Distribution line point cloud data interval time period timing classification system and method

Abstract

The invention discloses a distribution line point cloud data interval time period timing classification system and method. The system comprises a point cloud data acquisition module and an image acquisition module. The timed task module controls the point cloud data acquisition module and the image acquisition module to operate at preset time intervals, point cloud data acquisition of the distribution line is carried out through the point cloud data acquisition module, and image acquisition of the distribution line is carried out through the image acquisition module; the collected point cloud data is subjected to denoising, filtering and down-sampling processing through a point cloud preprocessing module, and then point cloud feature extraction and point cloud classification are carried out; the collected image data is subjected to image feature extraction and then is identified by using a target detection algorithm through an image identification module; the point cloud classification result and the image recognition result are input into a fusion verification module to perform corresponding verification on the point cloud classification result and the recognition result of the image recognition module, and point cloud classification data which are verified to be consistent are input into a data storage module to be stored; and the accuracy of point cloud data classification is greatly improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of point cloud data processing; in particular, it relates to a system and method for classifying point cloud data of distribution lines at regular intervals during intervals. Background Art

[0002] During the establishment, use, and maintenance of the three-dimensional management platform for distribution lines, it is necessary to regularly update the point cloud scanning of distribution lines to adapt to environmental changes and line renovations. Each time the point cloud data is scanned and updated, a huge amount of point cloud data will be generated. The point cloud data includes object information such as trees, poles, wires, and insulating terminals. When marking and classifying the object point cloud features in the point cloud data, manual marking, framing, and assignment classification are carried out by combining the point cloud object features. However, the workload is too large, and due to frequent updates, it is not a long-term strategy. Currently, the industry generally uses algorithms to extract and automatically identify the object features such as trees, poles, and wires in the point cloud, and achieve automatic classification and identification. In this way, the error rate is relatively high, and the error rate needs to be further reduced. Summary of the Invention

[0003] The technical problem to be solved by the present invention is: to provide a system and method for classifying point cloud data of distribution lines at regular intervals during intervals, so as to solve the technical problems that in the prior art, when processing point cloud data of distribution network lines, algorithms are used to extract and automatically identify object features such as trees, poles, and wires in the point cloud, and achieve automatic classification and identification, and the error rate is relatively high.

[0004] Technical Solution of the Present Invention

[0005] A system for classifying point cloud data of distribution lines at regular intervals during intervals, the system includes a point cloud data acquisition module and an image acquisition module; the timing task module controls the operation of the point cloud data acquisition module and the image acquisition module every preset interval. The point cloud data of the distribution line is acquired through the point cloud data acquisition module, and the image of the distribution line is acquired through the image acquisition module;

[0006] The point cloud data acquired by the point cloud data acquisition module is subjected to denoising, filtering, and downsampling processing through the point cloud preprocessing module. After the processing, the point cloud features are extracted from the point cloud data, and then the point cloud is classified; the image features are extracted from the image data acquired by the image acquisition module, and then the target detection algorithm is used by the image recognition module for recognition;

[0007] The point cloud classification result and the recognition result of the image recognition module are input into the fusion verification module. The fusion verification module performs corresponding verification on the point cloud classification result and the recognition result of the image recognition module. The point cloud classification data with consistent verification is input into the data storage module for storage.

[0008] The fusion verification module extracts the inconsistent point cloud classification data and the corresponding image data, and makes a judgment through manual error correction. Manually combine the extracted point cloud model and image information to judge the category of the point cloud data; when the point cloud classification is correct but the image recognition is incorrect, resulting in inconsistent verification, the point cloud data is directly input into the data storage module for storage; when the point cloud classification is incorrect but the image recognition is correct, resulting in inconsistent verification, the segment of point cloud data and the corresponding image data are input into the point cloud data supplement and optimization system; in the point cloud data supplement and optimization system, correct the point cloud classification information, and supplement and optimize the point cloud data according to the image data. After completion of the supplement and optimization, the point cloud data is input into the data storage module for storage.

