A power transmission corridor tree barrier identification method and device based on SIFT feature points and a medium

Abstract

The application discloses a power transmission corridor tree barrier identification method and device based on SIFT feature points, and a medium, and belongs to the field of image data identification. The method comprises the following steps: shooting image data of different types of trees in a power transmission line corridor area, wherein the image data comprises an RGB image and a multispectral image; pre-processing the collected image, extracting SIFT feature points in the image, evaluating and screening the contrast and stability of key feature points, generating feature descriptors according to the neighborhood information of each feature point, and performing feature matching; matching the SIFT feature points in the image with pre-stored tree feature descriptors in a database based on a nearest neighbor matching method of Lowe ratio test, performing classification and identification of the trees, judging whether the current trees constitute a tree barrier object, and evaluating whether the trees will affect the power transmission line through a detection rate. The application has high-precision tree barrier identification capability, low misjudgment rate and missed detection rate, and high-efficiency tree barrier identification and processing.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of image data recognition, and particularly relates to a power transmission corridor tree barrier recognition method, device and medium based on SIFT feature points. BACKGROUND

[0002] With the development of the power industry, the safe operation of power transmission lines has attracted more and more attention. The tree barrier problem in the power transmission corridor, especially when the trees grow too high or too dense, can cause the shutdown, damage or even safety accidents of the power transmission line. How to efficiently and accurately identify and monitor tree barriers in a wide power transmission corridor has become an important task in the operation and maintenance of the power system. Traditional tree barrier identification methods rely on manual inspection, although this method has achieved certain results in the early stage, but due to its high labor intensity, limited coverage, low efficiency and easy to be affected by human factors, it gradually exposes many problems.

[0003] Many technologies currently try to use image processing, LiDAR point cloud data and deep learning to improve the automation and accuracy of tree barrier identification. Image-based tree barrier detection methods usually rely on edge detection, image segmentation, color space transformation and other traditional image processing techniques. However, these methods often perform poorly when faced with complex backgrounds and low-quality images, and have low recognition accuracy and robustness. With the rise of deep learning, techniques using convolutional neural networks (CNN), object detection algorithms (such as YOLO, Faster R-CNN) for tree barrier identification have gradually emerged. Although deep learning methods perform well in image processing, their shortcomings are also obvious, mainly in the strong dependence on labeled data, high demand for computing resources, long training time and other issues. Especially in practical applications, the labeling work of many data sets is very tedious and expensive, and deep learning methods often require a large number of samples to achieve good recognition results, so in some data-lacking fields, the application of this method is limited.

[0004] LiDAR point cloud technology has also been widely used in recent years, especially in the precise positioning of tree barrier objects in three-dimensional space. LiDAR point cloud can provide detailed spatial data to help the system accurately identify the location and height of trees and obstacles. However, the cost of LiDAR technology is relatively high, and its data processing is complex, especially in the processing of high-density point cloud data, the accuracy of segmentation and classification is often limited. Moreover, the deployment and maintenance cost of LiDAR equipment is large, which is usually suitable for large-scale and long-term projects, but it is not economical for short-term and small-scale inspection tasks.

[0005] In view of the above technical deficiencies, the present technology proposes a tree barrier recognition method based on SIFT (Scale-Invariant Feature Transform) feature points, aiming to improve the accuracy and robustness of tree barrier recognition through classical feature point extraction techniques in image processing. SIFT algorithm has been widely applied in image recognition field, which extracts key feature points in image and realizes target detection and recognition through matching these feature points. Unlike traditional image segmentation methods, SIFT algorithm can effectively handle image rotation, scale change and partial occlusion problems, making it perform well in complex background and environment. At the same time, SIFT feature extraction has lower requirements for image quality, so it can obtain good recognition effect in relatively ordinary images.

[0006] The present technology combines the practical needs of SIFT feature point extraction and tree barrier recognition, and in the detection of tree barriers in power transmission corridors, SIFT algorithm is used to extract representative feature points from images, and matching algorithm is used to identify the position of trees or other obstacles. Compared with traditional image processing methods, SIFT feature extraction method can better cope with rotation, scale change and complex background in image, so as to improve the accuracy and stability of tree barrier detection. SIFT algorithm does not rely on large-scale labeled data set required by deep learning, so it has lower data requirement and is suitable for application in environment with difficult data acquisition.

[0007] This tree barrier recognition method based on SIFT feature points has obvious advantages compared with deep learning and LiDAR methods. It requires lower computing resources and can realize real-time processing on ordinary hardware, suitable for low-cost, fast deployment application scenarios. SIFT feature point method has lower requirements for image quality and can stably extract key features in low-resolution images, so it is particularly suitable for large-scale inspection tasks. SIFT feature point extraction algorithm has strong real-time performance and can realize fast recognition, meeting the demand for real-time and high efficiency in inspection. SUMMARY

[0008] In view of the above existing problems, the present technology aims to fill the gap in the prior art and provide a solution for efficiently and accurately identifying tree barriers in power transmission corridors in low-resource environments. By combining SIFT feature point extraction technology and tree barrier recognition requirements, this method can operate stably in complex background and occlusion conditions, and has strong application prospect.

[0009] To solve the above technical problems, a tree barrier recognition method based on SIFT feature points in power transmission corridors is proposed, which includes,

[0010] The high-definition RGB camera is carried by the unmanned aerial vehicle to take aerial photographs, image data of different types of trees in the power transmission line corridor area is taken, different growth stages and perspectives are covered, and the image data includes RGB images and multispectral images; the collected images are preprocessed, the median filter is used to remove noise generated under low light conditions, and the histogram equalization is used for image enhancement; SIFT feature points are extracted from the images, key feature points are extracted under different scales and rotation angles through the scale invariance characteristic, the contrast and stability of the key feature points are evaluated and screened, the feature descriptors are generated according to the neighborhood information of each feature point, and feature matching is performed; the nearest neighbor matching method based on the Lowe ratio test is used to match the SIFT feature points in the images and the pre-stored tree feature descriptors in the database, the classification and identification of the trees are performed, and it is determined whether each tree is a tree barrier; the spatial distribution of the trees is analyzed through the clustering algorithm, the tree area affecting the power transmission line is extracted, the distance between the spatial coordinates of the trees and the power transmission line is combined, and the geometric constraint model is used to determine whether the current trees constitute a tree barrier object; the performance is evaluated by using the evaluation index, and whether the trees will affect the power transmission line is further evaluated by the detection rate.

[0011] As a preferred scheme of the SIFT feature point-based power transmission corridor tree barrier identification method, the image data includes image data collected in the power transmission corridor area of different types of growing trees by the unmanned aerial vehicle platform with high-resolution shooting capability and the camera equipment carried by the unmanned aerial vehicle platform; the camera equipment adopts a multispectral camera; and the image data is collected simultaneously.

[0012] The constraint conditions when collecting the images are set as follows: condition one, the climate conditions and light changes are considered, the climate conditions include sunny days and cloudy days, and the light changes include morning, noon and evening; condition two, the flight height of the unmanned aerial vehicle is set as 80 meters according to the average height of the trees and the width of the power transmission corridor; condition three, the overlap degree between adjacent flight lines is set as 70% during the flight process; and condition four, one image is shot per second, and the flight line angle is adjusted every 30 meters during the flight process.

[0013] The tree objects in each image are labeled by means of artificial labeling, including the types, positions and sizes of the trees, the image labeling tool LabelImg is used for labeling, each tree object is framed as a rectangular region, and the corresponding tree types are labeled; during the image labeling process, the height of the tree and the crown width of the tree based on the pixel ratio of the measuring tool and the image are recorded for each labeled tree object, wherein the height of the tree is measured by using the RTK positioning system carried by the unmanned aerial vehicle, and the actual height of each tree is calculated in combination with the height data during the flight process of the unmanned aerial vehicle.

