Tree ranging method based on unmanned aerial vehicle point cloud semantic segmentation for power transmission corridor forest fire prevention and control

Abstract

The invention provides an unmanned aerial vehicle point cloud semantic segmentation tree distance measurement method for forest fire prevention and control of a power transmission corridor, and the method comprises the steps: obtaining the three-dimensional point cloud data of a power transmission line, trees and related facilities through the scanning of an unmanned aerial vehicle, and carrying out the preprocessing of the initial data; analyzing the preprocessed data by using a semantic segmentation model based on deep learning, improving semantic segmentation precision by using a deep learning algorithm so as to automatically identify point clouds of the power line and the tree, and extracting three-dimensional coordinate information of the point clouds; modeling the three-dimensional point cloud of the power line to obtain a complete power line model; measuring and calculating the distance between the power line and the tree through a segmented section method, and comparing the distance with a preset safety threshold value; the method can guarantee the safety of the power line through an automatic and high-precision distance measurement method.

Description

Technical Field

[0001] The present invention relates to the technical field of power facility monitoring, and in particular to a tree ranging method for unmanned aerial vehicle (UAV) point cloud semantic segmentation for wildfire prevention and control in power transmission corridors. Background Art

[0002] In mountainous areas, if trees grow too close to power lines, it is very easy to cause line failures in dry weather, especially during the peak season of wildfires, and even cause the wildfires to spread. Therefore, during power inspections, it is necessary to focus on and inspect overhead power lines.

[0003] Traditional power inspection methods mostly rely on manual on-site inspections, which often require inspectors to use their naked eyes or ordinary instruments to measure the distance between trees and power lines. This is not only inefficient and labor-intensive, but also limited by factors such as terrain and field of view, making it difficult to meet the needs of large-scale transmission line inspections.

[0004] With the development of UAV technology and 3D mapping technology, the image data of power corridors are acquired by manned aircraft or UAV equipped with multi-spectral airborne equipment, and the data is processed to realize inspection. However, many of these data acquisition devices are greatly affected by environmental factors, such as weather and light, which may lead to unstable data quality. An automated and high-precision distance measurement method is needed to ensure the safety of power lines. The present invention proposes a solution to this problem. Summary of the invention

[0005] The present invention proposes a tree ranging method based on semantic segmentation of drone point cloud for forest fire prevention and control in power transmission corridors, which can ensure the safety of power lines with an automated and high-precision ranging method.

[0006] The present invention adopts the following technical solutions.

[0007] A tree ranging method using semantic segmentation of drone point clouds for forest fire prevention and control in power transmission corridors, the method firstly acquires three-dimensional point cloud data of power transmission lines, trees and related facilities through drone scanning, and then pre-processes the initial data; applies a semantic segmentation model based on deep learning to analyze the pre-processed data, uses a deep learning algorithm to improve the accuracy of semantic segmentation to automatically identify the point clouds of power lines and trees, and extracts their three-dimensional coordinate information; then models the three-dimensional point clouds of power lines to acquire a complete power line model; measures and calculates the distance between power lines and trees through a segmented profile method, and compares it with a preset safety threshold.

[0008] The method comprises the following steps: Step S1: Use a drone equipped with a laser radar system to scan the mountainous forest area to obtain three-dimensional point cloud data of power transmission lines, trees and related facilities; Step S2: pre-process the collected point cloud data, extract target point cloud data, and remove ground points and irrelevant noise; Step S3: classify the target point cloud data based on the semantic segmentation model of deep learning, identify the point cloud data of wires and trees, and extract their three-dimensional coordinate information; Step S4, modeling the three-dimensional point cloud of the power lines to obtain a complete power line model; Step S5, measuring the distance between the power line and the tree by segmented profile analysis method; Step S6: Compare the measured distance with a preset safety threshold to determine whether it is lower than the safety threshold.

[0009] In step S1, when a drone equipped with a laser radar system is used to scan a mountain forest area, the distance between the drone and the power transmission line is kept within a safe range, and the laser radar's scanning angle of the power transmission line is mainly a side view angle.

[0010] When the drone scans the power transmission corridor in the mountainous area, if the terrain hinders the flight path, the drone adjusts the flight path so that the lidar's scanning angle of the transmission line can be combined with the bird's-eye view to collect point cloud data, and a segmented scanning strategy is used to collect point cloud data in complex terrain.