[0009] The methods of denoising include: setting a distance threshold T according to the overall distribution and noise level of the point cloud data d , if the average distance of the point p in the point cloud i is greater than T d , then the point is determined as an outlier and deleted from the point cloud; the filtering method includes: filtering according to the coordinate range of the point cloud in a certain axis direction. For the x-axis direction, set a minimum coordinate value x min and a maximum coordinate value x max . Traverse all points in the point cloud. Only when the x coordinate of the point satisfies x min ≤x≤x max , the point will be retained. The same operation is performed for the y-axis and z-axis directions to filter out the points not within the specified spatial range; the downsampling method includes: setting a downsampling ratio p, traversing all points in the point cloud, generating a random number r (0≤r≤1) for each point. If r≤p, then retain the point, otherwise delete the point.

[0010] When extracting point cloud features, extract the geometric features of the point cloud, including the coordinates, normal vectors and curvatures of the points, for classifying objects. When classifying the point cloud, use a deep learning model for classification, and classify the towers, trees, wires and objects in the point cloud data and label their attributes.

[0011] The point cloud data supplementary optimization system includes: constructing a Difference of Gaussian (DoG) pyramid to detect key points in the image. The DoG pyramid is obtained by subtracting Gaussian blurred images of different scales. The key point (x, y) is a local extreme point in the DoG space, that is, it satisfies that among the pixels at this point and its neighboring pixels, its DoG value is the largest or the smallest. For each key point, calculate its orientation histogram to determine the main orientation of the key point. Taking the pixels within a certain radius range centered on the key point, calculate its gradient direction and magnitude. Divide the 360° range into 8 intervals, and count the number of pixels with gradient directions in each interval to form an orientation histogram. The direction corresponding to the peak of the histogram is the main orientation of the key point. Generate a descriptor for the key point. Taking a neighborhood of a certain size centered on the key point, taking 16×16 pixels as an example, divide it into 4×4 sub-regions. For each sub-region, calculate the gradient orientation histogram in 8 directions, and a total of 4×4×8 = 128-dimensional feature descriptors are obtained.

[0012] Perform normal vector estimation for the point cloud data:

[0013] For a point p in the point cloud i (x i , y i , z i ), select the neighboring points {p i,j}(j = 1, 2,..., k, where k is the number of neighboring points) within a certain radius range of it. Use the least squares method to fit a plane to these neighboring points. The plane equation is ax + by + cz + d = 0. Determine the plane parameters by solving the following system of equations:

[0014] Take the minimum value. Take the partial derivatives of this formula with respect to a, b, c, and d respectively and set them to 0 to obtain a system of linear equations. Solve this system of equations to obtain the plane parameters a, b, c, and d. The normal vector n of point p i is n i = (a, b, c);

[0015] Perform curvature calculation for the point cloud data:

[0016] Calculate the curvature of a point according to the normal vector and the distribution of neighboring points. Let the projection height of the neighboring points of point p i in the direction of its normal vector be h i (j = 1, 2,..., k). Then the curvature C of point p i can be calculated by the following formula: i where where

[0017] Establish feature correspondence. For each SIFT feature point f in the image iFor each feature point pj in the point cloud, calculate the feature similarity between them. The similarity S can be calculated using the following formula ij :[[]]

[0018]

[0019] where (x fi , y fi , z fi ) and (x pj , y pj , z pj ) are the coordinates of the feature point in the image and the point cloud respectively. For the image feature point, the z coordinate can be set to 0 or determined according to the camera parameters. D desc (f i , p j ) is the distance between the feature point descriptors, θ fi,pj is the angle between the main direction of the image feature point and the normal vector of the point cloud feature point. α, β, γ are weight coefficients, and α + β + γ = 1. By setting the similarity threshold T S , select the feature point pairs with similarity greater than T S to establish a preliminary correspondence between the image feature points and the point cloud feature points.