[0014] The flight route is planned in advance using flight planning software, the flight height, route spacing and image shooting frequency are input in the software, the unmanned aerial vehicle is automatically flown, and the image is automatically shot according to the set parameters, after the flight is started, the camera shoots the image according to the flight planning and the set frequency, after each flight is finished, the image is uploaded to the storage server using the data transmission equipment, and the shooting angle and image brightness of the camera are monitored in real time during the flight process;

[0015] All the collected image data is transmitted to the local server and stored in a classified manner, meanwhile, the data redundancy backup is carried out using the RAID storage array, all the image data is classified according to the tree species, illumination condition and shooting time, the image index is established, and the image data quality is checked by visual inspection after the experimental data collection is completed.

[0016] As a preferred scheme of the power transmission corridor tree barrier recognition method based on SIFT feature points, the pre-processing of the collected image comprises,

[0017] The collected image data is input, a 3x3 and 5x5 window median filter is selected for image denoising, the median value of all pixels in the neighborhood is calculated for each pixel in the image data, and the original pixel value is replaced with the median value, and the step is repeated until the noise in the image is suppressed;

[0018] The low-contrast image is enhanced by histogram equalization to ensure the distinguishability of the tree profile in the complex background, the gray level histogram of the input image data is calculated, and the cumulative distribution function of the image is calculated , the gray value range of the image is set as , wherein L is the gray level:

[0019] ;

[0020] , wherein is the number of pixels of the gray value i, is the total number of pixels of the image, i is a variable index, and k is the number of gray values; each pixel value of the image is remapped using the cumulative distribution function to obtain an enhanced image: each pixel value is mapped to a new value:

[0021] ;

[0022] The local contrast limited adaptive histogram equalization method is used for histogram equalization in a local area, the image is divided into a plurality of 8x8 or 16x16 pixel local areas, histogram equalization is carried out on each area respectively, and the contrast growth is limited, all the areas are combined to obtain an enhanced overall image, wherein the contrast limit is:

[0023] ;

[0024] wherein, is the upper limit of the control contrast, is the maximum pixel value of the image in the region;

[0025] Different perspectives and postures of the collected images are standardized by geometric transformation, and for each image, random rotation and scaling operations are performed by an affine transformation matrix.

[0026] As a preferred scheme of the SIFT feature point-based power transmission corridor tree barrier recognition method, the SIFT feature points in the image are extracted by using a SIFT feature point extraction algorithm to extract robust feature points from the image after image preprocessing:

[0027] The image is subjected to multi-scale Gaussian blur processing to generate a blurred image and construct a scale space;

[0028] In the scale space, potential feature points are found by calculating the local extreme values of the image;

[0029] The feature points are accurately positioned using Gaussian difference;

[0030] A direction is assigned to each feature point according to the gradient direction of the image, and the feature point direction assignment is performed;

[0031] Based on the neighborhood information of each feature point, a 128-dimensional feature descriptor is generated.

[0032] As a preferred scheme of the SIFT feature point-based power transmission corridor tree barrier recognition method, the feature matching includes comparing the feature points in the newly collected image with the calibrated images in the database by a nearest neighbor matching method based on Lowe's ratio test, matching the feature descriptors generated by each feature point with the descriptors in the database, and judging the reliability of the matching by the ratio of the minimum matching distance to the second smallest matching distance;

[0033] Specifically, for each feature point in the new image, the corresponding generated feature descriptor For all tree images in the database, the descriptors of all feature points are calculated, and the Euclidean distance between the descriptors and the feature descriptors in the new image is calculated:

[0034] ;

[0035] wherein, and are the descriptors​ and the value of the i-th dimension;

[0036] selecting the two closest descriptors and wherein, is the minimum matching distance, is the second minimum matching distance;

[0037] According to and performing Lowe ratio test, i.e. filtering out false matches by comparing the ratio of and calculating the ratio :

[0038] ;

[0039] Setting the matching threshold τ = 0.7, when (set threshold τ is usually 0.7), it is considered that the current feature point matching is valid; when , it is considered that the current feature point matching is invalid;

[0040] According to the matching result, the tree classification is carried out by a machine learning classification algorithm, the feature points with valid matching are assigned a label, the label is the tree category of the corresponding image, and the trees with assigned labels form a calibration data set.

[0041] As a preferred scheme of the SIFT feature point-based power transmission corridor tree barrier identification method, wherein: the classification and identification comprises, based on the result of feature matching, analysis of the spatial distribution of trees, extraction of spatial coordinates of each tree , clustering the positions of the trees by a clustering algorithm, the position of each tree will be input data of the K-means algorithm, and the tree area threatening the power transmission line in the target image is identified,

[0042] Setting the number of clusters K, i.e. dividing the tree area into different groups according to the preset number of clusters, randomly initializing the centroids of the K clusters , each centroid is a two-dimensional coordinate , and is the two-dimensional coordinate of the k-th centroid;

[0043] Calculating the Euclidean distance of the tree points to the centroids of the clusters, for the i-th tree point, the distance between it and the k-th centroid is:

[0044] ;

[0045] The tree points are assigned to the cluster to which the nearest centroid belongs, and for the kth cluster, the new centroid is determined by the mean of all tree points in the cluster:

[0046] ;

[0047] wherein, and is the spatial coordinate of the ith tree, is the number of tree points in the cluster, until the center of the cluster does not change or a preset maximum number of iterations is reached, after the clustering is completed, the tree density and distribution in each cluster are analyzed, and the area affecting the power transmission line is identified, when the tree density in each cluster is greater than 110 trees per hectare, the current area is considered to be a tree barrier area;

[0048] For extracting the spatial coordinates of each tree, geometric constraints and classification judgments are performed to determine whether the trees constitute a tree barrier object, and the judgment rules include the horizontal distance of the trees from the power transmission line, the tree height, and the tree width

[0049] The horizontal distance of the trees from the power transmission line includes calculating the horizontal distance between the trees and the power transmission line, and according to the spatial position of the trees:

[0050] ;

[0051] wherein D is the horizontal distance of the trees from the power transmission line, is the spatial coordinate of the power transmission line;

[0052] The tree height, when the height of the tree exceeds a set threshold , and the horizontal distance of the power transmission line is less than a threshold , the current tree is determined to be a tree barrier object;

[0053] The width of the tree in the image is calculated through contour detection of the image When the tree width is greater than a set threshold , and the horizontal distance of the power transmission line is less than a threshold , the current tree is determined to be a tree barrier.

[0054] As a preferred scheme of the SIFT feature point-based power transmission corridor tree barrier identification method, the detection rate includes using the accuracy, recall, precision, and F1 value evaluation index evaluation method to evaluate the performance;

[0055] The tree barrier detection rate is further used to evaluate the performance of different tree species, and to determine whether the trees will affect the power transmission line;

[0056] wherein the tree barrier detection rate is calculated by comparing the detection result with the manually labeled tree barrier region:

[0057] ;

[0058] When the tree barrier detection rate is greater than 90%, it is considered that the current tree species performs well in the current identification, the identification is accurate, and it is considered that the current identification result can be adopted, wherein, is the number of regions correctly identified as tree barriers, is the total tree barrier region.

[0059] A computer device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the steps of the SIFT feature point-based power transmission corridor tree barrier identification method when executing the computer program.

[0060] A computer-readable storage medium having a computer program stored thereon, characterized in that the computer program is executed by a processor to implement the steps of the SIFT feature point-based power transmission corridor tree barrier identification method.