[0011] During the data collection process, the horizontal distance between the UAV equipped with the LiDAR system and the power transmission line is maintained between 5 meters and 20 meters to ensure that the UAV is not subject to electromagnetic interference from the power transmission line and can collect high-quality point cloud data at the same time. During the data collection process, the specific flight altitude of the drone is dynamically adjusted according to different terrains; when the drone scans the power lines, trees and related facilities from different sides at a side-view angle, the angle range is set at 10 to 35 degrees, and the specific angle is adjusted according to the direction of the power lines and terrain changes to cover the lines and the trees and terrain around them; When the drone needs to scan the power transmission lines and related environment from a bird’s-eye view, the angle is set between 45 and 90 degrees to enhance data integrity; When encountering complex terrain, the UAV adopts a segmented scanning strategy, dividing the scanning area into multiple segments according to terrain changes. Each segment is controlled within the length range of 50 to 100 meters to plan a flight path that suits it, and adjusts the UAV's flight altitude and scanning angle in real time according to the terrain.

[0012] In step S2, the preprocessing operation includes distinguishing power line segments, ground point filtering, and point cloud downsampling. The specific steps include: According to the geometric distribution characteristics of the point cloud and the location of the tower, the power line point cloud is divided into different sections between two adjacent towers; the tower point cloud is assumed to be , the center point of the tower is taken as the tower position ; Divide the power line point cloud into multiple sections according to the distance between adjacent towers, and determine whether the power line point cloud is within the range of two towers: in For the power line point cloud, is a specific point cloud in the segment, Represents the segment point cloud, and They are the center points of two adjacent towers respectively; In the initial point cloud, the CSF algorithm is used to filter the ground points and separate the ground points from the non-ground points. The main steps are as follows: the point cloud is inverted and the ground points are made to be the top of the point cloud through coordinate transformation; the simulated cloth is placed on top of the point cloud and divided into node grids. The number of nodes is determined by the resolution. Decision: Project the point cloud and cloth nodes onto the horizontal plane, based on the vertical distance between the node and the point cloud. Limit the range of node movement; calculate the gravity of each node and tension Action: Update node displacement and repeat iterations until the node height change meets the threshold. or the maximum number of iterations is reached ; Calculate the distance between the point cloud and the cloth model based on the node position , the distance is less than the set threshold Points are classified as ground points, and the rest are non-ground points; The voxel downsampling method is used for point cloud data to reduce the amount of data to improve computational efficiency. The point cloud is divided into voxel grids , the side length of each voxel is , the center point of the point cloud in each voxel grid is used as the representative point of the voxel , output the downsampled point cloud set .

[0013] In step S3, semantic segmentation is performed on the point cloud data based on the PointNet++ model, and the semantic segmentation model is trained on a large number of annotated power line and tree point cloud data sets. The specific steps include: first, grouping and sampling the input point cloud layer by layer to determine the local neighborhood of each representative point; then, in each local neighborhood, A shared multi-layer perceptron is used to extract local features of relative positions. The calculation formula is: , in It represents feature concatenation, and MLP is a multi-layer perceptron; is the coordinate of the current point cloud, are the coordinates of the center point of the local neighborhood where the point cloud is located, Represents the feature vector of the current point cloud; And feature aggregation is performed through symmetric functions: in, Yes Local features of Represents the maximum pooling operation; Then, multi-scale features from local to global are extracted recursively in a hierarchical manner. The feature transfer formula of each layer is: in It is Characteristics of the layer; It is grouping, which means grouping the surrounding points based on the sampled point cloud to form a new local area. It is sampling, which means downsampling the point cloud, reducing the number of points, and constructing a multi-level feature representation; Finally, the features are input into the fully connected network, the classification or segmentation results are calculated, and the network parameters are optimized using cross entropy loss. At the same time, the regularization term is combined to improve the generalization ability of the model. The loss function calculation formula for the semantic segmentation task is: in is the one-hot encoding of the real label, is the predicted probability output by the model, Represents the cross entropy loss, which is used to calculate the difference between the category distribution predicted by the model and the true distribution. represents the regularization term, which is used to prevent the model from overfitting. is the weight of the regularization term.

[0014] In step S4, the three-dimensional point cloud of the power line is modeled, and the method includes the following steps: first, a straight line model is used Calculate the slope , get the angle between the straight line model and the X axis ;in, represents the slope of the straight line, represents the intercept of the straight line on the Y axis, Indicates the angle between the straight line model and the X-axis; Then according to Construct a rotation matrix ,pass The power line point cloud can be rotated around the Z axis and roughly parallel to the X axis, so that the power line point cloud is more consistent with the parabola characteristics on the vertical projection plane; Selecting the XOZ axis as the vertical projection plane of the electric line, the parabolic model can be obtained. ; Assume the number of point clouds of a single power line is , for the parabolic model, the error equation calculation formula can be obtained as: ; The three parameters of the parabola model can be solved by taking the first-order partial derivative of the equation .

