A base station antenna posture analysis method

Abstract

The application provides a base station antenna posture analysis method, comprising: collecting a base station inspection video stream; performing equal-interval frame extraction processing on the video stream to generate a two-dimensional image sequence; solving a camera intrinsic parameter matrix and an external parameter posture of each frame of image in the two-dimensional image sequence, and constructing an initial three-dimensional Gaussian scene; performing reverse propagation optimization on the initial three-dimensional Gaussian scene to generate a target three-dimensional Gaussian scene; performing Poisson surface reconstruction or extracting a preset probability density section according to the opacity and distribution density of Gaussian points to generate a dense point cloud; separating base station antenna point cloud and base station antenna support point cloud by using the camera intrinsic parameter matrix and the camera posture; determining an antenna main radiation plane; obtaining a target unit normal vector of the antenna main radiation plane; and determining posture parameters of each base station antenna based on the target unit normal vector; thereby solving the problems of multi-dimensional information loss and poor three-dimensional reconstruction quality when a two-dimensional video is used to analyze the posture of a base station antenna.

Description

Technical Field

[0001] This application relates to the field of point cloud data technology, and in particular to a method for base station antenna attitude analysis. Background Technology

[0002] With the continuous expansion of mobile communication networks, the total number of base station antennas deployed in China has exceeded ten million. To ensure stable communication quality, the attitude parameters (pitch and azimuth angles) of base station antennas directly determine the signal coverage and user experience. Therefore, regular monitoring and maintenance of antenna attitude has become an indispensable core task in the daily network operation and maintenance of operators.

[0003] Currently, the mainstream technical solutions for base station antenna inspection in the industry are mainly divided into two categories. One is the manual inspection mode, in which maintenance personnel climb the tower to reach the antenna installation location and use tools such as tilt meters and compasses to measure the antenna's elevation and azimuth parameters on-site. The other is the drone-assisted inspection solution, which involves controlling a drone to fly near the antenna to collect two-dimensional video and image materials, and then the back-end maintenance personnel manually visually check the captured images to determine whether the antenna attitude has shifted.

[0004] However, the manual climbing measurement mode is not only extremely dangerous and inefficient, but also requires a lot of manpower and time to measure a single tower, making it difficult to meet the inspection needs of the current massive number of base stations. On the other hand, the analysis scheme based on UAV 2D video cannot accurately restore the actual orientation and stress state of the antenna in 3D space because 2D images naturally lack depth information. At the same time, when dealing with narrow and long objects with weak texture such as antennas, 3D reconstruction methods such as SfM point cloud often generate point clouds that are too sparse and accompanied by a lot of noise, making it difficult to extract an accurate physical plane model of the antenna. Ultimately, this leads to large attitude measurement errors and insufficient reliability of results, which cannot meet the actual requirements of high-precision network operation and maintenance. Summary of the Invention

[0005] This application provides a base station antenna attitude analysis method to solve the technical problems of missing multi-dimensional information and poor 3D reconstruction quality when using UAV 2D video for base station antenna attitude analysis.

[0006] This application provides a base station antenna attitude analysis method, including: Collect video streams from base station inspections; The video stream is subjected to equal-interval frame extraction to generate a two-dimensional image sequence; Using the structure-reconstruction-motion algorithm, feature extraction and matching are performed on the two-dimensional image sequence, the camera intrinsic parameter matrix and extrinsic parameter pose of each frame in the two-dimensional image sequence are calculated, and an initial three-dimensional Gaussian scene is constructed. Based on the camera intrinsic parameter matrix and camera pose, the initial 3D Gaussian scene is optimized by backpropagation using the photometric loss function to generate the target 3D Gaussian scene. Based on the opacity and distribution density of Gaussian points in the target 3D Gaussian scene, Poisson surface reconstruction or extraction of a preset probability density section is performed to generate a dense point cloud; the dense point cloud restores the edge contour of the base station antenna plane and the geometric features of the base station antenna support. Based on the dense point cloud, the base station antenna point cloud and the corresponding base station antenna support point cloud are separated using the camera intrinsic parameter matrix and camera pose. Based on the point cloud of the base station antenna, a plane fitting process is performed to determine the main radiation plane of the antenna; Obtain the target unit normal vector of the antenna's main radiating plane; Based on the target unit normal vector, the attitude parameters of each base station antenna are determined; the attitude parameters include: pitch angle and azimuth angle.

[0007] In some embodiments, the steps of using the structure-of-motion reconstruction algorithm to extract and match features from the two-dimensional image sequence, calculate the camera intrinsic parameter matrix and extrinsic pose of each frame in the two-dimensional image sequence, and construct an initial three-dimensional Gaussian scene include: The motion reconstruction structure algorithm is used to extract feature points from the two-dimensional image sequence; Based on the feature points, a sparse point cloud of the scene is constructed using feature matching, and the camera intrinsic parameter matrix and camera pose of each frame image relative to the scene coordinate system are calculated. An initial 3D Gaussian scene is constructed with the sparse point cloud of the scene as the centroid; each Gaussian point in the initial 3D Gaussian scene contains Gaussian properties; the Gaussian properties include: position, scaling length, rotation quaternion, and opacity.

[0008] In some embodiments, the step of performing backpropagation optimization on the initial 3D Gaussian scene based on the camera intrinsic parameter matrix and camera pose using a photometric loss function to generate a target 3D Gaussian scene includes: Based on the camera intrinsic parameter matrix and camera pose, the Gaussian points are projected onto the image plane using differentiable rendering to generate a rendered image. The rendered image and the two-dimensional image sequence are compared to calculate the photometric loss; Based on the photometric loss, the Gaussian properties are optimized by backpropagation using the photometric loss function to generate the target 3D Gaussian scene.