[0061] Advantages of the present application: (1) The present application uses scale-invariant feature transform (SIFT) technology for feature point extraction, describes the key feature points in the tree image, and ensures that the unique features of the tree can be stably extracted under different viewing angles and different lighting conditions. The scale and rotation invariance of SIFT makes the tree recognition more robust, effectively improving the accuracy of tree barrier identification.

[0062] Through the experimental test of various trees by the present application, the recognition accuracy and tree barrier detection rate are excellent. Especially in the face of trees in different growth stages and different environments, the present application can effectively reduce misjudgment and omission, and ensure the accuracy of the identified trees as potential tree barriers.

[0063] (2) Through the Lowe ratio test and the nearest neighbor matching algorithm, the probability of false matching is greatly reduced, thereby effectively reducing the false tree barrier identification result. This optimization strategy compares the matching quality of each feature point, and only when the distance ratio of the two closest matching points meets the set threshold, the matching is considered effective, further ensuring the identification accuracy and reliability. In the method, the geometric constraint relationship between the tree and the power transmission line is combined, and by setting the distance threshold and the tree height and width limit, the problems of omission and misjudgment are effectively avoided. Even for trees in different growth stages, the judgment of whether they are tree barriers can be accurate and accurate, reducing the common omission phenomenon in traditional methods.

[0064] (3) Based on computer vision and image processing algorithms, it can automatically identify whether the trees are tree barriers, reducing the workload and time cost of manual inspection. In large-scale power transmission corridor inspection, it can quickly and accurately process a large amount of image data, greatly improving work efficiency. Through unmanned aerial vehicle or satellite remote sensing technology to obtain image data, the invention can realize real-time image preprocessing, feature extraction and matching, tree classification and tree barrier judgment, greatly improving the real-time monitoring and rapid response ability, especially suitable for rapid inspection and abnormal detection of power transmission lines.

[0065] (4) The invention can effectively identify different tree species, not only suitable for different tree species, but also cope with tree identification under different growth environments and climate conditions. In practical application, different types of trees and tree growth environments may have different effects on tree barrier formation, and the use of the invention can ensure high universality.

[0066] The scale invariance and rotation invariance of SIFT feature points make the invention still able to stably extract effective features under different lighting conditions and different viewing angles. Especially in complex climate conditions or mountainous areas, it can better adapt to different environmental challenges.

[0067] (5) Not only can it process RGB images, but also can analyze combined with depth image data, enhancing the accuracy and stability of tree barrier detection. For occluded or low-contrast tree regions, the additional spatial information provided by the depth image helps improve the accuracy of identification. By integrating SIFT features and machine learning algorithms (such as support vector machine SVM, k- nearest neighbor KNN, etc.), the method can automatically classify and judge trees, and combine the spatial distribution of trees with the geometric relationship of power transmission lines to intelligently determine which trees are tree barrier objects, further improving the intelligent level of tree barrier management.

[0068] (6) The automatic tree barrier identification system greatly reduces the demand and cost of manual inspection, and through unmanned aerial vehicle or other automatic equipment for image acquisition and using the invention for tree barrier identification, not only improves the inspection efficiency, but also saves a lot of labor and time cost. Through efficient and accurate identification of tree barriers, it can effectively prevent trees from affecting the safety of power transmission lines due to excessive height or excessive density, reduce power outages caused by tree barriers, and thus ensure the stable operation of the power system and improve the economic benefits of power companies.

[0069] (7) The application is not only suitable for power transmission corridor tree barrier identification, but also can be applied to other fields such as vegetation monitoring in agriculture and forestry. Due to the generality of the feature extraction and identification method, the tree classification model can be adjusted according to different application scenarios in the future to improve the applicability. The application can also be combined with other technologies such as unmanned aerial vehicle flight control system, three-dimensional modeling, map information system, etc., to further expand its application range. The tree barrier identification result can be combined with the geographic information of the power transmission line by combining GIS (Geographic Information System) technology, to realize more accurate positioning and management. BRIEF DESCRIPTION OF DRAWINGS

[0070] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0071] Figure 1 The overall flowchart of a power transmission corridor tree barrier identification method based on SIFT feature points provided by an embodiment of the present application.

[0072] Figure 2 The SIFT feature point extraction principle diagram of a power transmission corridor tree barrier identification method based on SIFT feature points provided by an embodiment of the present application.

[0073] Figure 3 The image matching algorithm flowchart of a power transmission corridor tree barrier identification method based on SIFT feature points provided by an embodiment of the present application.

[0074] Figure 4 The geometric constraint and classification judgment architecture diagram of a power transmission corridor tree barrier identification method based on SIFT feature points provided by an embodiment of the present application.

[0075] Figure 5 The multi-dimensional feature extraction module structure diagram of a power transmission corridor tree barrier identification method based on SIFT feature points provided by an embodiment of the present application. DETAILED DESCRIPTION

[0076] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.

[0077] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. However, it will be apparent to one skilled in the art that the present application can be practiced without the specific details set forth in this description, that the present application can be practiced with other than the described embodiments, and that the present application can be practiced with other than the described embodiments, and that the present application can be practiced with other than the described embodiments. Therefore, the scope of the present application is indicated by the appended claims rather than the foregoing description, and all changes that come within the meaning and range of equivalents are intended to be embraced therein.

[0078] Embodiment 1, Reference Figures 1-5 As a first embodiment of the present application, the embodiment provides a power transmission corridor tree barrier recognition method based on SIFT feature points, comprising:

[0079] S1: Through an unmanned aerial vehicle carrying a high-definition RGB camera for aerial photography, image data of different types of trees in the power transmission line corridor area is photographed, covering different growth stages and viewing angles. The image data includes RGB images and multispectral images.

[0080] It should be noted that through the unmanned aerial vehicle platform with high resolution shooting capability and the camera equipment carried:

[0081] The unmanned aerial vehicle platform uses DJI Matrice 300 RTK unmanned aerial vehicle, which has a precise positioning system and can maintain stable flight in complex environments;

[0082] The camera equipment uses DJI Zenmuse H20T multispectral camera, which has 20MP high-definition RGB image shooting capability, and also supports near-infrared (NIR) image acquisition, which is convenient for subsequent multispectral image processing;

[0083] The device can simultaneously collect RGB images and multispectral images, has high resolution, and is suitable for a wide range of applications, especially for tree barrier monitoring in large areas such as power transmission corridors.

[0084] Set the constraints when collecting images:

[0085] Condition one, considering the climate conditions and light changes, the climate conditions include sunny days, cloudy days, and the light changes include morning, noon, and evening.

[0086] Condition two, according to the average height of the trees and the width of the power transmission corridor, the flight height of the unmanned aerial vehicle is set to 80 meters; this flight height can ensure that the image collection covers the range, and at the same time, it will not cause the trees to be blocked due to the height being too low, and it is sufficient to ensure that the overall shape of the trees can be clearly presented in the image.

[0087] Condition three, in order to obtain the complete information of the trees and ensure that the image has good overlap, the overlap between adjacent flight lines is set to 70% during the flight process.

[0088] Condition four, set every second to take a picture, and in the process of flight every 30 meters to adjust the flight path angle, so as to cover the tree information under different angles and fields of view.

[0089] By artificial labeling, the tree objects in each image are labeled, including the species, position and size of the tree, and the image labeling tool LabelImg is used for labeling. Each tree object is framed as a rectangular region, and the corresponding tree species is labeled. During the image labeling process, for each labeled tree object, the tree height and crown width based on the pixel ratio of the measuring tool and the image are recorded. The RTK positioning system carried by the unmanned aerial vehicle is used to measure the tree height, and the actual height of each tree is calculated by combining the height data during the flight of the unmanned aerial vehicle.