[0015] In step S5, the segmented profile analysis method includes selecting multiple profile positions along the power line path, and the specific steps include: firstly extracting the connecting lines of adjacent points in the power line model, and the calculation formula of the connecting lines is: This means that the vectors between adjacent points , that is, the line segment connecting two points; Connect the wire As the axis of the section, the plane normal vector perpendicular to the axis is It can be obtained by normalizing the coordinates of the axis , Section plane It is expressed as: in is any point on the plane, It is the starting point of the power line.

[0016] In step S5, the shortest distance between the tree and the power line in the section plane is calculated, and the measured distance is compared with a preset safety threshold. The specific steps include: assuming that the tree point cloud , which is projected onto the section plane , the projection formula is: in, is the normal vector of the plane, is a point in the tree point cloud, It is the starting point of the power line; The calculation formula for the shortest distance from the tree point cloud to the power line is: in is the closest point on the electric line segment within the section.

[0017] In step S6, the calculated shortest distance With the preset safety threshold For comparison, if , an alarm or notification will be issued to remind relevant personnel to take timely measures to prevent the occurrence of wildfires.

[0018] Compared with the prior art, the present invention has the following advantages: 1. Use a deep learning-based semantic segmentation model to automatically identify trees and power lines in point clouds, extract their 3D coordinate information, and achieve efficient and automated distance measurement; 2. The semantic segmentation model based on PointNet++ is trained on a large-scale annotated data set. Compared with the traditional segmentation method based on feature thresholds, it can adapt to various point cloud densities and complex environments and improve the recognition accuracy of power line and tree point clouds; 3. The least squares method is used to fit the power line point cloud, and combined with the segmented profile modeling method, a continuous and complete three-dimensional power line model can be generated, providing an accurate data basis for subsequent distance measurement; 4. The use of drones and efficient deep learning algorithms, combined with fast point cloud preprocessing and ranging calculation methods, can achieve real-time monitoring of large-scale transmission lines and provide timely response for wildfire prevention. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments: Attached Figure 1 This is a schematic diagram of the overall workflow of a method for automatic tree distance measurement based on semantic segmentation of drone point clouds for forest fire prevention in power transmission corridors according to the present invention. Attached Figure 2 This is a schematic diagram of the working principle of the laser radar device carried by the drone in the automatic tree ranging method of the drone point cloud semantic segmentation for the prevention and control of forest fires in power transmission corridors of the present invention. Attached Figure 3 This is a schematic diagram of the point cloud data preprocessing process in a method for automatic tree ranging based on semantic segmentation of drone point cloud for forest fire prevention and control in power transmission corridors according to the present invention. Attached Figure 4 This is a schematic diagram of point cloud data annotation results in a tree automatic ranging method for UAV point cloud semantic segmentation for power transmission corridor forest fire prevention and control in the present invention. DETAILED DESCRIPTION

[0020] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments. The present invention can be implemented or applied through other different specific embodiments, and the details in this specification can also be modified and changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. In addition, the drawings of the present invention are only simple schematic illustrations and are not depicted according to actual dimensions. It is stated in advance. The following implementation methods will further describe the relevant technical content of the present invention in detail, but the disclosed content is not intended to limit the scope of protection of the present invention.

[0021] As shown in the figure, a tree ranging method based on semantic segmentation of drone point cloud for forest fire prevention and control in power transmission corridors is described. The method first obtains three-dimensional point cloud data of power transmission lines, trees and related facilities through drone scanning, and then preprocesses the initial data; applies a semantic segmentation model based on deep learning to analyze the preprocessed data, and uses a deep learning algorithm to improve the accuracy of semantic segmentation to automatically identify the point clouds of power lines and trees, and extract their three-dimensional coordinate information; then models the three-dimensional point cloud of the power lines to obtain a complete power line model; and measures and calculates the distance between the power lines and the trees through a segmented profile method, and compares it with a preset safety threshold.

[0022] The method comprises the following steps: Step S1: Use a drone equipped with a laser radar system to scan the mountainous forest area to obtain three-dimensional point cloud data of power transmission lines, trees and related facilities; Step S2: pre-process the collected point cloud data, extract target point cloud data, and remove ground points and irrelevant noise; Step S3: classify the target point cloud data based on the semantic segmentation model of deep learning, identify the point cloud data of wires and trees, and extract their three-dimensional coordinate information; Step S4, modeling the three-dimensional point cloud of the power lines to obtain a complete power line model; Step S5, measuring the distance between the power line and the tree by segmented profile analysis method; Step S6: Compare the measured distance with a preset safety threshold to determine whether it is lower than the safety threshold.