[0009] In some embodiments, the step of backpropagating and optimizing the Gaussian properties based on the photometric loss using the photometric loss function to generate the target 3D Gaussian scene includes: Based on the photometric loss, the Gaussian property is optimized by backpropagation using the photometric loss function, and three scaling axes are calculated for each Gaussian point. A penalty term is applied to the scaling axis, causing the scaling ratio of one of the scaling axes to approach zero, while the remaining two scaling axes expand in the local tangent plane, so that the Gaussian points form a thin sheet distribution on the surface of the object to be reconstructed, and ensuring that the direction of the flat axis of the Gaussian points is the same as the direction of the local geometric normal of the object to be reconstructed. After the alignment optimization is completed on the surface of the object to be reconstructed, the target 3D Gaussian scene is generated.

[0010] In some embodiments, the step of separating the base station antenna point cloud and the corresponding base station antenna support point cloud based on the dense point cloud, using the camera intrinsic parameter matrix and camera pose, includes: Using a deep learning detection model, target detection is performed on the two-dimensional image sequence to obtain a directed bounding box containing the base station antenna and the bracket; the directed bounding box includes: the center coordinates, length and width dimensions, and rotation angle information of the base station antenna and the bracket; Using the camera intrinsic parameter matrix and extrinsic pose, each three-dimensional coordinate point in the dense point cloud is back-projected into the two-dimensional image sequence of the corresponding viewpoint; The number of times each three-dimensional coordinate point falls within the corresponding directed bounding box is determined. If the number of times is greater than a preset number, the corresponding target three-dimensional coordinate point is retained; if the number of times is less than or equal to the preset number, the corresponding three-dimensional coordinate point is discarded. A spatial clustering algorithm is performed on the target three-dimensional coordinate points to separate the base station antenna point cloud and the corresponding base station antenna support point cloud.

[0011] In some embodiments, the step of performing a spatial clustering algorithm on the target three-dimensional coordinate points to separate the base station antenna point cloud and the corresponding base station antenna support point cloud includes: A spatial clustering algorithm is performed on the target three-dimensional coordinate points to determine the point cloud of the base station antenna components and the number of base station antenna components based on the Euclidean distance; Based on the directed bounding box, the base station antenna identification box is determined; The base station antenna identification frame in the two-dimensional image sequence is labeled; The base station antenna identification box after the label is reprojected back to the target three-dimensional Gaussian scene using the camera intrinsic parameter matrix and extrinsic parameter pose, and the base station antenna point cloud is extracted using a multi-view and plane fitting strategy. Based on the point cloud of the base station antenna component and the point cloud of the base station antenna, Boolean difference operation is performed to separate the corresponding point cloud of the base station antenna support.

[0012] In some embodiments, the step of performing plane fitting processing based on the base station antenna point cloud to determine the main radiating plane of the antenna includes: A candidate planar parameter set for the point cloud of a single base station antenna is determined using a random sample consensus algorithm; the candidate planar parameter set is used to separate the front, side, and noise of the base station antenna. Calculate the number of points in each candidate plane in the candidate plane parameter set; Project the three-dimensional point set of each candidate plane onto its corresponding fitting plane, and convert the three-dimensional point set into two-dimensional point coordinates; Based on the two-dimensional point coordinates, the projected area of ​​the circumscribed polygon corresponding to the three-dimensional point set is calculated using the convex hull. Based on the number of points and the projected area, a scoring function is used to determine the score of each candidate plane, and the plane with the highest score that meets the antenna aspect ratio characteristics is extracted as the main radiation plane of the antenna.

[0013] In some embodiments, the scoring function is: Score i = α · f ( N i )+ β · g ( S i ); In the formula, f (·)and g (·) represents the normalization processing function; α and β Preset weighting coefficients; N i Number of points; S i This represents the projected area.

[0014] In some embodiments, the step of obtaining the target unit normal vector of the antenna's main radiating plane includes: Using the point cloud of the base station antenna bracket, the representative location point of the bracket is determined; Using the point cloud of the base station antenna, the location point represented by the antenna is determined; Based on the location point represented by the bracket and the location point represented by the antenna, a reference direction vector is constructed pointing from the location point represented by the bracket to the location point represented by the antenna; the reference direction vector represents the spatial layout relationship of the base station antenna relative to the base station antenna bracket. Obtain the initial unit normal vector of the main radiating plane of the antenna; When the direction of the initial unit normal vector is consistent with the direction of the reference direction vector, the initial unit normal vector is reversed to obtain the target unit normal vector.

[0015] In some embodiments, the step of determining the attitude parameters of each base station antenna based on the target unit normal vector includes: Define a vertical reference axis; the direction of the vertical reference axis is consistent with the direction of gravity; Calculate the spatial angle between the target unit normal vector and the vertical reference axis; The spatial angle is converted into a pitch angle according to a preset conversion relationship; The target unit normal vector is projected onto a horizontal reference plane orthogonal to the vertical reference axis to obtain a horizontal projection vector. Calculate the directional angle between the horizontal projection vector and the preset horizontal zero-degree reference direction in the scene; Calculate the radian deviation between the base station antenna projection vector and the relative reference vector based on a reference antenna with a known preset azimuth angle; Using the radian deviation, absolute angle compensation is performed on the directional included angle to obtain the absolute azimuth angle of the base station antenna.