[0090] The flight planning software is used to plan the route in advance. After inputting the flight height, route spacing and image shooting frequency in the software, the unmanned aerial vehicle automatically flies and automatically shoots images according to the set parameters. After the flight starts, the camera shoots images according to the flight planning and the set frequency. Each image is saved as a JPEG format with a resolution of 20MP and an image size of about 5MB.

[0091] After each flight, the image is uploaded to the storage server using the data transmission device. The shooting angle and image brightness of the camera are monitored in real time during the flight.

[0092] All collected image data is transmitted to the local server and stored in categories. RAID storage array is used for data redundancy backup. All image data is classified according to tree species, lighting conditions and shooting time to establish an image index. After the experimental data collection is completed, the image data quality is checked visually to ensure that the images are clear, without obvious distortion, excessive exposure or shadow problems. For images with poor quality due to weather, flight jitter or other factors, appropriate screening is performed to ensure the high quality of the final data set.

[0093] S2: Preprocess the collected images. Use median filtering to remove noise produced under low light conditions, and use histogram equalization for image enhancement.

[0094] It should be noted that the collected image data is input, and a 3x3 and 5x5 window median filter is selected for image denoising. For each pixel in the image data, the median value of all pixels in the neighborhood is calculated, and the original pixel value is replaced with the median value. Repeat the steps until the noise in the image is suppressed. Let the pixel value of a point in the image be , the neighborhood of which is a k x k window, and the pixel values in the window are , where and , the replaced value For:

[0095] ;

[0096] wherein, represents the median of the pixel values in the neighborhood.

[0097] It should also be noted that the purpose of image enhancement is to improve the contrast and brightness of the image, so that the tree features in the image are more prominent, which helps the subsequent extraction of feature points. For tree images, especially in poor lighting or complex background conditions, enhancement processing is particularly important.

[0098] Specifically, the low-contrast image is enhanced by histogram equalization to ensure the recognizability of the tree contour in the complex background. The gray level histogram of the input image data is calculated, and the cumulative distribution function of the image is calculated , the gray value range of the image is set as , wherein L is the gray level:

[0099] ;

[0100] wherein, is the number of pixels with gray value i, is the total number of pixels of the image, i is the variable index, and k is the number of gray values; the cumulative distribution function is used to re-map each pixel value of the image to obtain the enhanced image: each pixel value is mapped to a new value :

[0101] ;

[0102] Further, using the local contrast limited adaptive histogram equalization method, in some complex background, directly using global histogram equalization may cause over-enhancement or over-exposure, resulting in loss of details. Therefore, the CLAHE method is also used to perform histogram equalization in the local region to control the problem of over-enhancement.

[0103] Specifically, the image is divided into multiple local regions of 8x8 or 16x16 pixels, and histogram equalization is performed on each region, and the contrast is limited to control the growth of the contrast, and all regions are combined to obtain the enhanced overall image, wherein the limited contrast is:

[0104] ;

[0105] wherein, is the upper limit of the control contrast, is the maximum pixel value of the image in the region;

[0106] Further, different perspectives and poses of the collected images are standardized by geometric transformation, and for each image, random rotation and scaling operations are performed by an affine transformation matrix;

[0107] Specifically, the rotation transformation formula is as follows:

[0108] ;

[0109] where (x, y) is the coordinate in the original image, θ is the rotation angle, and (x', y') is the coordinate after rotation.

[0110] Scaling transformation formula:

[0111] ;

[0112] where, and are the scaling factors along the x and y axes; through the above denoising and enhancement processing, the image quality can be significantly improved. For images under different lighting conditions (such as overcast and evening), the brightness and contrast of the preprocessed images are more balanced, and the outlines of trees are clearer. Especially in the case of complex background or tree partial occlusion, the detailed information in the image is effectively enhanced. The processed image provides more accurate input for subsequent SIFT feature point extraction.

[0113] Further, after the image preprocessing is completed, the SIFT feature point extraction algorithm is used to extract robust feature points from the image:

[0114] ① Construct a scale space:

[0115] To achieve scale invariance of features, first, the original image is subjected to layer-by-layer Gaussian blur processing and forms a multi-scale pyramid structure. Specifically, different standard deviations σ of the Gaussian kernel are applied to the input image to generate a sequence of progressively blurred images from low to high scales; each scale group contains several Gaussian blurred images, and the σ between adjacent scales increases by a fixed multiple. The complete scale space is formed by concatenating multiple scale groups, enabling feature points to be stably detected at different spatial resolutions.

[0116] ②Detect scale space extreme points:

[0117] In the Gaussian scale space, by comparing each pixel point in its 26-neighborhood at its scale and adjacent upper and lower scales, the local maximum or minimum of the DoG (Difference of Gaussian) response is found as a potential feature point.

[0118] ③Feature point precise positioning:

[0119] ​The Taylor expansion is used to optimize the sub-pixel level position and scale of the candidate extreme points, the points with low contrast or located in unstable edge structure are removed, and only the key feature points with high stability are reserved.

[0120] IV. Feature point direction assignment:

[0121] The direction histogram is constructed according to the gradient amplitude and direction information of the feature point neighborhood, and the main direction is selected as the main direction of the feature point, so that the rotation invariance is realized.

[0122] V. Generating feature descriptors:

[0123] A plurality of sub-regions are divided with the feature point as the center, the gradient direction histogram is counted in each region, and finally the 128-dimensional SIFT descriptor is spliced to form a 128-dimensional SIFT descriptor, which is used for subsequent matching and classification recognition.

[0124] The Gaussian blur processing includes Gaussian blur processing of the original image with different standard deviations to generate images of different scales, and applying a Gaussian kernel function to each layer of image:

[0125] ;

[0126] Wherein, sigma is the standard deviation of blur, which controls the degree of blur; for each image of different scales, a set of blurred images is constructed using different scales (such as from sigma = 1 to sigma = 10), each set of images as a pyramid level, and the scale gradually increases;

[0127] The Gaussian difference includes that according to the generated blurred images, SIFT highlights the local features of the image by calculating the difference between adjacent scales, and the difference between two images of adjacent scales is calculated For:

[0128] ;

[0129] Wherein, And The Gaussian kernel function of the two images of adjacent scales; the Gaussian difference image is obtained, and the extreme point search is performed on the Gaussian difference image, that is, the local maximum and minimum are found in the 26 neighborhood pixels around each pixel, and the current local extreme point is the potential feature point;

[0130] The positioning feature point includes correcting the position of the feature point by Taylor expansion, optimizing the coordinates of each feature point, and deleting the feature points with low contrast and unstable edges;

[0131] The feature point direction assignment includes that the SIFT algorithm determines the direction of each feature point based on the gradient amplitude and gradient direction of its neighborhood pixels, first calculates the gradient according to the pixel intensity difference, and the calculation formula of the gradient is as follows:

[0132] ;

[0133] wherein, and are the gradients of the image in the x and y directions respectively; , , , are the intensities of the four most immediate neighboring pixels of the pixel respectively, , are calculated by comparing the intensities of the left and right neighbors of the pixel respectively; , are calculated by comparing the intensities of the upper and lower neighbors of the pixel respectively;

[0134] The present application employs a two-variable arctangent function to calculate the true gradient direction of each pixel ;

[0135] The present application can return an absolute direction angle within the range of (-pi, pi], accurately distinguish the four quadrants, and thus ensure the uniqueness and accuracy of the direction of the feature point;

[0136] Subsequently, the gradient directions of all the pixels in the neighborhood of the feature point are accumulated into a direction histogram at a set angular interval, and the direction with the largest amplitude in the histogram is selected as the main direction of the feature point; when there are secondary peaks with close amplitudes, multiple directions can be additionally generated to enhance the rotational invariance of the feature.