[0020] The point cloud data supplementary and optimization system also includes: RANSAC filtering. Use the Random Sample Consensus (RANSAC) algorithm to further filter accurate feature matching pairs. Randomly select a group of feature point pairs and calculate the transformation model between them; according to this transformation model, project or transform other feature point pairs and calculate the distance error between the transformed feature points. If the error is less than the set threshold T e , then regard this feature point pair as an inlier. Repeat the above process and select the transformation model with the largest number of inliers as the final matching model, and obtain the corresponding accurate matching feature point pairs.

[0021] When supplementing the point cloud data, for the missing areas in the point cloud data, supplement according to the information of the matching image feature points. Assume that there are matching image feature points f m and the corresponding point cloud feature points p n . Based on p n , generate new point cloud points according to the neighborhood information of the image feature points. If there are k feature points {f m}(j = 1, 2,..., k) in the neighborhood of f m,j in the image, and their relative position relationship with f m is known, then in the point cloud space, with p n as the center, generate k new point cloud points {p new,j}, the coordinates of the new point can be calculated by the following formula: p new,j = p n + λ(f m,j - f m ), where λ is the scaling coefficient, which is determined according to the spatial proportional relationship between the image and the point cloud.

[0022] When optimizing the point cloud data, the least squares method is used to optimize the position of the point cloud. Suppose there are n points {p i} (i = 1, 2,..., n) in the point cloud. The goal is to make the point cloud match the image features better. Define an objective function E:

[0023] where m is the number of image feature points for matching, l is the number of neighboring point cloud points of point p i , S ij is the similarity between point p i and the matching image feature point, (p i - p i,k ) 2 is the square of the distance between point p i and its neighboring point p i,k ; Take the partial derivative of the objective function E with respect to the coordinates of the point cloud points and set the partial derivative to 0 to obtain a system of linear equations. By solving this system of linear equations, update the coordinates of the point cloud points to achieve the optimization of the point cloud.

[0024] Advantages of the present invention:

[0025] The timing classification system for point cloud data of distribution lines of the present invention can verify the classification correctness of point cloud data in combination with image recognition, greatly improving the classification correctness of point cloud data;

[0026] When the verification of the present invention system is inconsistent due to incorrect point cloud classification, it can use image data to supplement and optimize the point cloud data, improving the quality of the point cloud data of the corresponding object;

[0027] The present invention processes the image through an algorithm, combines multi-dimensional features such as SIFT features, normal vectors, and curvatures of the image, and can supplement and optimize the point cloud data, making the optimized point cloud more conform to the shape and features of the actual object.

[0028] It solves the technical problems in the prior art that when processing point cloud data of distribution network lines, algorithms are used to extract and automatically identify the features of objects such as trees, poles, and wires in the point cloud to achieve automatic classification and identification, and the error rate is relatively high. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 is a schematic flowchart of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0030] The present invention provides a technical solution: a time-interval timing classification system for point cloud data of distribution lines, as Figure 1 shown in the figure, which includes a timing task module, a point cloud data acquisition module, an image acquisition module, a point cloud preprocessing module, point cloud feature extraction, point cloud classification, image feature extraction, an image recognition module, a fusion verification module, and a data storage module. The timing task module controls the operation of the point cloud data acquisition module and the image acquisition module at preset time intervals. The point cloud data of the distribution line is collected through the point cloud data acquisition module, and the image of the distribution line is collected through the image acquisition module. The point cloud data acquisition module is a three-dimensional laser scanner, and the image acquisition module is a camera device. Both are installed on an unmanned aerial vehicle and are set to take off regularly through the timing task module to collect point cloud data and images of the distribution line along the set route.