[0016] This application provides a base station antenna attitude analysis method, comprising: acquiring a base station inspection video stream; performing equal-interval frame extraction on the video stream to generate a two-dimensional image sequence; using a structure-of-motion (SOG) algorithm to extract and match features from the two-dimensional image sequence, calculating the camera intrinsic parameter matrix and extrinsic pose of each frame in the two-dimensional image sequence, and constructing an initial three-dimensional Gaussian scene; based on the camera intrinsic parameter matrix and camera pose, performing backpropagation optimization on the initial three-dimensional Gaussian scene using a photometric loss function to generate a target three-dimensional Gaussian scene; and performing Poisson surface reconstruction or extracting a preset probability based on the opacity and distribution density of Gaussian points in the target three-dimensional Gaussian scene. A density cross-section is used to generate a dense point cloud. This dense point cloud reconstructs the edge contour of the base station antenna plane and the geometric features of the base station antenna support. Based on this dense point cloud, the base station antenna point cloud and the corresponding base station antenna support point cloud are separated using the camera intrinsic parameter matrix and camera pose. Based on the base station antenna point cloud, plane fitting processing is performed to determine the main radiating plane of the antenna. The target unit normal vector of the main radiating plane of the antenna is obtained. Based on the target unit normal vector, the attitude parameters of each base station antenna are determined. The attitude parameters include pitch angle and azimuth angle to ensure the uniqueness and accuracy of the azimuth angle calculation results, and to improve the stability and repeatability accuracy of the pitch angle calculation. Attached Figure Description

[0017] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of the base station antenna attitude analysis method in this application. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0020] Because some technologies suffer from the problems of missing multi-dimensional information and poor 3D reconstruction quality when using UAV 2D video for base station antenna attitude analysis, this application provides a base station antenna attitude analysis method to solve these problems. The base station antenna attitude analysis method is described below: like Figure 1 The diagram shown is a flowchart of the base station antenna attitude analysis method in this application.

[0021] This application provides a base station antenna attitude analysis method, including the following steps: S100: Collects video streams from base station inspections.

[0022] S200: Perform equal-interval frame extraction on the video stream to generate a two-dimensional image sequence.

[0023] S300: Using the structure-of-motion algorithm, feature extraction and matching are performed on the two-dimensional image sequence, the camera intrinsic parameter matrix and extrinsic parameter pose of each frame in the two-dimensional image sequence are calculated, and an initial three-dimensional Gaussian scene is constructed.

[0024] The steps of using the structure-of-motion reconstruction algorithm to extract and match features from the two-dimensional image sequence, calculate the camera intrinsic parameter matrix and extrinsic pose of each frame in the two-dimensional image sequence, and construct an initial three-dimensional Gaussian scene include the following sub-steps: S310: Use the Structure from Motion (SfM) algorithm to extract feature points from the two-dimensional image sequence.

[0025] S320: Based on the feature points, construct a sparse point cloud of the scene using feature matching, and calculate the camera intrinsic parameter matrix and camera pose (extrinsic parameter matrix) of each frame image relative to the scene coordinate system.

[0026] S330: Construct an initial 3D Gaussian scene with the sparse point cloud of the scene as the centroid; each Gaussian point in the initial 3D Gaussian scene contains Gaussian attributes; the Gaussian attributes include: position, scaling length, rotation quaternion and opacity.

[0027] S400: Based on the camera intrinsic parameter matrix and camera pose, backpropagation optimization of the initial 3D Gaussian scene is performed using the photometric loss function to generate the target 3D Gaussian scene. S410: The step of generating a target 3D Gaussian scene by backpropagating the initial 3D Gaussian scene based on the camera intrinsic parameter matrix and camera pose using a photometric loss function includes the following sub-steps: S420: Based on the camera intrinsic parameter matrix and camera pose, the Gaussian points are projected onto the image plane using differentiable rendering to generate a rendered image.

[0028] S430: Compare the rendered image and the two-dimensional image sequence to calculate the photometric loss.

[0029] S440: Based on the photometric loss, the Gaussian property is optimized by backpropagation using the photometric loss function to generate the target 3D Gaussian scene.

[0030] Specifically, a scene representation based on 3D Gaussian splatting is established using sparse point clouds as initial centroids. Each Gaussian point contains attributes such as position, scaling length, rotation quaternion, and opacity. The Gaussian points are projected onto the image plane using differentiable rendering, and the sampled image is used as the ground truth. The Gaussian attributes are then optimized using backpropagation with a photometric loss function.

[0031] The step of generating a target 3D Gaussian scene by backpropagating the Gaussian property based on the photometric loss function includes the following sub-steps: S441: Based on the photometric loss, the Gaussian property is optimized by backpropagation using the photometric loss function, and the three scaling axes of each Gaussian point are calculated.

[0032] S442: Apply a penalty term to the scaling axes, causing the scaling ratio of one of the scaling axes to approach zero, and the remaining two scaling axes to expand in the local tangent plane, so that the Gaussian points form a thin sheet distribution on the surface of the object to be reconstructed, and ensure that the direction of the flat axis of the Gaussian points is the same as the direction of the local geometric normal of the object to be reconstructed.

[0033] S443: After completing alignment optimization on the surface of the object to be reconstructed, generate the target 3D Gaussian scene.

[0034] Specifically, a surface alignment regularization term is introduced during the optimization process. By calculating the three scaling axes of each Gaussian ellipsoid (Gaussian point), a penalty term is applied to make the scaling ratio of one axis approach zero, while the other two axes expand within the local tangent plane. This process forces the Gaussian distribution to form a sheet-like distribution on the object surface, ensuring that its flat axis is consistent with the local geometric normal direction of the object to be reconstructed. After the alignment optimization is completed on the surface of the object to be reconstructed, the target 3D Gaussian scene is generated.

[0035] S500: Based on the opacity and distribution density of Gaussian points in the target 3D Gaussian scene, perform Poisson surface reconstruction or extract a preset probability density section to generate a dense point cloud; the dense point cloud restores the edge contour of the base station antenna plane and the geometric features of the base station antenna support.

[0036] Specifically, after surface alignment optimization, based on the opacity and distribution density of Gaussian points in the target 3D Gaussian scene, Poisson surface reconstruction or direct extraction of high-probability-density sections is performed to generate a geometrically continuous "skin-level" dense point cloud. This dense point cloud accurately reconstructs the edge contours of the base station antenna plane and the subtle geometric features of the support, providing a high-resolution data foundation for subsequent angle calculations.