[0137] The feature descriptor comprises: after determining the direction of the feature point, SIFT generates a feature descriptor according to the neighborhood information of each feature point, divides the neighborhood of the feature point into 16 sub-regions, and the size of each sub-region is 16x16 pixels; in each sub-region, the amplitude and direction of the gradient are calculated, and an 8-direction histogram is generated; and the 8-direction information of all 16 sub-regions is spliced into a 128-dimensional descriptor in order , is the gradient direction histogram of the i-th sub-region;

[0138] In order to ensure the robustness of the feature descriptor to illumination changes and noise, each descriptor is subjected to L2 normalization, so as to avoid the influence of the changes in the brightness of some regions on the accuracy of matching.

[0139] Through the above SIFT feature point extraction, stable and reliable feature points can be obtained from each tree image, and generally hundreds of feature points are extracted; for example, 200 feature points can be extracted from an eucalyptus image, and 250 feature points can be extracted from a pine image. The number of feature points is related to the morphology of the tree, the image resolution and the shooting angle.

[0140] The feature points are usually distributed in the contour of the tree, the junction of branches and leaves, and other structural rich areas. For example, the feature points of eucalyptus are mostly concentrated around the crown, and the feature points of Chinese arborvitae are mostly concentrated at the junction of the trunk and branches. These feature points are of great significance for subsequent tree category identification, position estimation, and modeling of the relative position relationship between the tree and the power transmission line, and can realize effective identification of the tree barrier in the power transmission corridor.

[0141] In summary, the present application adopts multi-source data fusion, not only processes standard RGB images, but also analyzes multi-dimensional data such as depth images. In the depth image, the depth information of each pixel point can provide additional support for the spatial relationship between the tree and the power transmission line, thereby improving the identification accuracy of the tree. Through multi-source data fusion, the present application can overcome the influence of factors such as light and shielding, and better detect the tree barrier.

[0142] The existing tree barrier identification method generally only relies on a single type of image data (such as an RGB image), and ignores depth information. By combining additional information such as a depth image, the present application can better handle complex scenes, especially areas with characteristics such as shielding and dense trees.

[0143] S3: Extract SIFT feature points in the image, extract key feature points at different scales and rotation angles through the scale invariance property, evaluate and screen the contrast and stability of the key feature points, generate feature descriptors according to the neighborhood information of each feature point, and perform feature matching.

[0144] The evaluation and screening include: ① Contrast evaluation: calculate the local response value of the candidate feature point in the DoG (Gaussian difference) space, if the response amplitude is lower than the preset threshold, it is considered that the point is in a low-contrast area (such as a weak texture or noise area), such points are prone to failure in matching, and are therefore removed.

[0145] ② Stability screening (edge response suppression): to avoid retaining unstable feature points located on the strip-shaped edge, the principal curvature ratio H of the Hessian matrix is used for judgment:

[0146] ;

[0147] Wherein, is the second-order partial derivative of the image in the x direction, reflecting the curvature of the feature point in the x direction; is the second-order partial derivative of the image in the y direction, reflecting the curvature of the feature point in the y direction; is the mixed second-order partial derivative of the image in the x and y directions, reflecting the cross curvature or rotational invariance of the feature point.

[0148] The trace and determinant are calculated, and when the principal curvature ratio exceeds a threshold value (i.e., obvious unidirectional edge structure is presented) or the determinant is negative, it is determined that the feature point is unstable and is deleted.

[0149] ③ Generating feature descriptors: for the feature points passed through the screening, local feature descriptors are constructed according to the amplitude and direction of the neighborhood gradient, which are used for subsequent feature matching.

[0150] Through the above contrast filtering and edge response suppression, feature points with significant structure, stable positioning and high distinguishability can be effectively retained, thereby improving the subsequent matching accuracy and the reliability of tree barrier recognition.

[0151] Further, when performing feature matching, the feature points in the newly collected image are compared with the calibrated images in the database. Through the nearest neighbor matching method based on Lowe's ratio test, the feature descriptor generated for each feature point is matched with the descriptors in the database, and the reliability of the matching is judged by the ratio of the minimum matching distance to the second minimum matching distance.

[0152] Specifically, for each feature point in the new image, the corresponding generated feature descriptor is calculated For all tree images in the database, the descriptors of all feature points are calculated, and the Euclidean distance with the feature points in the new image is calculated

[0153]

[0154] wherein, and are the values of the i-th dimension in the descriptors and

[0155] The two closest descriptors and are selected, wherein, is the minimum matching distance, is the second minimum matching distance.

[0156] According to and , Lowe's ratio test is performed, that is, by comparing the ratio of and , the wrong matching is filtered out, and the ratio is calculated.

[0157]

[0158] According to experience, the matching threshold τ=0.7 is set, and when (set threshold τ is usually 0.7), it is considered that the current feature point is matched effectively; when ​​​​, consider that the current feature point matching is invalid;

[0159] According to the matching result, the tree classification is carried out through a machine learning classification algorithm, a label is assigned to the effective feature point in the matching, the label is the tree category of the corresponding image, and the trees with the assigned label form a calibration data set.

[0160] In summary, the traditional tree barrier detection method depends on simple image difference analysis or traditional template matching algorithm, and these methods are easily disturbed by light changes, tree shape differences and view angle changes in complex environments, resulting in low recognition accuracy and high misjudgment rate.

[0161] The present application adopts SIFT (Scale Invariant Feature Transform) feature point extraction technology to realize efficient recognition of tree images, so that the shape of the tree can be stably recognized in different environments, greatly improving the reliability of detection. SIFT technology has scale invariance and rotation invariance, and can extract stable and high-discrimination feature points under different viewing angles and different lighting conditions. Compared with traditional image recognition methods, SIFT can avoid the influence of environmental changes and shooting angle changes on the recognition result, thereby improving the robustness of tree barrier recognition.

[0162] S4: nearest neighbor matching method based on Lowe ratio test, matching SIFT feature points in the image and pre-stored tree feature descriptors in the database, classifying and identifying the tree, and determining whether each tree is a tree barrier.

[0163] Specifically, when performing feature matching, the feature points in the newly collected image are compared with the calibrated images in the database, the feature descriptors generated by each feature point are matched with the descriptors in the database through the nearest neighbor matching method based on Lowe ratio test, and the reliability of the matching is judged by the ratio of the minimum matching distance to the second smallest matching distance.

[0164] Specifically, for each feature point in the new image, the corresponding generated feature descriptor is calculated For all tree image in the database, the descriptors of all feature points are calculated, and the Euclidean distance between the new image and

[0165] ;

[0166] Wherein, And The i-th value in the descriptor And

[0167] The two closest descriptors And , wherein, ​​the minimum matching distance, the second minimum matching distance;

[0168] According to and perform Lowe ratio test, that is, filter out false matches by comparing the ratio of and Calculate the ratio :

[0169] ;

[0170] Set the matching threshold τ = 0.7, when (set threshold τ is usually 0.7), it is considered that the current feature point matching is valid; when , it is considered that the current feature point matching is invalid;

[0171] According to the matching result, the tree is classified by a machine learning classification algorithm, the feature points with valid matching are assigned a label, the label is the tree category of the corresponding image, and the trees with assigned labels form a calibration dataset;

[0172] In an optional embodiment of the present application, the machine learning classification algorithm can be a support vector machine (SVM) for tree classification, specifically:

[0173] The 128-dimensional SIFT descriptor of each feature point is input into the SVM model as a feature vector;

[0174] According to the calibration dataset, the SVM classifier is trained to learn to classify tree species according to the SIFT feature descriptor;

[0175] After training, the SVM model is used to classify the feature points in the newly collected image and predict the tree category;

[0176] The optimization goal of the SVM model training is to maximize the interval between the two categories, so that the classification boundary has the smallest classification error rate, and the SVM optimization is:

[0177] ;

[0178] ;

[0179] where w is the normal vector of the hyperplane, b is the bias term, is the slack variable, is the training sample, is the class label.