[0031] The point cloud data collected by the point cloud data acquisition module is subjected to denoising, filtering, and downsampling through the point cloud preprocessing module. Denoising: Set a distance threshold T according to the overall distribution and noise level of the point cloud data. d If the average distance of the point p i in the point cloud is greater than T d , then the point is determined as an outlier and deleted from the point cloud. Filtering: Filter according to the coordinate range of the point cloud in a certain coordinate axis direction. For the x-axis direction, set a minimum coordinate value x min and a maximum coordinate value x max . Traverse all the points in the point cloud. Only when the x coordinate of the point satisfies x min ≤x≤x max , the point will be retained. The same operation is performed for the y-axis and z-axis directions to filter out the points not within the specified spatial range. Downsampling: Set a downsampling ratio p. Traverse all the points in the point cloud, generate a random number r (0≤r≤1) for each point. If r≤p, then retain the point; otherwise, delete the point.

[0032] When extracting the features of the processed point cloud data, the geometric features of the point cloud are extracted, including the coordinates, normal vectors, and curvatures of the points, which are used to classify objects, and then point cloud classification is performed. For point cloud classification, deep learning models, such as networks specialized in processing point cloud data like PointNet++ and DGCNN, are used for classification. These networks can directly process point cloud data, learn the feature representations of the point cloud, and perform classification to classify objects such as poles, trees, wires, etc. in the point cloud data and assign attributes.

[0033] The image data collected by the image acquisition module is subjected to image feature extraction. The deep convolutional neural network is used to extract the high-level features of the image, and then the object detection algorithm is used by the image recognition module for recognition, such as Faster R-CNN, YOLOv4, etc. to recognize the tower poles, insulating terminals, trees, etc. in the image;

[0034] The point cloud classification result and the recognition result of the image recognition module are input into the fusion verification module. In the fusion verification module, the point cloud data and the image data are expressed in a map form, the point cloud data map and the image data map are aligned, and the objects are mutually corresponding according to the map position coordinates, and it is judged whether the recognition attributes of the objects in the same coordinate position in the point cloud classification and the image recognition attributes are consistent; thus, the point cloud classification result and the recognition result of the image recognition module are correspondingly verified, and the consistent point cloud classification data is input into the data storage module for storage.

[0035] The fusion verification module extracts the inconsistent point cloud classification data and the corresponding image data, and makes a judgment through manual error correction. Manually combine the extracted point cloud model and image information to judge the category of the point cloud data; when the point cloud classification is correct but the image recognition is incorrect, resulting in inconsistent verification, the point cloud data is directly input into the data storage module for storage; when the point cloud classification is incorrect but the image recognition is correct, resulting in inconsistent verification, the section of point cloud data and the corresponding image data are input into the point cloud data supplement and optimization system again; in the point cloud data supplement and optimization system, correct the point cloud classification information, and supplement and optimize the point cloud data according to the image data. After completion of the supplement and optimization, the point cloud data is input into the data storage module for storage.

[0036] The point cloud data supplement and optimization system includes SIFT feature extraction for image data, specifically as follows:

[0037] Construct a Difference of Gaussian pyramid (DoG) to detect the key points in the image. The DoG pyramid is obtained by subtracting Gaussian blurred images of different scales. The key point (x, y) is a local extreme point in the DoG space, that is, it satisfies that in this point and its neighboring pixels, its DoG value is the largest or the smallest; for each key point, calculate its direction histogram to determine the main direction of the key point; taking the key point as the center, take the pixels within a certain radius range, calculate its gradient direction and amplitude, divide the 360° range into 8 intervals, and count the number of pixels in each interval with the gradient direction to form a direction histogram; the direction corresponding to the peak of the histogram is the main direction of the key point, generate the descriptor of the key point, taking the key point as the center, take a neighborhood of a certain size, taking 16×16 pixels as an example, divide it into 4×4 sub-regions, for each sub-region, calculate the gradient direction histogram of 8 directions, and a total of 4×4×8 = 128-dimensional feature descriptors are obtained;

[0038] Normal vector estimation for point cloud data:

[0039] For a point p in the point cloud i (x i ,y i ,z i ), select its neighborhood points within a certain radius {p i,j}(j = 1, 2,..., k, where k is the number of neighborhood points), and use the least squares method to fit a plane to these neighborhood points. The plane equation is ax + by + cz + d = 0, and the plane parameters are determined by solving the following system of equations:

[0040] Take the minimum value, take the partial derivatives of this formula with respect to a, b, c, and d respectively and set them to 0 to obtain a system of linear equations. Solving this system of equations can obtain the plane parameters a, b, c, and d; the normal vector n of point p i =(a, b, c); i

[0041] Curvature calculation for point cloud data:

[0042] Calculate the curvature of a point according to the distribution of the normal vector and neighborhood points. Let the projection height of the neighborhood points of point p i in the direction of its normal vector be h i (j = 1, 2,..., k), then the curvature C of point p i can be calculated by the following formula: i where

[0043] Establish feature correspondence. For each SIFT feature point f in the image i and each feature point pj in the point cloud, calculate the feature similarity between them. The similarity S can be calculated using the following formula ij :

[0044]

[0045] where (x fi ,y fi ,z fi ) and (x pj ,y pj ,z pj ) are the coordinates of the feature points in the image and the point cloud respectively. For image feature points, the z coordinate can be set to 0 or determined according to the camera parameters. D desc (f i ,p j ) is the distance between the feature descriptors, and θ fi,pj ​​is the angle between the main direction of the image feature point and the normal vector of the point cloud feature point. α, β, and γ are weight coefficients, and α + β + γ = 1. By setting the similarity threshold T S , feature point pairs with a similarity greater than T S are selected to establish a preliminary correspondence between the image feature points and the point cloud feature points;

[0046] RANSAC filtering: Use the Random Sample Consensus (RANSAC) algorithm to further filter accurate feature matching pairs. Randomly select a set of feature point pairs and calculate the transformation model between them. According to this transformation model, project or transform other feature point pairs and calculate the distance error between the transformed feature points. If the error is less than the set threshold T e , then consider this feature point pair as an inlier. Repeat the above process multiple times and select the transformation model with the largest number of inliers as the final matching model, and obtain the corresponding accurate matching feature point pairs;

[0047] Supplementary of point cloud data is as follows: For the missing areas in the point cloud data, they are supplemented according to the information of the matching image feature points. Assume that there are matching image feature points f m and the corresponding point cloud feature points p n . Based on p n , new point cloud points are generated according to the neighborhood information of the image feature points. If there are k feature points {f m}(j = 1, 2,..., k) in the neighborhood of f m,j in the image, and their relative position relationship with f m is known, then in the point cloud space, with p n as the center, generate k new point cloud points {p new,j} according to the same relative position relationship. The coordinates of the new points can be calculated by the following formula: p new,j = p n +λ(f m,j - f m ), where λ is a scaling coefficient determined according to the spatial proportional relationship between the image and the point cloud;

[0048] Optimization of point cloud data is as follows: Use the least squares method to optimize the position of the point cloud. Suppose there are n points {p i}(i = 1, 2,..., n) in the point cloud. The goal is to make the point cloud match the image features better. Define an objective function E:

[0049] where m is the number of matching image feature points, l is the number of neighboring point cloud points of point p i , S ij is the similarity between point p i and the matching image feature point, (pi -p i,k ) 2 is the square of the distance between point p i and its neighboring point p i,k ; taking the partial derivative of the objective function E with respect to the coordinates of the point cloud points and setting the partial derivative to 0, a linear equation system is obtained. By solving this linear equation system, the coordinates of the point cloud points are updated to achieve the optimization of the point cloud.

[0050] An operation method for a time-interval timing classification system of distribution line point cloud data, the operation method comprising the following steps:

[0051] S1. Set the interval duration through the timing task module;

[0052] S2. When the point cloud classification data and the corresponding image data with inconsistent verification output by the fusion verification module need manual error correction, judge the result.