[0037] This application addresses the problems of numerous background noise points, distortion of slender target structures, and the difficulty of directly outputting high-purity target point clouds in the 3D reconstruction of antenna base station videos. It proposes a 3D point cloud reconstruction and purification method for antenna base station targets. This method is based on a 3D Gaussian representation and surface-aligned reconstruction framework. It selects an appropriate coarse reconstruction optimization strategy according to the structural characteristics and scene features of the antenna base station target. In the point cloud extraction stage, multi-attribute joint screening is performed on the reconstructed Gaussian elements, combined with outlier removal based on neighborhood consistency, to suppress non-target background components. Subsequently, target surface point clouds are generated based on surface-level sampling, and the point cloud results are directly output when the target representation requirements are met, thus avoiding redundant background and structural distortion introduced by the subsequent meshing process.

[0038] The above technical solution can generate a "skin layer" surface point cloud with high geometric continuity, which significantly improves the restoration accuracy of antenna edges and planar regions, provides high-quality input data for subsequent point cloud segmentation and angle analysis, and solves the problem that point clouds cannot be directly used for precise plane fitting.

[0039] S600: Based on the dense point cloud, using the camera intrinsic parameter matrix and camera pose, separate the base station antenna point cloud and the corresponding base station antenna support point cloud.

[0040] The step of separating the base station antenna point cloud and the corresponding base station antenna support point cloud based on the dense point cloud, using the camera intrinsic parameter matrix and camera pose, includes the following sub-steps: S610: Using a deep learning detection model, target detection is performed on the two-dimensional image sequence to obtain a directed bounding box containing the base station antenna and the bracket; the directed bounding box includes: the center coordinates, length and width dimensions, and rotation angle information of the base station antenna and the bracket.

[0041] Specifically, a deep learning detection model optimized for rotating objects (such as YOLO-OBB) is used to detect objects in a two-dimensional image sequence. The model outputs a directed bounding box for antennas and active antenna units (AAUs) such as supports, containing the object's center coordinates, length, width, and rotation angle information.

[0042] S620: Using the camera intrinsic parameter matrix and camera pose, project each three-dimensional coordinate point P(x,y,z) in the dense point cloud backwards into the two-dimensional image sequence of the corresponding viewpoint.

[0043] S630: Determine the number of times each of the three-dimensional coordinate points falls within the corresponding directed bounding box. If the number of times is greater than a preset number, retain the corresponding target three-dimensional coordinate point; if the number of times is less than or equal to the preset number, discard the corresponding three-dimensional coordinate point.

[0044] Specifically, a voting evaluation mechanism is constructed, traversing all visible images for each stated 3D coordinate point. If the projected mapping point p(u,v) falls within the corresponding 2D directed bounding box, a valid vote is recorded. The system sets a threshold T (preset number of votes) and only retains 3D coordinate points with a voting rate greater than T. This mechanism effectively eliminates background occlusion, shadows, and non-target areas (i.e., the corresponding 3D coordinate points) caused by point cloud drift.

[0045] For example, current methods typically perform Euclidean distance segmentation or semantic segmentation directly in 3D space, which struggles to handle the complex topological relationships where base station towers and mounted equipment are closely connected. This application utilizes camera poses generated by COLMAP to back-project the 3D point cloud onto multiple frames of original 2D images, and combines this with a YOLO-OBB detection model optimized for rotating objects for cross-validation. Through multi-view... Figure 1 With its consistent voting mechanism, this application can effectively identify and eliminate interference from shadows, occlusions, background noise, and point cloud drift, achieving precise separation of individual antennas and pole supports in complex and cluttered backgrounds, thus improving the robustness of target extraction.

[0046] S640: Perform a spatial clustering algorithm on the target three-dimensional coordinate points to separate the base station antenna point cloud and the corresponding base station antenna support point cloud.

[0047] The step of performing a spatial clustering algorithm on the target three-dimensional coordinate points to separate the base station antenna point cloud and the corresponding base station antenna support point cloud includes the following sub-steps: S641: Perform a spatial clustering algorithm on the target three-dimensional coordinate points to determine the point cloud of the base station antenna components and the number of base station antenna components based on the Euclidean distance.

[0048] Specifically, the DBSCAN (density-based spatial clustering with noise) algorithm is executed on the preserved point cloud to aggregate adjacent points into independent antenna individuals based on Euclidean distance constraints. At the same time, the number of antennas, n, is counted for use in the next step.

[0049] S642: Based on the directed bounding box, determine the base station antenna identification box.

[0050] S643: Label the base station antenna identification frame in the two-dimensional image sequence.

[0051] S644: The base station antenna identification box after the label is reprojected back to the target three-dimensional Gaussian scene using the camera intrinsic parameter matrix and extrinsic parameter pose, and the base station antenna point cloud is extracted using a multi-view and plane fitting strategy.

[0052] Specifically, the antenna identification boxes in each image are labeled 1-n using the camera pose. For each labeled antenna, its 2D label identification box (antenna identification box) is reprojected back into the 3D point cloud, and a clean base station antenna point cloud (excluding the support) is obtained using a multi-view and plane fitting strategy.

[0053] S645: Based on the point cloud of the base station antenna assembly and the point cloud of the base station antenna, perform Boolean difference operation to separate the corresponding point cloud of the base station antenna support. By performing Boolean difference operation on the coarse segmented point cloud (point cloud of the base station antenna assembly) with the support and the fine segmented point cloud (point cloud of the base station antenna) after removing the support, separate and save the point cloud of the base station antenna support corresponding to each antenna for subsequent vector correction reference.