[0023] In step S1, when a drone equipped with a laser radar system is used to scan a mountain forest area, the distance between the drone and the power transmission line is kept within a safe range, and the laser radar's scanning angle of the power transmission line is mainly a side view angle.

[0024] When the drone scans the power transmission corridor in the mountainous area, if the terrain hinders the flight path, the drone adjusts the flight path so that the lidar's scanning angle of the transmission line can be combined with the bird's-eye view to collect point cloud data, and a segmented scanning strategy is used to collect point cloud data in complex terrain.

[0025] During the data collection process, the horizontal distance between the UAV equipped with the LiDAR system and the power transmission line is maintained between 5 meters and 20 meters to ensure that the UAV is not subject to electromagnetic interference from the power transmission line and can collect high-quality point cloud data at the same time. During the data collection process, the specific flight altitude of the drone is dynamically adjusted according to different terrains; when the drone scans the power lines, trees and related facilities from different sides at a side-view angle, the angle range is set at 10 to 35 degrees, and the specific angle is adjusted according to the direction of the power lines and terrain changes to cover the lines and the trees and terrain around them; When the drone needs to scan the power transmission lines and related environment from a bird’s-eye view, the angle is set between 45 and 90 degrees to enhance data integrity; When encountering complex terrain, the UAV adopts a segmented scanning strategy, dividing the scanning area into multiple segments according to terrain changes. Each segment is controlled within the length range of 50 to 100 meters to plan a flight path that suits it, and adjusts the UAV's flight altitude and scanning angle in real time according to the terrain.

[0026] In step S2, the preprocessing operation includes distinguishing power line segments, ground point filtering, and point cloud downsampling. The specific steps include: According to the geometric distribution characteristics of the point cloud and the location of the tower, the power line point cloud is divided into different sections between two adjacent towers; the tower point cloud is assumed to be , the center point of the tower is taken as the tower position ; Divide the power line point cloud into multiple sections according to the distance between adjacent towers, and determine whether the power line point cloud is within the range of two towers: in For the power line point cloud, is a specific point cloud in the segment, Represents the segment point cloud, and They are the center points of two adjacent towers respectively; In the initial point cloud, the CSF algorithm is used to filter the ground points and separate the ground points from the non-ground points. The main steps are as follows: the point cloud is inverted and the ground points are made to be the top of the point cloud through coordinate transformation; the simulated cloth is placed on top of the point cloud and divided into node grids. The number of nodes is determined by the resolution. Decision: Project the point cloud and cloth nodes onto the horizontal plane, based on the vertical distance between the node and the point cloud. Limit the range of node movement; calculate the gravity of each node and tension Action: Update node displacement and repeat iterations until the node height change meets the threshold. or the maximum number of iterations is reached ; Calculate the distance between the point cloud and the cloth model based on the node position , the distance is less than the set threshold Points are classified as ground points, and the rest are non-ground points; The voxel downsampling method is used for point cloud data to reduce the amount of data to improve computational efficiency. The point cloud is divided into voxel grids , the side length of each voxel is , the center point of the point cloud in each voxel grid is used as the representative point of the voxel , output the downsampled point cloud set .

[0027] In step S3, semantic segmentation is performed on the point cloud data based on the PointNet++ model, and the semantic segmentation model is trained on a large number of annotated power line and tree point cloud data sets. The specific steps include: first, grouping and sampling the input point cloud layer by layer to determine the local neighborhood of each representative point; then, in each local neighborhood, A shared multi-layer perceptron is used to extract local features of relative positions. The calculation formula is: , in It represents feature concatenation, and MLP is a multi-layer perceptron; is the coordinate of the current point cloud, are the coordinates of the center point of the local neighborhood where the point cloud is located, Represents the feature vector of the current point cloud; And feature aggregation is performed through symmetric functions: in, Yes Local features of Represents the maximum pooling operation; Then, multi-scale features from local to global are extracted recursively in a hierarchical manner. The feature transfer formula of each layer is: in It is Characteristics of the layer; It is grouping, which means grouping the surrounding points based on the sampled point cloud to form a new local area. It is sampling, which means downsampling the point cloud, reducing the number of points, and constructing a multi-level feature representation; Finally, the features are input into the fully connected network, the classification or segmentation results are calculated, and the network parameters are optimized using cross entropy loss. At the same time, the regularization term is combined to improve the generalization ability of the model. The loss function calculation formula for the semantic segmentation task is: in is the one-hot encoding of the real label, is the predicted probability output by the model, Represents the cross entropy loss, which is used to calculate the difference between the category distribution predicted by the model and the true distribution. represents the regularization term, which is used to prevent the model from overfitting. is the weight of the regularization term.