Claims

1. A time-interval timed classification system for point cloud data of a distribution line, characterized in that: The system includes a point cloud data acquisition module and an image acquisition module; the timing task module controls the operation of the point cloud data acquisition module and the image acquisition module at intervals of a preset period, acquires the point cloud data of the power distribution line through the point cloud data acquisition module, and acquires the image of the power distribution line through the image acquisition module; The point cloud data acquired by the point cloud data acquisition module is denoised, filtered and downsampled through the point cloud preprocessing module, the point cloud features are extracted from the processed point cloud data, and then the point cloud classification is performed; the image features are extracted from the image data acquired by the image acquisition module, and then the target detection algorithm is used by the image recognition module for recognition; The point cloud classification result and the recognition result of the image recognition module are input into the fusion verification module, and the fusion verification module performs corresponding verification on the point cloud classification result and the recognition result of the image recognition module, and the point cloud classification data with consistent verification is input into the data storage module for storage.

2. A time - interval - timed classification system for point - cloud data of a distribution line according to claim 1, characterized in that: The fusion verification module extracts the point cloud classification data and the corresponding image data with inconsistent verification, and makes a judgment through manual error correction. The human combines the extracted point cloud model and image information to judge the category of the point cloud data; When the verification is inconsistent due to correct point cloud classification but incorrect image recognition, the point cloud data is directly input into the data storage module for storage; when the verification is inconsistent due to incorrect point cloud classification but correct image recognition, the segment of point cloud data and the corresponding image data are input into the point cloud data supplementary and optimization system; in the point cloud data supplementary and optimization system, the point cloud classification information is corrected, and the point cloud data is supplemented and optimized according to the image data. After completion of the supplementation and optimization, the point cloud data is input into the data storage module for storage.

3. A point cloud data interval period timing classification system for a distribution line according to claim 1, characterized in that: The methods for denoising include: setting a distance threshold T according to the overall distribution and noise level of the point cloud data d , if the average distance of the point p in the point cloud i is greater than T d , then the point is determined as an outlier and deleted from the point cloud; The filtering method includes: filtering according to the coordinate range of the point cloud in a certain axis direction. For the x-axis direction, set a minimum coordinate value x min and a maximum coordinate value x max . Traverse all the points in the point cloud. Only when the x coordinate of the point satisfies x min ≤x≤x max , the point will be retained. The same operation is performed for the y-axis and z-axis directions to filter out the points not within the specified spatial range; The downsampling method includes: setting a downsampling ratio p, traversing all the points in the point cloud, generating a random number r (0≤r≤1) for each point. If r≤p, then retain the point, otherwise delete the point.

4. A point cloud data interval - period timing classification system for a distribution line according to claim 1, characterized in that: When extracting point cloud features, the geometric features of the point cloud are extracted, including the coordinates, normal vectors and curvatures of the points, which are used to classify objects. When performing point cloud classification, a deep learning model is used for classification, and the poles, trees, wires and objects in the point cloud data are classified and labeled with attributes.