[0054] For example, existing 3D plane fitting algorithms (such as standard RANSAC) extract normal vectors with a 180° directional uncertainty (i.e., the normal vector may point towards the front or back of the object). This application extracts the point cloud of the mast on the back of the antenna using a differential algorithm and constructs a physical reference vector between the mast center and the antenna center. This reference vector is then used to determine and correct the polarity of the fitted normal vector, ensuring that the final fitted normal vector always points outward from the tower. This solves the problem of reverse deviation that may occur in azimuth angles during automated measurement, ensuring the uniqueness and objectivity of the measurement results, greatly reducing the system's misjudgment rate, and enabling UAV inspection data to directly replace manual measurements with engineering reliability.

[0055] S700: Based on the point cloud of the base station antenna, perform plane fitting processing to determine the main radiation plane of the antenna.

[0056] The step of performing plane fitting processing based on the base station antenna point cloud to determine the main radiation plane of the antenna includes the following sub-steps: S710: Using a random sampling consensus algorithm, determine a candidate planar parameter set for the point cloud of a single base station antenna; the candidate planar parameter set is used to separate the front, side, and noise of the base station antenna.

[0057] Specifically, for the segmented point cloud of a single base station antenna, the Random Sample Consensus Algorithm (RANSAC) is used for plane parameter estimation. This application performs a multi-iterative extraction process: fitting the optimal plane parameters in the current point cloud set and extracting the inliers belonging to that plane; then removing these inliers from the original point cloud; and continuing the next plane fitting in the remaining point cloud subset. This process generates multiple candidate plane parameter sets π. i ={A i B i C i D i This ensures that noise from the front, sides, and surrounding areas of the antenna can be effectively separated.

[0058] S720: Calculate the number of points in each candidate plane in the candidate plane parameter set; for each candidate plane π i Calculate the number of interior points in the plane. N i And it is used as the first indicator reflecting the reliability of the fit.

[0059] S730: Project the three-dimensional point set of each candidate plane onto its corresponding fitting plane, and convert the three-dimensional point set into two-dimensional point coordinates.

[0060] S740: Based on the coordinates of the two-dimensional points, calculate the projected area of ​​the circumscribed polygon corresponding to the three-dimensional point set using the convex hull.

[0061] Specifically, the 3D point set of each candidate plane is projected onto its corresponding fitting plane, transforming it into 2D point coordinates. Convex hull calculation is then used to generate the circumscribed polygon of the point set and calculate its projected area. S i This metric is used to filter out small planes (such as local or residual noise on the support structure) that, although the point cloud is dense, do not conform to the antenna's physical dimensions.

[0062] S750: Based on the number of points and the projected area, a scoring function is used to determine the score of each candidate plane, and the plane with the highest score that conforms to the antenna aspect ratio is extracted as the main radiation plane of the antenna. The scoring function is: Score i = α · f ( N i )+ β · g ( S i ); In the formula, f (·)and g (·) represents the normalization processing function; α and β Preset weighting coefficients (set according to antenna geometry); N i Number of points; S i This represents the projected area.

[0063] S800: Obtain the target unit normal vector of the main radiating plane of the antenna.

[0064] The step of obtaining the target unit normal vector of the antenna's main radiating plane includes the following sub-steps: S810: Use the point cloud of the base station antenna bracket to determine the representative location point of the bracket.

[0065] S820: Use the base station antenna point cloud to determine the antenna representative location point.

[0066] S830: Based on the location point represented by the bracket and the location point represented by the antenna, a reference direction vector is constructed pointing from the location point represented by the bracket to the location point represented by the antenna; the reference direction vector represents the spatial layout relationship of the base station antenna relative to the base station antenna bracket.

[0067] Specifically, the representative position point A of the support frame is determined using the segmented point cloud of the support frame, and the representative position point B of the antenna is determined using the point cloud of the antenna body. A reference direction vector is generated by pointing from the representative position point A of the support frame to the representative position point B of the antenna. This reference direction vector represents the spatial layout relationship of the antenna relative to the support frame.

[0068] S840: Obtain the initial unit normal vector of the main radiating plane of the antenna.

[0069] S850: When the direction of the initial unit normal vector is consistent with the direction of the reference direction vector, the initial unit normal vector is reversed to obtain the target unit normal vector.

[0070] Specifically, the initial unit normal vector of the antenna's main radiating surface is obtained. Based on the reference direction vector, a consistency check is performed on the direction of the initial unit normal vector; when it is determined that the current direction of the initial unit normal vector corresponds to the side facing the support pole, it is reversed to obtain the corrected unit normal vector (target unit normal vector). The target unit normal vector is constrained to point outward from the antenna surface in free space away from the support.

[0071] S900: Based on the target unit normal vector, determine the attitude parameters of each base station antenna; the attitude parameters include: pitch angle and azimuth angle.

[0072] The step of determining the attitude parameters of each base station antenna based on the target unit normal vector includes the following sub-steps: S910: Define a vertical reference axis; the direction of the vertical reference axis is consistent with the direction of gravity.

[0073] S920: Calculate the spatial angle between the target unit normal vector and the vertical reference axis.

[0074] S930: Convert the spatial angle into a pitch angle according to a preset conversion relationship.

[0075] Specifically, a vertical reference axis associated with the direction of gravity is defined in the reconstructed scene. The spatial angle between the target unit normal vector and the vertical reference axis is calculated, and the spatial angle is converted into an elevation angle parameter reflecting the degree of tilt of the antenna relative to the horizontal plane according to a preset conversion relationship; wherein, it is agreed that a positive value of the elevation angle corresponds to outward tilt and a negative value corresponds to inward tilt.

[0076] S940: Project the target unit normal vector onto a horizontal reference plane orthogonal to the vertical reference axis to obtain a horizontal projection vector.

[0077] S950: Calculate the directional angle between the horizontal projection vector and the preset horizontal zero-degree reference direction in the scene; the directional angle is used as the azimuth information of the antenna.