[0180] In an optional embodiment of the present application, the machine learning classification algorithm can be a k-Nearest Neighbor algorithm (KNN) for tree classification, which classifies new samples based on their distance from all samples in the training set. Specifically:

[0181] For each feature point in the newly collected image, the KNN algorithm calculates the Euclidean distance between each feature point and each feature point in the training set, selects the K nearest training samples, counts their class labels, and determines the class of the current feature point based on the label. The majority class label is taken as the predicted class of the feature point:

[0182] ;

[0183] wherein, is the predicted class.

[0184] By using the nearest neighbor matching method of Lowe's ratio test, the feature points in the new image can be accurately matched with the feature points in the database, and the trees in the new image can be classified by Support Vector Machine (SVM) or k-Nearest Neighbor algorithm (KNN).

[0185] In summary, the prior art often relies on simple image features or color information in the process of tree classification, which lacks sufficient robustness and accuracy. By introducing the combination of SIFT feature matching and machine learning algorithm, the accuracy of tree classification can be significantly improved, and the trees in complex environment can be accurately distinguished.

[0186] In the process of tree recognition, a feature point matching method based on Lowe's ratio test and a machine learning algorithm (such as SVM or KNN) are used for tree classification. Lowe's ratio test analyzes the ratio of the smallest matching distance and the second smallest matching distance of each feature point, effectively filtering out incorrect matching points, thereby ensuring high accuracy of the matching result. For tree classification, a classification model can be trained to automatically identify trees based on their feature descriptors, thereby intelligently classifying trees as potential tree barriers or non-tree barriers.

[0187] S5: Perform spatial distribution analysis of the positions of the trees by clustering algorithm, extract the tree area that affects the power transmission line, and use geometric constraint model to determine whether the current tree constitutes a tree barrier object in combination with the distance between the spatial coordinates of the tree and the power transmission line.

[0188] Further, based on the results of feature matching, the spatial distribution of trees is analyzed, and the spatial coordinates of each tree are extracted The positions of the trees are clustered by clustering algorithm, and the position of each tree serves as input data for the K-means clustering algorithm to identify the tree area that poses a threat to the power transmission line in the target image.

[0189] Set the number of clusters K, i.e. divide the tree area into different groups according to the preset number of clusters, randomly initialize the centroids of K clusters Each centroid is a two-dimensional coordinate , , and is the two-dimensional coordinate of the kth centroid;

[0190] Calculate the Euclidean distance of the tree point to each cluster centroid, for the ith tree point, the distance of the ith tree point to the kth centroid is:

[0191] ;

[0192] Cluster assignment: assign the tree point to the cluster to which the nearest centroid belongs;

[0193] Centroid update: for the kth cluster, the new centroid is determined by the mean of all tree points in the cluster:

[0194] ;

[0195] wherein, and is the spatial coordinate of the ith tree, is the number of tree points in the cluster, until the cluster center no longer changes significantly or reaches the preset maximum iteration number, after clustering is completed, analyze the tree density and distribution in each cluster, identify the area affecting the power transmission line, when the tree density in each cluster is greater than 110 trees per hectare, the current area is considered as a tree barrier area;

[0196] For extracting the spatial coordinates of each tree, geometric constraints and classification judgments are performed to determine whether the tree constitutes a tree barrier object, the judgment rules include the horizontal distance of the tree from the power transmission line, the tree height, and the tree width

[0197] The horizontal distance of the tree from the power transmission line includes calculating the horizontal distance between the tree and the power transmission line, according to the spatial position of the tree:

[0198] ;

[0199] wherein D is the horizontal distance of the tree from the power transmission line, is the spatial coordinate of the power transmission line;

[0200] The tree height, when the height of the tree exceeds the set threshold , and the horizontal distance of the power transmission line is less than the threshold , the current tree is determined as a tree barrier object;

[0201] Calculate the width of the tree in the image by contour detection of the image When the tree width is greater than a set threshold , and the horizontal distance of the power line is less than the threshold , the current tree is determined to be a tree barrier, and the threshold is set empirically to 5-12 meters and automatically adjusted according to the voltage level of the project line, and the threshold is set empirically to 30-50px (pixels) to filter out meaningless small vegetation or noise targets.

[0202] Through the extraction of tree barrier area and geometric constraint judgment, it can accurately identify which trees may constitute a tree barrier and thus affect the safety of the power line. Based on the clustering analysis of spatial position and the constraint conditions of geometric model, combined with the height and width of the tree, the potentially dangerous trees can be effectively screened out, and the tree barrier identification is further judged.

[0203] In summary, the traditional tree barrier judgment method relies mainly on the visual features of the image (such as tree size, color, etc.), and these methods are easily affected by external factors (such as tree overlap, image blur, etc.). This method introduces a geometric constraint model, combines the spatial position of the tree with the relative position of the power line, and determines whether the tree is a tree barrier by setting a threshold. In particular, by using the spatial coordinates of the tree and the distance relationship of the power line, the horizontal distance D and the tree height are calculated, and if the distance is less than the set threshold and the height is greater than the threshold, the tree is determined to be a tree barrier.

[0204] The introduction of this geometric model enhances the accuracy of tree barrier judgment and avoids the shortcomings of relying solely on image features. The tree barrier judgment combined with geometric constraints not only considers the appearance features of the tree itself, but also considers the spatial relationship between the tree and the power line, making the tree barrier identification more accurate, especially suitable for complex terrain and multi-tree species environment.

[0205] S6: Evaluate performance using evaluation indicators, and further evaluate whether the tree will affect the power line by detection rate.

[0206] It should also be noted that the performance evaluation method using accuracy, recall, precision, and F1 value evaluation indicators: accuracy Acc=(TP+TN) / (TP+TN+FP+FN);

[0207] Recall Rec=TP / (TP+FN);

[0208] Precision Pre=TP / (TP+FP); ​​

[0209] F1 value F1-Score = 2 x [(Pre x Rec) / (Pre + Rec)];

[0210] Wherein: TP is the tree correctly identified as a tree barrier, TN is the tree correctly identified as a non-tree barrier, FP is the non-tree incorrectly identified as a tree barrier, and FN is the tree incorrectly identified as a non-tree barrier;

[0211] The performance of different tree species is further evaluated by the tree barrier detection rate to determine whether the tree will affect the power transmission line;

[0212] Wherein, the tree barrier detection rate is calculated by comparing the detection results with the manually labeled tree barrier area:

[0213] ;

[0214] When the tree barrier detection rate is greater than 90%, it is considered that the current tree species performs well in the current identification, the identification is accurate, and the current identification result can be used, wherein, is the number of areas correctly identified as a tree barrier, which refers to a continuous tree area (such as a tree crown range or a tree community area) in the manual annotation that is identified as a tree barrier and is also judged as a tree barrier in the system output. Each area is usually taken as a statistical unit in the form of a connected domain in an image, a tree clustering cluster, or a risk unit based on spatial coordinates; is the total tree barrier area, and the manual annotation personnel identify all tree areas that may affect the power transmission line according to the tree height, distance from the conductor, tree crown coverage range, etc. Each area corresponds to an independent tree barrier judgment unit, which can be obtained based on image connected domains, polygon annotation frames, or geographic space partitioning.

[0215] When the tree barrier detection rate is greater than 90%, it indicates that the tree species performs well in the current model identification, and the system can stably and accurately identify the risk area, and the identification result can be directly used for power transmission line tree barrier risk assessment.

[0216] Embodiment 2 is a second embodiment of the present application, which provides a power transmission corridor tree barrier identification method based on SIFT feature points. In order to verify the beneficial effects of the present application, scientific demonstration is carried out through experiments.

