one, the number of neighbors (K) must be selected according to the appropriate multiple of the number of neighbors (a), and the number of points must be sufficiently large.

(Example 2)

Experimental conditions and parameters

In one embodiment of the present invention, we aim to prove the number of neighbors (K) most suitable for river cross section estimation. Experimental conditions and parameters are the same as Example 1. Additionally, when the neighboring number (K) was calculated as a times the square root of n, the range of neighboring number multiples (a) was set to 0.5 to 5 at 0.5 intervals. And in each condition, the root mean square error (RMSE) was calculated between KLR and the actual river cross section (Measured) according to an embodiment of the present invention, and the root mean square error (RMSE) for the number of neighbors (K) ) was displayed in a graph.

Figure 4 shows the root mean square deviation (A) for each multiple (a) of the number of neighbors of the KLR for a 3-dimensional point twice the reference and the neighbors of the KLR for a 3-dimensional point 10 times the reference according to an embodiment of the present invention. This is a diagram showing the root mean square deviation (B) for each number multiple (a). Figure 5 shows the root mean square deviation (A) for each multiple of the number of neighbors (a) of KLR for a three-dimensional point M times the reference and the optimal number of neighbors for each multiple of the number of points (M) according to an embodiment of the present invention. This is a drawing showing multiples (B).

Experiment result

In Figure 4 (A), the root mean square deviation for each number of neighbors of KLR for a three-dimensional point twice the standard according to an embodiment of the present invention shows the smallest deviation when the number of neighbors multiple (a) is around 1.5. Here, the standard for the number of points is 161, the same as in Example 1, and the deviation is output for 322 points, which is twice the multiple (M) of the number of points.

In addition, in Figure 4 (B), the root mean square deviation for each neighbor multiple (a) of KLR for a three-dimensional point 10 times the standard according to an embodiment of the present invention is around 2.5. It shows the least deviation. Here, the number of dots is 1,610.

Figure 5 (A) shows that the larger the multiple (M) of the number of points ranging from 1 to 12, that is, the larger the number of points, the actual river cross section (Measured) and the KLR estimated according to an embodiment of the present invention. It can be seen that the deviation is small. Figure 5 (B) shows the most suitable neighboring number multiple (a) for each multiple (M) of the number of points.

Therefore, in order to be able to estimate the actual river cross section most closely according to an example of the present invention, the multiple (M) of the number of points may be within the range of 3 to 12, and the multiple (a) of the neighboring number may be within the range of 1.5 to 2.5. You can.

As described above, although the embodiments have been described with limited examples and drawings, various modifications and variations can be made by those skilled in the art from the above description. For example, the described techniques are performed in an order different from the described method, and/or the components of the described system, structure, device, circuit, etc. are combined or combined in a different form than the described method, or in a different configuration. Appropriate results may be achieved through substitution or substitution by elements or equivalents.

Therefore, other implementations, other embodiments, and equivalents to the claims also fall within the scope of the claims described below.

100.. Unmanned Aerial Vehicle (UAV)
110.. Photogrammetry camera
200.. Server

Claims (5)

A point cloud acquisition step in which a point cloud, which is a set of three-dimensional points scanned from a photogrammetry camera installed on an unmanned aerial vehicle (UAV), is acquired by the server; and
A river cross section estimation step in which the river cross section is estimated by the server using the point cloud in a Local Linear Regression (KLR) model based on the K-Nearest Neighbor Algorithm. do,
The river cross section estimation step is,
A distance (D j ) calculation step in which the distance (D _j ) between the observed _point (x _j ) and the predicted point (x _t ) for n 3D points selected in the point cloud is calculated;
An altitude index storage step in which the prediction point (x _t ) is used as a standard and the elevation index for the k distances (D _j ) with the minimum value are respectively stored;
A fitting step in which a Local Linear Regression model (KLR) is fitted with the elevation index for the k distances (D _j ) to the dataset ([x(p), y(p)]). ; and
A three-dimensional point obtained from an unmanned aerial vehicle (UAV), comprising: a dependent variable (y _t ) estimation step in which the dependent variable (y _t ) of the Local Linear Regression model (KLR) is estimated; River section estimation method using cloud.
Here, j is from 1 to n, p is from 1 to K, n is a positive integer, and k is a positive rational number. According to claim 1,
The fitting step is,
A weight matrix generation step in which a weight matrix is generated using Equation 1 below;
[Equation 1]

A design matrix generation step in which a design matrix is generated using the k distances (D _j ) in the dataset ([x(p), y(p)]) using the following [Equation 2]; and
[Equation 2]

here, is a design matrix, xt is a prediction point, x(k) is the kth closest point to the prediction point (xt) in the dataset, and k is the number of neighbors (K), which is a positive rational number.
Obtained from an unmanned aerial vehicle (UAV) including a parameter estimation step in which the weight matrix and the design matrix are substituted into [Equation 3] below to estimate the parameters of the Local Linear Regression model (KLR). A river cross section estimation method using a 3D point cloud.
[Equation 3]

here, is a parameter, is the transpose matrix of the design matrix, is the weight matrix generated using [Equation 1] above. According to clause 2,
The dependent variable (yt) estimation step is,
A river cross-section estimation method using a 3D point cloud obtained from an unmanned aerial vehicle (UAV), wherein the dependent variable (yt) is estimated by substituting the parameters in [Equation 4] below.
[Equation 4]

Here, y _t is the dependent variable, is a parameter. According to clause 1,
The river cross section estimation step is,
Using a heuristic analysis method, the number of k neighbors (K) is selected in the K-Nearest Neighbor Algorithm using the following [Equation 5]. River section estimation method using 3D point cloud.
[Equation 5]

Here, k is the number of neighbors (K), a is a multiple of the number of neighbors, and n is the number of 3D points selected in the point cloud. k, a are positive rational numbers, and n is a positive integer. delete