[0078] S960: Calculate the radian deviation between the base station antenna projection vector and the relative reference vector based on a reference antenna with a known preset azimuth angle; S970: Using the radian deviation, perform absolute angle compensation on the directional included angle to obtain the absolute azimuth angle of the base station antenna.

[0079] Specifically, by selecting a reference antenna with a known preset azimuth angle within the scene as a benchmark, the radian deviation of the projection vector of the antenna under test relative to the reference vector is calculated. Finally, combined with the reference azimuth angle compensation value, the system automatically outputs the absolute azimuth angle data of each antenna and writes all attitude parameters (ID, elevation angle, azimuth angle) into a structured results file.

[0080] For example, in the plane fitting stage, this application abandons the single criterion of using "maximum number of interior points" as the optimal plane. Instead, it employs an iterative RANSAC strategy to extract multiple candidate planes and constructs a weighted scoring function that comprehensively considers the size of the interior points, the fitting variance, and the two-dimensional projected area distribution of the point cloud on the plane. This allows for the automatic and accurate identification of the effective main radiating surface from the antenna's main surface, sides, back, and support surface, effectively avoiding the confusion of fitting targets caused by the complexity of the antenna geometry, and further improving the stability and accuracy of the vertical elevation angle calculation.

[0081] This application provides a base station antenna attitude analysis method, aiming to solve the problems of high risk of manual measurement, inaccurate estimation of two-dimensional images, and difficulty in automatically determining the orientation of objects in three-dimensional point cloud processing during base station inspection. Specifically, it includes: 1. High-fidelity 3D reconstruction: Solve the problems of missing and sparse object edges in point cloud reconstruction, as well as the problem of shielding the surrounding environment during reconstruction, such as clouds and surrounding buildings; achieve high-quality 3D reconstruction with surface alignment of base station antennas through video streams.

[0082] 2. Accurate target extraction in complex environments: Solve the problem of fully automatic extraction of 3D point clouds of a single antenna and its corresponding support (pole) in complex base station environments (including towers, cables, and various mounted devices).

[0083] 3. Azimuth Orientation Issue: This addresses the problem that after fitting the antenna principal plane, the system cannot automatically determine whether the normal vector "points to the front of the antenna" or "points to the back of the mast," thereby eliminating the reverse error in azimuth calculation.

[0084] 4. Accuracy of plane fitting under multi-factor interference: For the noise and uneven regions in the antenna point cloud, optimize the fitting algorithm to improve the robustness of elevation angle calculation.

[0085] This application provides a base station antenna attitude analysis method, which has the following advantages: 1. Significantly Improved Geometric Reconstruction Accuracy: By introducing a surface-aligned Gaussian Splashing (SuGaR) technique, the problem of sparse point clouds (especially on the antenna panel surface) caused by SfM point clouds in processing weakly textured, narrow and elongated antenna components is overcome. This application can generate high-resolution surface skin layer point clouds, providing an accurate geometric basis for sub-degree-level antenna attitude measurements.

[0086] 2. Effectively eliminates logical defects in azimuth angle calculation: This application proposes a normal vector correction mechanism based on the mast position reference. By automatically calculating the relative position vector between the mast and the antenna body, it solves the technical bottleneck of not being able to automatically distinguish the positive and negative directions of the normal vector in three-dimensional plane fitting, ensuring the uniqueness and accuracy of the azimuth angle calculation results and eliminating engineering risks caused by reverse deviation.

[0087] 3. High anti-interference capability in complex environments: Employing multi-view directed bounding box (OBB) projection and statistical voting algorithms, it deeply integrates 2D semantic information with 3D spatial topology. This mechanism effectively filters out occlusion interference from complex backgrounds such as base station towers, cables, and lightning rods, ensuring that target extraction accuracy in cluttered environments is higher than that of single-view segmentation methods.

[0088] 4. Full-process automation and high operational efficiency: This application achieves fully automated closed-loop processing from UAV video input to attitude parameter output. Compared with manual climbing operations, it significantly reduces safety risks and labor costs; compared with traditional blind estimation of two-dimensional images, it provides quantifiable three-dimensional physical evidence, greatly improving the digitalization level and data consistency of communication base station inspection.

[0089] 5. Robustness of the fitting optimization algorithm: By establishing a dual-objective evaluation model that combines the number of interior points and the projected area, this application can automatically extract the optimal radiating surface from multiple candidate planes, effectively avoiding the interference of the antenna back, side and support planes on the fitting results, and further improving the stability and repeatability accuracy of the elevation angle calculation.

[0090] This application provides a base station antenna attitude analysis method, and specific embodiments are as follows: 1. System environment and architecture preparation: The execution environment of this application embodiment includes: a drone data acquisition unit, a high-performance GPU computing node, and a storage unit.

[0091] Drone data acquisition unit: A drone equipped with a high-resolution camera is used to fly around the base station in all directions to acquire 4K high-definition video streams covering all dimensions of the antenna.

[0092] Computing Unit: Equipped with a CUDA-accelerated graphics card (such as NVIDIA RTX 3090 or above), pre-installed with algorithm libraries such as Open3D, NumPy, and PyTorch, as well as point cloud analysis tools such as COLMAP.

[0093] 2. Data preprocessing and pose recovery: In practice, the first step is to automatically extract frames from the input video file.

[0094] Implementation details: Set the video duration to T, the preset number of frame extractions to N, calculate the sampling interval, and use the ffmpeg tool to export the image.

[0095] Pose calculation: The image sequence is input into the COLMAP analysis module. The system performs feature extraction and sequential matching, followed by incremental mapping. Finally, a binary model file containing the camera intrinsic matrix, rotation and translation vectors for each frame, and sparse point cloud is exported.