[0217] A typical power transmission corridor area is selected for the experiment, which is about 3000 meters long and about 50 meters wide. Different types of trees grow in the corridor, including eucalyptus, camphor tree, broussonetia papyrifera, blue oak, tung tree, Chinese arborvitae, pine tree, cedar and bamboo, etc. The number, distribution and tree species in the area are complex, which can better simulate the actual power transmission corridor environment.

[0218] To ensure the diversity of the dataset, nine typical tree species were selected, which have different growth characteristics and morphologies, including tall eucalyptus, pine, and fir trees, as well as shorter bamboo and cypress trees. Each tree species has different tree height, crown shape, and growth density, representing common tree types in power transmission corridors.

[0219] (1) The recognition results of the nine common tree species are shown in Table 1, including the accuracy, precision, recall, and F1-score of each tree species.

[0220] Table 1 Recognition results

[0221]

[0222] Result analysis:

[0223] ① From the experimental results, it can be seen that the method performs very well on different tree species, with an accuracy rate generally exceeding 90%. Among them, the recognition result of bamboo is the best, with an accuracy rate of 96.3% and an F1 value of 95.9%. This indicates that the method can well identify whether most tree species are tree barriers, and has good versatility.

[0224] ② Accuracy and precision: The accuracy and precision of almost all tree species are high, especially for eucalyptus, tung tree, and bamboo, with precision rates of 96.2%, 96.4%, and 97.1%, respectively. The identification of these trees can reduce misjudgment and improve the reliability of tree barrier detection in power transmission corridors.

[0225] ③ Recall rate and F1 value: The recall rate and F1 value are also excellent, indicating that the method not only effectively identifies whether a tree constitutes a tree barrier, but also minimizes missed detections. Especially for eucalyptus and tung tree, their recall rates are close to 90% or above, and their F1 values remain above 94%.

[0226] ④ Performance of specific tree species: For some trees, such as balsa and fir, although their precision and recall rates are slightly lower, their overall recognition ability is still reliable, which is closely related to the external characteristics and growth environment of different trees. For some highly similar tree species, the method may need more image data for training and adjustment to improve its classification accuracy.

[0227] (2) Analysis of tree barrier detection rate

[0228] The tree barrier detection rate results for the nine typical tree species are shown in Table 2:

[0229] Table 2 Detection rate results

[0230]

[0231] Result analysis:

[0232] ① High detection rate: From the experimental results, for oil camellia, bamboo, eucalyptus, pine and other tree species, the tree barrier detection rate is more than 90%, especially for bamboo, the tree barrier detection rate reaches 95%. This shows that the recognition method based on SIFT feature points performs very well on these tree species, and can accurately identify the area that may constitute a tree barrier.

[0233] ② Lower detection rate: For Chinese hackberry and fir trees, the tree barrier detection rate is relatively low, which is 78% and 85% respectively. This phenomenon may be related to the morphological characteristics and distribution density of these trees. The trunks and crowns of Chinese hackberry and fir trees are relatively slender and sparse, which may lead to inaccurate feature matching in the recognition process, thereby affecting the accurate identification of tree barriers.

[0234] ③ Influencing factors: The difference in tree barrier detection rate is related to multiple factors, mainly including the growth form of trees, the similarity between trees, the spatial distribution of trees, and the image quality, etc. For complex or similar tree species, the matching of SIFT features may be affected, resulting in a lower tree barrier detection rate.

[0235] This method shows high detection rate in most tree species, especially for tree species with thick trunks and obvious morphological characteristics, such as bamboo and oil camellia, the tree barrier detection rate is very high. For images taken under complex background, partially occluded trees or different lighting conditions, the system can still maintain high recognition accuracy, with a false negative rate of less than 5%, proving that this method has high reliability in dealing with these trees. It can achieve efficient tree barrier detection in most trees, especially in transmission line maintenance, which can effectively identify tree areas that may affect line safety, and improve the efficiency and accuracy of tree barrier management.

[0236] In summary, this embodiment shows that the tree barrier recognition method based on SIFT feature points has high accuracy and good robustness in transmission corridor tree identification and tree barrier detection. It has shown high accuracy and stability in tree barrier detection of most trees, especially under different tree species, different growth stages and different viewing angles. It can provide an effective tree barrier recognition scheme. Compared with traditional image processing methods and deep learning methods, this method has obvious advantages in data demand, computing efficiency and real-time performance. By integrating with other sensor data (such as LiDAR point cloud), the system can further improve the accuracy and stability of tree barrier detection, and can adapt to changing environments and conditions.

[0237] Embodiment 3, the third embodiment of the present application, which is different from the first two embodiments:

[0238] If the functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application or the part of the technical solutions that essentially contribute to the prior art or the part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0239] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered a list of executable instructions for implementing logic functions, and can be specifically embodied in any computer-readable medium for use by an instruction execution system, apparatus, or device, such as a computer-based system, a system including a processor, or other system that can fetch the instructions from the instruction execution system, apparatus, or device and execute the instructions, or in conjunction with these instructions execution systems, apparatuses, or devices. For the purpose of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport programs for use by an instruction execution system, apparatus, or device, or in conjunction with these instruction execution systems, apparatuses, or devices.

[0240] More specific examples (a non-exhaustive list) of the computer-readable medium include the following: an electrical connection having one or more wires (electrical devices), a portable computer diskette (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disc read-only memory (CD ROM). In addition, the computer readable medium can even be paper or other suitable medium on which the program can be printed, as the program can be electronically obtained, for example, by optical scanning of the paper or other medium, followed by editing, interpreting, or otherwise processing, if necessary, in other suitable ways to be electronically obtained, and then stored in the computer memory.

[0241] It should be understood that aspects of the application can be implemented in hardware, software, firmware or combinations thereof. In the embodiments described above, various steps or methods can be implemented, for example, by software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, and in another embodiment, any of the following technologies, known in the art, and combinations thereof, can be used: a discrete logic circuit having logic gates for implementing logic functions on data signals, an application specific integrated circuit having appropriate combinational logic gates, a programmable gate array (PGA), a field programmable gate array (FPGA), and / or the like.

Claims

1. A SIFT feature point-based power transmission corridor tree barrier identification method, characterized in that: The utility model relates to a kind of tree barrier detection method based on SIFT feature point, including, Including, Through the aerial photography of high-definition RGB camera carried by unmanned aerial vehicle, the image data of different kinds of trees in transmission line corridor area are photographed, covering different growth stages and visual angle, and the image data includes RGB image and multispectral image; The collected image is preprocessed, and the noise generated under low light condition is removed using median filter, and the image is enhanced by histogram equalization; SIFT feature points are extracted in the image, key feature points are extracted under different scales and rotation angles through scale invariance, the contrast and stability of key feature points are evaluated and screened, feature descriptors are generated according to the neighborhood information of each feature point, and feature matching is performed; Wherein, the SIFT feature points are extracted in the image after image preprocessing, and the feature points are extracted from the image using SIFT feature point extraction algorithm: The image is subjected to multi-scale Gaussian blur processing to generate a blurred image and construct a scale space; In the scale space, potential feature points are found by calculating the local extreme of the image; The feature points are accurately positioned using Gaussian difference; Each feature point is assigned a direction according to the gradient direction of the image, and the feature point direction is assigned a value; Based on the neighborhood information of each feature point, a 128-dimensional feature descriptor is generated; When performing feature matching, the feature points in the newly collected image are compared with the calibrated images in the database. The nearest neighbor matching method based on Lowe ratio test is used to match the feature descriptors generated by each feature point with the descriptors in the database, and the reliability of the match is determined by the ratio of the minimum matching distance to the second smallest matching distance. The nearest neighbor matching method based on Lowe ratio test is used to match the SIFT feature points in the image with the pre-stored tree feature descriptors in the database to classify and identify the trees, and determine whether each tree is a tree barrier. The spatial distribution of the trees is analyzed by clustering algorithm, and the tree area affecting the transmission line is extracted. The distance between the spatial coordinates of the trees and the transmission line is used to determine whether the current trees constitute a tree barrier object using a geometric constraint model.