[0028] In step S4, the three-dimensional point cloud of the power line is modeled, and the method includes the following steps: first, a straight line model is used Calculate the slope , get the angle between the straight line model and the X axis ;in, represents the slope of the straight line, represents the intercept of the straight line on the Y axis, Indicates the angle between the straight line model and the X-axis; Then according to Construct a rotation matrix ,pass The power line point cloud can be rotated around the Z axis and roughly parallel to the X axis, so that the power line point cloud is more consistent with the parabola characteristics on the vertical projection plane; Selecting the XOZ axis as the vertical projection plane of the electric line, the parabolic model can be obtained. ; Assume the number of point clouds of a single power line is , for the parabolic model, the error equation calculation formula can be obtained as: ; The three parameters of the parabola model can be solved by taking the first-order partial derivative of the equation .

[0029] In step S5, the segmented profile analysis method includes selecting multiple profile positions along the power line path, and the specific steps include: firstly extracting the connecting lines of adjacent points in the power line model, and the calculation formula of the connecting lines is: This means that the vectors between adjacent points , that is, the line segment connecting two points; Connect the wire As the axis of the section, the plane normal vector perpendicular to the axis is It can be obtained by normalizing the coordinates of the axis , Section plane It is expressed as: in is any point on the plane, It is the starting point of the power line.

[0030] In step S5, the shortest distance between the tree and the power line in the section plane is calculated, and the measured distance is compared with a preset safety threshold. The specific steps include: assuming that the tree point cloud , which is projected onto the section plane , the projection formula is: in, is the normal vector of the plane, is a point in the tree point cloud, It is the starting point of the power line; The calculation formula for the shortest distance from the tree point cloud to the power line is: in is the closest point on the electric line segment within the section.

[0031] In step S6, the calculated shortest distance With the preset safety threshold For comparison, if , an alarm or notification will be issued to remind relevant personnel to take timely measures to prevent the occurrence of wildfires.

[0032] Example: This embodiment discloses a method for automatic tree distance measurement based on semantic segmentation of drone point cloud for forest fire prevention in power transmission corridors. Figure 1 The specific situation is as follows: Figure 2 The working principle diagram of the UAV equipped with LiDAR equipment was shown. The UAV equipped with high-precision LiDAR equipment scans the mountain forest area and flies according to the preset flight path and parameters, so that the UAV can cover the tree area along the transmission line.

[0033] During the data collection process, the horizontal distance between the drone equipped with the LiDAR system and the transmission line is maintained between 5 meters and 20 meters to ensure that the drone is not subject to electromagnetic interference from the transmission line, and can collect high-quality point cloud data at the same time. The specific height is dynamically adjusted according to different terrains. The drone scans the transmission line, trees and related facilities from different sides at a side-view angle, with the angle set at 10 to 35 degrees. The specific angle is adjusted according to the direction of the transmission line and the terrain changes to cover the line and the trees and terrain around it; when necessary, the drone scans the transmission line and related environment at a top-down angle, with the angle set between 45 and 90 degrees to enhance data integrity. When encountering complex terrain, a segmented scanning strategy is adopted to divide the scanned area into multiple sections according to terrain changes, with each section controlled within a length range of 50 to 100 meters. At the same time, the drone's flight altitude and scanning angle are adjusted in real time according to the terrain.

[0034] The drone flew at a speed of 8 meters per second, the scanning frequency of the lidar was 120 Hz, the flight path length was 15 kilometers, the flight time was 1,875 seconds, and the point cloud data volume generated was 1.25 million points. During the flight, the lidar continuously emitted laser pulses, received reflected signals, generated 3D point cloud data of transmission lines, trees and related facilities, and constructed a 3D model basis of the real scene.

[0035] The acquired point cloud data is imported into the point cloud processing software CloudCompare. First, the three-dimensional point cloud data is divided into different segments. According to the geometric distribution characteristics of the point cloud and the location of the tower, the power line point cloud is divided into different segments between two adjacent towers. , the center point of the tower is taken as the tower position ; Divide the power line point cloud into multiple sections according to the distance between adjacent towers, and determine whether the power line point cloud is within the range of two towers: in For the power line point cloud, and They are the center points of two adjacent towers.

[0036] Next, the CSF algorithm is used to filter the ground points in the initial point cloud to remove noise points and outliers. The main steps are as follows: the point cloud is inverted and the ground points are made to be the top of the point cloud through coordinate transformation; the simulated cloth is placed on top of the point cloud and divided into node grids, with the number of nodes determined by the resolution. Meters determine; project the point cloud and cloth nodes onto the horizontal plane, and according to the vertical distance between the node and the point cloud Meters limit the node movement range; calculate the gravity of each node and tension Action: Update node displacement and repeat iterations until the node height change meets the threshold. or the maximum number of iterations is reached times; calculate the distance between the point cloud and the cloth model based on the node position , the distance is less than the set threshold The points with a size of 1 meter are classified as ground points, and the rest are non-ground points.