[0096] 3. Surface-aligned dense point cloud reconstruction: Implementation details: Using the pose derived from COLMAP as a priori, a 3D Gaussian splash optimization procedure is initiated. During the training iterations, SuGaR regularization loss is introduced. The system adjusts the scale properties of the Gaussian distribution to make it exhibit a flattened distribution on the antenna surface.

[0097] Point cloud generation: When the training reaches a convergence tendency, the skin layer extraction command is executed. The generated point cloud file contains the microscopic geometric details of the antenna plane, and the point cloud density must meet the requirement of no less than 10 sampling points per square centimeter to ensure the accuracy of subsequent plane fitting.

[0098] 4. Individual antenna segmentation: Projection mapping: Traversing every point in a dense point cloud P j Using formulas p ij = K [ R i | t i ] P j Project the point onto the first... i Frame image, in which K For internal reference purposes. R , t For external reference.

[0099] OBB Detection and Voting: Calling a pre-trained YOLO-OBB model to evaluate the first...i Inference is performed on the frames. The projection point is determined. p ij Does it fall within a directed bounding box of category "Antenna" or "AAU"? For each point... P j Maintain a vote counter.

[0100] Fine segmentation: The antenna identification boxes in the image are labeled based on the prior condition of the number of antennas. The identification boxes are projected back into the point cloud. The intersection of different images is used, combined with the RANSAC plane fitting method to obtain the antenna panel point cloud.

[0101] Clustering Separation: Set the effective voting rate threshold to T. Perform DBSCAN clustering on the set of high-vote points, setting the neighborhood radius and minimum number of points. The clustering results label the individual antennas as Ant_1, Ant_2, ..., Ant_n. Simultaneously, obtain the point cloud region of the support structure immediately adjacent to the back of each antenna through differential calculation.

[0102] 5. Automated parsing of attitude parameters: Taking the identified Ant_1 individual point cloud as an example, the calculation process is as follows: Main radiating surface identification: A multi-plane extraction strategy is adopted (inner points are removed after each extraction). The main radiating plane that conforms to the antenna structure characteristics is selected from the point cloud first according to the size and shape constraints of the point set.

[0103] Normal vector pointing correction: Construct a spatial reference vector pointing from the bracket to the antenna, and use its direction consistency with the initial normal vector to check, and automatically flip the normal vector pointing to the bracket side to the opposite direction.

[0104] Attitude parameter calculation: The corrected normal vector is geometrically correlated with the vertical axis of the scene to obtain the pitch angle (outward pitch is positive); at the same time, the normal vector is projected onto the horizontal plane to obtain the directional angle between the horizontal projection and the zero-degree reference direction, and the known reference azimuth is fused to obtain the absolute azimuth angle of the antenna.

[0105] 6. Result storage and post-processing: After execution, the system stores the local point cloud of each antenna, the corresponding 2D verification map, and a result.json file containing the ID, elevation angle, and azimuth angle to the task directory. Users can perform a final review of the analysis results through a visual interface.

[0106] The above detailed embodiments further illustrate the purpose, technical solution, and beneficial effects of the embodiments of this application. It should be understood that the above are merely specific embodiments of the embodiments of this application and are not intended to limit the protection scope of the embodiments of this application. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solutions of the embodiments of this application should be included within the protection scope of the embodiments of this application.

Claims

1. A method for base station antenna attitude analysis, characterized in that, include: Collect video streams from base station inspections; The video stream is subjected to equal-interval frame extraction to generate a two-dimensional image sequence; Using the structure-reconstruction-motion algorithm, feature extraction and matching are performed on the two-dimensional image sequence, the camera intrinsic parameter matrix and camera pose of each frame in the two-dimensional image sequence are calculated, and an initial three-dimensional Gaussian scene is constructed. Based on the camera intrinsic parameter matrix and camera pose, the initial 3D Gaussian scene is optimized by backpropagation using the photometric loss function to generate the target 3D Gaussian scene. Based on the opacity and distribution density of Gaussian points in the target 3D Gaussian scene, Poisson surface reconstruction or extraction of a preset probability density section is performed to generate a dense point cloud; the dense point cloud restores the edge contour of the base station antenna plane and the geometric features of the base station antenna support. Based on the dense point cloud, the base station antenna point cloud and the corresponding base station antenna support point cloud are separated using the camera intrinsic parameter matrix and camera pose. Based on the point cloud of the base station antenna, a plane fitting process is performed to determine the main radiation plane of the antenna; Obtain the target unit normal vector of the antenna's main radiating plane; Based on the target unit normal vector, the attitude parameters of each base station antenna are determined; The attitude parameters include pitch angle and azimuth angle.

2. The base station antenna attitude analysis method according to claim 1, characterized in that, The steps of using the structure-of-motion (SOG) algorithm to extract and match features from the two-dimensional image sequence, calculate the camera intrinsic parameter matrix and camera pose of each frame in the two-dimensional image sequence, and construct an initial three-dimensional Gaussian scene include: The motion reconstruction structure algorithm is used to extract feature points from the two-dimensional image sequence; Based on the feature points, a sparse point cloud of the scene is constructed using feature matching, and the camera intrinsic parameter matrix and camera pose of each frame image relative to the scene coordinate system are calculated. An initial 3D Gaussian scene is constructed with the sparse point cloud of the scene as the centroid; each Gaussian point in the initial 3D Gaussian scene contains Gaussian properties; the Gaussian properties include: position, scaling length, rotation quaternion, and opacity.

3. The base station antenna attitude analysis method according to claim 1, characterized in that, The step of generating a target 3D Gaussian scene by backpropagating the initial 3D Gaussian scene based on the camera intrinsic parameter matrix and camera pose using a photometric loss function includes: Based on the camera intrinsic parameter matrix and camera pose, the Gaussian points are projected onto the image plane using differentiable rendering to generate a rendered image. The rendered image and the two-dimensional image sequence are compared to calculate the photometric loss; Based on the photometric loss, the Gaussian properties are optimized by backpropagation using the photometric loss function to generate the target 3D Gaussian scene.