2. The SIFT feature point-based power transmission corridor tree barrier identification method of claim 1, wherein: The performance is evaluated using evaluation indicators, and the impact of trees on the transmission line is evaluated by detection rate. The image data includes image data collected in the transmission corridor area of different tree species by an unmanned aerial vehicle platform with high-resolution shooting capability and a camera device mounted thereon: the unmanned aerial vehicle platform; the camera device uses a multispectral camera; image data of different tree species growing in the transmission corridor area is collected simultaneously; The constraint conditions for collecting images are set as follows: condition one, considering climate conditions and light changes, the climate conditions include sunny and cloudy days, and the light changes include morning, noon and evening; condition two, according to the average height of the trees and the width of the transmission corridor, the flight height of the unmanned aerial vehicle is set to 80 meters; condition three, the overlap degree between adjacent flight lines is set to 70% during flight; condition four, one image is taken per second, and the flight line angle is adjusted every 30 meters during flight; The tree objects in each image are labeled by manual labeling, including the species, position and size of the trees, and the image labeling tool LabelImg is used for labeling, each tree object is framed as a rectangular region, and the corresponding tree species is labeled; during the image labeling process, for each labeled tree object, the tree height and crown width based on the measurement tool and the pixel ratio of the image are recorded, wherein the RTK positioning system carried by the unmanned aerial vehicle is used to measure the tree height, and the actual height of each tree is calculated by combining the height data during the flight of the unmanned aerial vehicle; The flight route is planned in advance using flight planning software, after the flight height, route spacing and image shooting frequency are input into the software, the unmanned aerial vehicle automatically flies and automatically shoots images according to the set parameters, after the flight starts, the camera shoots images according to the flight planning and the set frequency, after each flight ends, the image is uploaded to the storage server using the data transmission device, and the shooting angle and image brightness of the camera are monitored in real time during the flight; All collected image data is transmitted to a local server and stored in categories, and a RAID storage array is used for data redundancy backup, all image data is classified according to tree species, lighting conditions and shooting time, and an image index is established, and after the experimental data collection is completed, the image data quality is checked visually.

3. The SIFT feature point-based power transmission corridor tree barrier identification method of claim 2, wherein: The pre-processing of the collected images comprises, The collected image data is input, a 3x3 and 5x5 window median filter is selected for image denoising, for each pixel in the image data, the median of all pixels in the neighborhood is calculated, and the original pixel value is replaced with the median, and the step is repeated until the noise in the image is suppressed; The low-contrast image is enhanced by histogram equalization to ensure the distinguishability of the tree profile in the complex background, a gray level histogram of the input image data is calculated, and a cumulative distribution function of the image is calculated , the gray value range of the image is set as , wherein L is the gray level number: wherein, is the number of pixels with gray value i, is the total number of pixels of the image, i is the variable index, k is the number of gray values; each pixel value of the image is re-mapped using the cumulative distribution function to obtain an enhanced image: each pixel value is mapped to a new value : The local contrast limited adaptive histogram equalization method is used to perform histogram equalization in local regions, the image is divided into multiple local regions of 8x8 or 16x16 pixels, histogram equalization is performed on each region respectively, and the growth of contrast is limited, all regions are combined, and an enhanced overall image is obtained, wherein the limited contrast is: : wherein, is an upper limit for the contrast ratio, is the maximum pixel value of the image within the region; Different angles and postures of the collected images are standardized by geometric transformation, for each image, random rotation and scaling operations are performed through an affine transformation matrix.

4. The SIFT feature point-based power transmission corridor tree barrier identification method of claim 3, wherein: The feature matching includes, for each feature point in the new image, computing a corresponding generated feature descriptor For all tree images in the database, compute descriptors for all feature points and compute the Euclidean distance to the new image 's feature descriptor: wherein, and are values of the i-th dimension in descriptors and respectively. selecting the two descriptors closest in distance and wherein, is the smallest matching distance, is the second smallest matching distance; According to and Lowe ratio test, i.e. filtering out false matches by comparing and the ratio : Set the matching threshold τ = 0.7, when , it is considered that the current feature point matching is effective; when , it is considered that the current feature point matching is ineffective; According to the matching result, a machine learning classification algorithm is used to classify the trees, a label is assigned to the matched effective feature point, the label is the tree category of the corresponding image, and the trees with assigned labels form a calibration data set.

5. The SIFT feature point-based power transmission corridor tree barrier identification method of claim 4, wherein: The classification recognition includes analysis of spatial distribution of trees based on results of feature matching, extraction of spatial coordinates of each tree , clustering of positions of trees by a clustering algorithm, the position of each tree serving as input data of a K-means algorithm, and identification of a tree region in the target image that constitutes a threat to the power transmission line, Setting the number of clusters K, i.e. dividing the tree area into different groups according to the preset number of clusters, randomly initializing the centroids of K clusters Each centroid is a two-dimensional coordinate: , and is the two-dimensional coordinate of the kth centroid; The Euclidean distance of the tree points to each cluster center is calculated, and the distance between the ith tree point and the kth cluster center is: Cluster assignment: assign the tree points to the cluster to which the nearest cluster center belongs; Centroid update: for the kth cluster, the new centroid is determined by the mean of all tree points in the cluster: wherein, and is the spatial coordinate of the i-th tree, is the number of tree points in the cluster until the center of the cluster no longer changes significantly or the preset maximum number of iterations is reached, after the clustering is completed, the tree density and distribution in each cluster are analyzed, and the area affecting the power transmission line is identified, when the tree density in each cluster is greater than 110 trees per hectare, the current area is considered to be a tree barrier area; For the spatial coordinates of each tree, geometric constraints and classification judgments are performed to determine whether the tree constitutes a tree barrier object, and the judgment rules include the horizontal distance between the tree and the power line, the tree height and the tree width.

6. The SIFT feature point-based power transmission corridor tree barrier identification method of claim 5, wherein: The horizontal distance between the tree and the power line comprises calculating the horizontal distance between the tree and the power line according to the spatial position of the tree: where D is the horizontal distance of the tree to the power line, is the spatial coordinate of the power line; The tree height, when the height of the tree exceeds a set threshold , and the horizontal distance of the power transmission line is less than the threshold , the current tree is determined as a tree obstacle object. The width of the tree in the image is calculated by contour detection of the image When the tree width is greater than a set threshold and the horizontal distance D of the power line is less than a threshold , the current tree is determined to be a tree obstacle.

7. The SIFT feature point-based power transmission corridor tree barrier identification method of claim 6, wherein: The detection rate comprises using the accuracy, recall, precision and F1 value evaluation index to evaluate the performance of the method; The performance of different tree species is evaluated by the tree barrier detection rate to determine whether the trees will affect the power line; wherein the tree barrier detection rate is calculated by comparing the detection result with the manually labeled tree barrier region: When the tree barrier detection rate is greater than 90%, it is considered that the current tree species performs well in the current identification, the identification is accurate, and the current identification result can be adopted, wherein, is the number of regions correctly identified as tree barriers, is the total tree barrier region. 8.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-7. The processor executes the computer program to realize the steps of the tree barrier identification method for power corridor based on SIFT feature points according to any one of claims 1-7.

9. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by a processor to implement the steps of the SIFT feature point-based power transmission corridor tree barrier identification method in any one of claims 1 to 7.

Images