[0037] Then, the data is normalized and the coordinate values ​​of the point cloud data are converted to the interval [0,1] to facilitate subsequent model training and calculation.

[0038] Finally, the voxel downsampling method is used to reduce the density of the point cloud data to the original . Divide the point cloud into a voxel grid , the side length of each voxel is The center point of the point cloud in each voxel grid is used as the representative point of the voxel. , output the downsampled point cloud set .

[0039] The point cloud processing software CloudCompare was used to annotate the collected point cloud data such as power lines and trees. The annotation content included information such as power lines, trees, poles, and buildings. The power section dataset was aligned with the S3DIS dataset format, and the annotated data was divided into a training set and a test set, with a ratio of 70% and 30% respectively. During the training process, data enhancement techniques such as random rotation, scaling, and translation were used to randomly transform the data each time training to expand the number of training samples. Appropriate hyperparameters, such as learning rate and batch size, were set, and the weights of the model were continuously adjusted through the back propagation algorithm to optimize the performance of the model on the test set. When the accuracy of the model on the test set reached more than 90%, the model training was considered complete. The three-dimensional coordinate information of the point cloud marked as power lines and trees was extracted to provide accurate data for subsequent distance measurement.

[0040] The point cloud data is semantically segmented based on the PointNet++ model, and the semantic segmentation model is trained on a large number of annotated power line and tree point cloud data sets. The specific steps include: First, the input point cloud is grouped and sampled layer by layer to determine the local neighborhood of each representative point; then, in each local neighborhood A shared multi-layer perceptron is used to extract local features of relative positions. The calculation formula is: , in It represents feature concatenation, and MLP is a multi-layer perceptron; And feature aggregation is performed through symmetric functions: in, Yes Local features of Then, multi-scale features from local to global are extracted recursively in a hierarchical manner. The feature transfer formula of each layer is: in It is Characteristics of the layer; Finally, the features are input into the fully connected network, the classification or segmentation results are calculated, and the network parameters are optimized using cross entropy loss. At the same time, the regularization term is combined to improve the generalization ability of the model. The loss function calculation formula for the semantic segmentation task is: in is the one-hot encoding of the real label, is the predicted probability output by the model, is the weight of the regularization term.

[0041] The training data set contains 35 labeled power stage samples. During the training process, the learning rate is set to 0.0008, the batch size is 64, and the training is performed for 120 epochs.

[0042] The three-dimensional point cloud of the power line is modeled by first using the straight line model Calculate the slope , get the angle between the straight line model and the X axis ;in, represents the slope of the straight line, represents the intercept of the straight line on the Y axis, Indicates the angle between the straight line model and the X-axis; Then according to Construct a rotation matrix ,pass The power line point cloud can be rotated around the Z axis and roughly parallel to the X axis, so that the power line point cloud is more consistent with the parabola characteristics on the vertical projection plane; Selecting the XOZ axis as the vertical projection plane of the electric line, the parabolic model can be obtained. ; Assume the number of point clouds of a single power line is , for the parabolic model, the error equation calculation formula can be obtained as: ; The three parameters of the parabola model can be solved by taking the first-order partial derivative of the equation .

[0043] The constructed power line model is segmented along its length, and the length of each segment is set to 10 meters. For each segment of the power line, a section perpendicular to the segment of the power line is generated. The steps of segmenting the power line include: first, extracting the connecting lines of adjacent points in the power line model, and the calculation formula of the connecting lines is: This means that the vectors between adjacent points , that is, the line segment connecting two points; Connect the wire As the axis of the section, the plane normal vector perpendicular to the axis is It can be obtained by normalizing the coordinates of the axis , Section plane It is expressed as: in is any point on the plane, It is the starting point of the power line.

[0044] Calculate the shortest distance between the tree and the power line in the section plane, and compare the measured distance with the preset safety threshold. The specific steps include: set the tree point cloud , which is projected onto the section plane , the projection formula is: in, is the normal vector of the plane, is a point in the tree point cloud, It is the starting point of the power line; The calculation formula for the shortest distance from the tree point cloud to the power line is: in is the closest point on the electric line segment within the section.

[0045] The measured distance is compared with a preset safety threshold. If the measured distance is less than the safety threshold, an alarm signal is sent out through the alarm device to prevent the occurrence of wildfires.