5. A point cloud data interval period timing classification system for a distribution line according to claim 2, characterized in that: The point cloud data supplementary and optimization system includes: constructing a Difference of Gaussian (DoG) pyramid to detect key points in the image. The DoG pyramid is obtained by subtracting Gaussian blurred images of different scales. The key point (x, y) is a local extreme point in the DoG space, that is, it satisfies that in this point and its neighboring pixels, its DoG value is the largest or the smallest; for each key point, calculate its direction histogram to determine the main direction of the key point; taking the pixels within a certain radius range centered on the key point, calculate its gradient direction and amplitude, divide the 360° range into 8 intervals, and count the number of pixels with gradient directions in each interval to form a direction histogram; the direction corresponding to the peak of the histogram is the main direction of the key point, and generate a descriptor of the key point. Taking a neighborhood of a certain size centered on the key point, taking 16×16 pixels as an example, divide it into 4×4 sub-regions. For each sub-region, calculate the gradient direction histogram in 8 directions, and a total of 4×4×8 = 128-dimensional feature descriptors are obtained; Perform normal vector estimation for the point cloud data: For point p in the point cloud i (x i , y i , z i ), select the neighboring points within a certain radius range {p i,j}(j = 1, 2,..., k, where k is the number of neighboring points), and use the least squares method to fit a plane to these neighboring points. The plane equation is ax + by + cz + d = 0, and the plane parameters are determined by solving the following system of equations: Take the minimum value, take the partial derivatives of this formula with respect to a, b, c, and d respectively and set them to 0 to obtain a system of linear equations. Solving this system of equations can obtain the plane parameters a, b, c, and d; the point p i The normal vector n i =(a, b, c); Perform curvature calculation for the point cloud data: Calculate the curvature of a point based on the distribution of the normal vector and neighboring points. Let point p i The projection height of the neighboring points in the direction of its normal vector is h i (j = 1, 2, ..., k), then point p i The curvature C i Can be calculated by the following formula: Where Establish feature correspondence. For each SIFT feature point f in the image i and each feature point p in the point cloud j , calculate the feature similarity between them. The similarity S can be calculated using the following formula ij : where (x fi , y fi , z fi ) and (x pj , y pj , z pj ) are the coordinates of the feature points in the image and the point cloud respectively. For the image feature points, the z - coordinate can be set to 0 or determined according to the camera parameters. D desc (f i , p j ) is the distance between the feature point descriptors, θ fi,pj is the angle between the main direction of the image feature point and the normal vector of the point cloud feature point, α, β, γ are weight coefficients, and α + β+γ = 1. By setting the similarity threshold T S , the feature point pairs with similarity greater than T S are selected to establish the preliminary correspondence between the image feature points and the point cloud feature points.

6. A time-interval timing classification system for point cloud data of a distribution line according to claim 5, characterized in that: The point cloud data supplement and optimization system further includes: RANSAC screening, which uses the Random Sample Consensus (RANSAC) algorithm to further screen accurate feature matching pairs. A set of feature point pairs is randomly selected, and the transformation model between them is calculated. According to this transformation model, other feature point pairs are projected or transformed, and the distance error between the transformed feature points is calculated. If the error is less than the set threshold T e , then this feature point pair is regarded as an inlier. Repeat the above process, and select the transformation model with the largest number of inliers as the final matching model, and obtain the corresponding accurate matching feature point pairs.

7. A point cloud data interval period timing classification system for a distribution line according to claim 2, characterized in that: When supplementing the point cloud data, the missing areas in the point cloud data are supplemented according to the matching image feature point information. Assuming that there are matching image feature points f near the missing area m And the corresponding point cloud feature point p n , with p n Based on the neighborhood information of the image feature points, new point cloud points are generated. If f m The neighborhood of has k feature points {f m,j }(j=1,2,...,k), and they are related to f m The relative position relationship of is known, then in the point cloud space, p n As the center, generate k new point cloud points {p new,j }, the coordinates of the new point can be calculated by the following formula: new,j =p n +λ(f m,j -f m ), where λ is the scaling factor, which is determined according to the spatial proportional relationship between the image and the point cloud.

8. A time-interval timing classification system for point cloud data of a distribution line according to claim 2, characterized in that: When optimizing the point cloud data, the least squares method is used to optimize the position of the point cloud. Suppose there are n points {p i}(i = 1, 2,..., n) in the point cloud. The goal is to make the point cloud match the image features better. Define an objective function E: where m is the number of matched image feature points, l is the number of point cloud points in the neighborhood of point p i , S ij is the similarity between point p i and the matched image feature points, (p i - p i,k ) 2 is the square of the distance between point p i and its neighborhood point p i,k ; Take the partial derivative of the objective function E with respect to the coordinates of the point cloud points and set the partial derivative to 0 to obtain a system of linear equations. By solving this system of linear equations, update the coordinates of the point cloud points to achieve the optimization of the point cloud.

Images