4. The base station antenna attitude analysis method according to claim 3, characterized in that, The step of generating a target 3D Gaussian scene by backpropagating and optimizing the Gaussian properties using the photometric loss function based on the photometric loss includes: Based on the photometric loss, the Gaussian property is optimized by backpropagation using the photometric loss function, and three scaling axes are calculated for each Gaussian point. A penalty term is applied to the scaling axis, causing the scaling ratio of one of the scaling axes to approach zero, while the remaining two scaling axes expand in the local tangent plane, so that the Gaussian points form a thin sheet distribution on the surface of the object to be reconstructed, and ensuring that the direction of the flat axis of the Gaussian points is the same as the direction of the local geometric normal of the object to be reconstructed. After the alignment optimization is completed on the surface of the object to be reconstructed, the target 3D Gaussian scene is generated.

5. The base station antenna attitude analysis method according to claim 1, characterized in that, The step of separating the base station antenna point cloud and the corresponding base station antenna support point cloud based on the dense point cloud, using the camera intrinsic parameter matrix and camera pose, includes: Using a deep learning detection model, target detection is performed on the two-dimensional image sequence to obtain a directed bounding box containing the base station antenna and the bracket; the directed bounding box includes: the center coordinates, length and width dimensions, and rotation angle information of the base station antenna and the bracket; Using the camera intrinsic parameter matrix and extrinsic pose, each three-dimensional coordinate point in the dense point cloud is back-projected into the two-dimensional image sequence of the corresponding viewpoint; The number of times each three-dimensional coordinate point falls within the corresponding directed bounding box is determined. If the number of times is greater than a preset number, the corresponding target three-dimensional coordinate point is retained; if the number of times is less than or equal to the preset number, the corresponding three-dimensional coordinate point is discarded. A spatial clustering algorithm is performed on the target three-dimensional coordinate points to separate the base station antenna point cloud and the corresponding base station antenna support point cloud.

6. The base station antenna attitude analysis method according to claim 5, characterized in that, The step of performing a spatial clustering algorithm on the target three-dimensional coordinate points to separate the base station antenna point cloud and the corresponding base station antenna support point cloud includes: A spatial clustering algorithm is performed on the target three-dimensional coordinate points to determine the point cloud of the base station antenna components and the number of base station antenna components based on the Euclidean distance; Based on the directed bounding box, the base station antenna identification box is determined; The base station antenna identification frame in the two-dimensional image sequence is labeled; The base station antenna identification box after the label is reprojected back to the target three-dimensional Gaussian scene using the camera intrinsic parameter matrix and extrinsic parameter pose, and the base station antenna point cloud is extracted using a multi-view and plane fitting strategy. Based on the point cloud of the base station antenna component and the point cloud of the base station antenna, Boolean difference operation is performed to separate the corresponding point cloud of the base station antenna support.

7. The base station antenna attitude analysis method according to claim 1, characterized in that, The step of performing plane fitting processing based on the base station antenna point cloud to determine the main radiating plane of the antenna includes: A candidate planar parameter set for the point cloud of a single base station antenna is determined using a random sample consensus algorithm; the candidate planar parameter set is used to separate the front, side, and noise of the base station antenna. Calculate the number of points in each candidate plane in the candidate plane parameter set; Project the three-dimensional point set of each candidate plane onto its corresponding fitting plane, and convert the three-dimensional point set into two-dimensional point coordinates; Based on the two-dimensional point coordinates, the projected area of ​​the circumscribed polygon corresponding to the three-dimensional point set is calculated using the convex hull. Based on the number of points and the projected area, a scoring function is used to determine the score of each candidate plane, and the plane with the highest score that meets the antenna aspect ratio characteristics is extracted as the main radiation plane of the antenna.

8. The base station antenna attitude analysis method according to claim 7, characterized in that, The scoring function is: Score i = α · f ( N i )+ β · g ( S i )； In the formula, f (·)and g (·) represents the normalization processing function; α and β Preset weighting coefficients; N i Number of points; S i This represents the projected area.

9. The base station antenna attitude analysis method according to claim 1, characterized in that, The step of obtaining the target unit normal vector of the antenna's main radiating plane includes: Using the point cloud of the base station antenna bracket, the representative location point of the bracket is determined; Using the point cloud of the base station antenna, the location point represented by the antenna is determined; Based on the location point represented by the bracket and the location point represented by the antenna, a reference direction vector is constructed pointing from the location point represented by the bracket to the location point represented by the antenna; the reference direction vector represents the spatial layout relationship of the base station antenna relative to the base station antenna bracket. Obtain the initial unit normal vector of the main radiating plane of the antenna; When the direction of the initial unit normal vector is consistent with the direction of the reference direction vector, the initial unit normal vector is reversed to obtain the target unit normal vector.

10. The base station antenna attitude analysis method according to claim 1, characterized in that, The steps for determining the attitude parameters of each base station antenna based on the target unit normal vector include: Define a vertical reference axis; the direction of the vertical reference axis is consistent with the direction of gravity; Calculate the spatial angle between the target unit normal vector and the vertical reference axis; The spatial angle is converted into a pitch angle according to a preset conversion relationship; The target unit normal vector is projected onto a horizontal reference plane orthogonal to the vertical reference axis to obtain a horizontal projection vector. Calculate the directional angle between the horizontal projection vector and the preset horizontal zero-degree reference direction in the scene; Calculate the radian deviation between the base station antenna projection vector and the relative reference vector based on a reference antenna with a known preset azimuth angle; Using the radian deviation, absolute angle compensation is performed on the directional included angle to obtain the absolute azimuth angle of the base station antenna.

Images