[0046] The contents disclosed above are only preferred feasible embodiments of the present invention, and do not limit the protection scope of the present invention. Therefore, all equivalent technical changes made using the contents of the present invention specification and drawings are included in the protection scope of the present invention. In addition, the elements therein can be updated as technology develops.

Claims

1. A tree ranging method based on semantic segmentation of drone point cloud for forest fire prevention in power transmission corridors, characterized by: The method first obtains three-dimensional point cloud data of power transmission lines, trees and related facilities through unmanned aerial vehicle scanning, and then preprocesses the initial data; applies a semantic segmentation model based on deep learning to analyze the preprocessed data, uses a deep learning algorithm to improve the accuracy of semantic segmentation to automatically identify the point clouds of power lines and trees, and extracts their three-dimensional coordinate information; then models the three-dimensional point clouds of power lines to obtain a complete power line model; and measures and calculates the distance between power lines and trees through a segmented profile method, and compares it with a preset safety threshold.

2. The tree ranging method of UAV point cloud semantic segmentation for forest fire prevention in power transmission corridor according to claim 1 is characterized by: The method comprises the following steps: Step S1: Use a drone equipped with a laser radar system to scan the mountainous forest area to obtain three-dimensional point cloud data of power transmission lines, trees and related facilities; Step S2: pre-process the collected point cloud data, extract target point cloud data, and remove ground points and irrelevant noise; Step S3: classify the target point cloud data based on the semantic segmentation model of deep learning, identify the point cloud data of wires and trees, and extract their three-dimensional coordinate information; Step S4, modeling the three-dimensional point cloud of the power lines to obtain a complete power line model; Step S5, measuring the distance between the power line and the tree by segmented profile analysis method; Step S6: Compare the measured distance with a preset safety threshold to determine whether it is lower than the safety threshold.

3. The tree ranging method of UAV point cloud semantic segmentation for forest fire prevention in power transmission corridor according to claim 2 is characterized by: In step S1, when a drone equipped with a laser radar system is used to scan a mountain forest area, the distance between the drone and the power transmission line is kept within a safe range, and the laser radar's scanning angle of the power transmission line is mainly a side view angle; When the drone scans the power transmission corridor in the mountainous area, if the terrain hinders the flight path, the drone adjusts the flight path so that the lidar's scanning angle of the transmission line can be combined with the bird's-eye view to collect point cloud data, and a segmented scanning strategy is used to collect point cloud data in complex terrain.

4. The tree ranging method of UAV point cloud semantic segmentation for power transmission corridor forest fire prevention according to claim 3 is characterized by: During the data collection process, the horizontal distance between the UAV equipped with the LiDAR system and the power transmission line is maintained between 5 meters and 20 meters to ensure that the UAV is not subject to electromagnetic interference from the power transmission line and can collect high-quality point cloud data at the same time. During the data collection process, the specific flight altitude of the drone is dynamically adjusted according to different terrains; when the drone scans the power lines, trees and related facilities from different sides at a side-view angle, the angle range is set at 10 to 35 degrees, and the specific angle is adjusted according to the direction of the power lines and terrain changes to cover the lines and the trees and terrain around them; When the drone needs to scan the power transmission lines and related environment from a bird’s-eye view, the angle is set between 45 and 90 degrees to enhance data integrity; When encountering complex terrain, the UAV adopts a segmented scanning strategy, dividing the scanning area into multiple segments according to terrain changes. Each segment is controlled within the length range of 50 to 100 meters to plan a flight path that suits it, and adjusts the UAV's flight altitude and scanning angle in real time according to the terrain.

5. The tree ranging method of UAV point cloud semantic segmentation for power transmission corridor forest fire prevention according to claim 2 is characterized by: In step S2, the preprocessing operation includes distinguishing power line segments, ground point filtering, and point cloud downsampling. The specific steps include: According to the geometric distribution characteristics of the point cloud and the location of the tower, the power line point cloud is divided into different sections between two adjacent towers; the tower point cloud is assumed to be , the center point of the tower is taken as the tower position ; Divide the power line point cloud into multiple sections according to the distance between adjacent towers, and determine whether the power line point cloud is within the range of two towers: in For the power line point cloud, is a specific point cloud in the segment, Represents the segment point cloud, and They are the center points of two adjacent towers respectively; In the initial point cloud, the CSF algorithm is used to filter the ground points and separate the ground points from the non-ground points. The main steps are as follows: the point cloud is inverted and the ground points are made to be the top of the point cloud through coordinate transformation; the simulated cloth is placed on top of the point cloud and divided into node grids. The number of nodes is determined by the resolution. Decision: Project the point cloud and cloth nodes onto the horizontal plane, based on the vertical distance between the node and the point cloud. Limit the range of node movement; calculate the gravity of each node and tension Action: Update node displacement and repeat iterations until the node height change meets the threshold. or the maximum number of iterations is reached ; Calculate the distance between the point cloud and the cloth model based on the node position , the distance is less than the set threshold Points are classified as ground points, and the rest are non-ground points; The voxel downsampling method is used for point cloud data to reduce the amount of data and improve computational efficiency; Divide the point cloud into a voxel grid , the side length of each voxel is , the center point of the point cloud in each voxel grid is used as the representative point of the voxel , output the downsampled point cloud set .

6. The tree ranging method of UAV point cloud semantic segmentation for power transmission corridor forest fire prevention according to claim 2 is characterized by: In step S3, semantic segmentation is performed on the point cloud data based on the PointNet++ model, and the semantic segmentation model is trained on a large number of annotated power line and tree point cloud data sets. The specific steps include: first, grouping and sampling the input point cloud layer by layer to determine the local neighborhood of each representative point; then, in each local neighborhood, A shared multi-layer perceptron is used to extract local features of relative positions. The calculation formula is: ， in It represents feature concatenation, and MLP is a multi-layer perceptron; is the coordinate of the current point cloud, are the coordinates of the center point of the local neighborhood where the point cloud is located, Represents the feature vector of the current point cloud; And feature aggregation is performed through symmetric functions: in, Yes Local features of Represents the maximum pooling operation; Then, multi-scale features from local to global are recursively extracted in a hierarchical manner; The feature transfer formula of each layer is: in It is Characteristics of the layer; It is grouping, which means grouping the surrounding points based on the sampled point cloud to form a new local area. It is sampling, which means downsampling the point cloud, reducing the number of points, and constructing a multi-level feature representation; Finally, the features are input into the fully connected network, the classification or segmentation results are calculated, and the network parameters are optimized using cross entropy loss. At the same time, the regularization term is combined to improve the generalization ability of the model. The loss function calculation formula for the semantic segmentation task is: in is the one-hot encoding of the real label, is the predicted probability output by the model, Represents the cross entropy loss, which is used to calculate the difference between the category distribution predicted by the model and the true distribution. represents the regularization term, which is used to prevent the model from overfitting. is the weight of the regularization term.

7. The tree ranging method of UAV point cloud semantic segmentation for forest fire prevention in power transmission corridor according to claim 2 is characterized by: In step S4, the three-dimensional point cloud of the power line is modeled, and the method includes the following steps: first, a straight line model is used Calculate the slope , get the angle between the straight line model and the X axis ;in, represents the slope of the straight line, represents the intercept of the straight line on the Y axis, Indicates the angle between the straight line model and the X-axis; Then according to Construct a rotation matrix ,pass The power line point cloud can be rotated around the Z axis and roughly parallel to the X axis, so that the power line point cloud is more consistent with the parabola characteristics on the vertical projection plane; Selecting the XOZ axis as the vertical projection plane of the electric line, the parabolic model can be obtained. ; Assume the number of point clouds of a single power line is , for the parabolic model, the error equation calculation formula can be obtained as: ； The three parameters of the parabola model can be solved by taking the first-order partial derivative of the equation .

8. The tree ranging method of UAV point cloud semantic segmentation for power transmission corridor forest fire prevention according to claim 2 is characterized by: In step S5, the segmented profile analysis method includes selecting multiple profile positions along the power line path, and the specific steps include: firstly extracting the connecting lines of adjacent points in the power line model, and the calculation formula of the connecting lines is: This means that the vectors between adjacent points , that is, the line segment connecting two points; Connect the wire As the axis of the section, the plane normal vector perpendicular to the axis is It can be obtained by normalizing the coordinates of the axis , Section plane It is expressed as: in is any point on the plane, It is the starting point of the power line.

9. The tree ranging method of UAV point cloud semantic segmentation for power transmission corridor forest fire prevention according to claim 2 is characterized by: In step S5, the shortest distance between the tree and the power line in the section plane is calculated, and the measured distance is compared with a preset safety threshold. The specific steps include: assuming that the tree point cloud , which is projected onto the section plane , the projection formula is: in, is the normal vector of the plane, is a point in the tree point cloud, It is the starting point of the power line; The calculation formula for the shortest distance from the tree point cloud to the power line is: in is the closest point on the electric line segment within the section.

10. The tree ranging method of UAV point cloud semantic segmentation for forest fire prevention in power transmission corridor according to claim 9, characterized in that: In step S6, the calculated shortest distance With the preset safety threshold For comparison, if , an alarm or notification will be issued to remind relevant personnel to take timely measures to prevent the occurrence of wildfires.

Images