Visual robot identification and hooking method and system suitable for overhead lines

Abstract

The application belongs to the field of overhead line hanging, and discloses a visual robot identification and hanging method and system suitable for overhead lines, which comprises the following steps: generating point cloud data by fusing multiple exposure images collected by vision through TSDF; extracting the main shaft and radius features to construct a cylindrical template, separating the line area and obtaining the initial pose; combining geometric constraints and normal consistency to iteratively optimize the accurate pose through the ICP algorithm; and establishing a multi-target model to solve the optimal attitude and drive the mechanical arm to complete the hanging. The application realizes high-precision geometric identification and positioning of overhead lines, improves the efficiency and operation safety of automatic hanging, and effectively solves the problem of manual hanging in complex environments.

Description

Technical Field

[0001] This invention relates to the field of overhead line splicing technology, and in particular to a visual robot recognition splicing method and system applicable to overhead lines. Background Technology

[0002] During power system maintenance, grounding wire connection is necessary to release residual charge and induced voltage on overhead lines. Current technology primarily relies on manual labor using aerial work platforms to complete this connection, which presents the following significant problems: First, workers face risks of falls from heights and accidental electric shock; second, the grounding wire is heavy and the operation is cumbersome, making it easy to become fatigued and make mistakes while holding it at heights; third, there is a lack of condition monitoring methods, making it difficult to detect potential problems such as loose connections or poor contact in a timely manner; and fourth, each operation is time-consuming and inefficient. In response to these problems with manual grounding methods, the industry has conducted some research on automated grounding technologies.

[0003] Existing automatic splicing solutions mostly employ general template matching or single-frame point cloud registration methods to identify and locate overhead lines. However, due to the strong reflective properties of the overhead line surface material, under actual field operating conditions, the point cloud data acquired by the sensing system often suffers from problems such as glare interference, missing point clouds, or partial occlusion. This makes it difficult for the aforementioned methods to meet the requirements for positioning accuracy and algorithm robustness. It is precisely because existing automatic splicing solutions lack accuracy and adaptability under complex sensing conditions that their application in actual power plant field operations still faces significant challenges and cannot effectively replace manual labor.

[0004] Therefore, both existing manual splicing methods and existing automatic splicing research schemes have obvious limitations, and there is a need for an automatic overhead line splicing technology that is highly safe, reliable in operation, and adaptable to complex field environments. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a visual robot identification and attachment method and system suitable for overhead power lines. By employing multimodal visual perception and robust point cloud registration, it solves the issues of insufficient positioning accuracy and poor robustness in existing automatic attachment schemes under conditions such as strong surface reflection, complex backgrounds, and missing local point clouds on overhead power lines. Furthermore, by constructing a multi-objective optimization mechanism that integrates robotic arm reachability analysis, joint constraints, and collision detection, it addresses the lack of automatic attachment posture planning capabilities for complex constraints in existing attachment schemes. This invention is achieved through the following scheme: A vision robot recognition and attachment method for overhead power lines includes the following steps: S1: The robotic arm drives a binocular camera to acquire multi-exposure depth maps and confidence maps of overhead lines under preset shooting poses. Through confidence, inter-frame registration, and TSDF voxel weighted average fusion, a complete point cloud of the overhead lines is generated. S2: Extract the principal axis direction and combine it with radius interval filtering and one-dimensional projection clustering to separate and identify the overhead line point cloud region from the fused point cloud, and construct the cylindrical template of the overhead line; S3: Based on the pre-established overhead line cylindrical template, the template axis is spatially aligned with the measured overhead line axis. The initial position and attitude of the overhead line in three-dimensional space are obtained by fitting the cross-sectional radius and solving the least square translation. S4: Based on the ICP registration method, cylindrical geometric constraints, normal consistency constraints and robust kernel functions are introduced to iteratively optimize the pose of the overhead line point cloud, thereby obtaining the accurate pose of the overhead line; S5: Based on the local geometric structure of the overhead line, analyze the available hanging area, and establish a multi-objective optimization model by combining the movement range of the robotic arm, the joint range limitation, and the collision detection constraint. Solve the optimal hanging posture through multi-objective optimization, and drive the robotic arm to complete the grounding wire hanging operation.

[0006] Furthermore, the collection of multi-exposure depth maps and confidence maps for overhead power lines specifically includes: The robotic arm moves the binocular camera to a preset shooting pose and continuously acquires multiple frames of depth maps and corresponding confidence maps by setting different exposure times. Using a feature-point-based inter-frame registration method, the extrinsic parameters between adjacent depth frames are calculated, and all depth maps are aligned to a unified coordinate system. The depth maps are then weighted pixel by pixel based on the confidence map, reducing the weight of reflective and noisy areas. The aligned depth frames are then input into the TSDF fusion module, and a fusion voxel is constructed using a truncation distance function to generate a 3D point cloud of the overhead line.

[0007] Furthermore, TSDF fusion uses the following expression for updating: ; ; The local TSDF value for the k-th observation is: ; Let v be the depth measured by projecting v onto the k-th frame. Let be the measured depth from the center of a single precursor voxel to the camera, be the truncation distance, trunc(.) indicates truncation to [-1,1], and the weights are... This can be taken as the overall confidence level of the observation. This is the upper limit of the weight, used to stabilize the value.

[0008] Furthermore, separating and identifying overhead line point cloud regions specifically includes: Principal component analysis is performed on the fused point cloud to obtain the first principal direction vector of the point cloud, which is used as the initial direction of the overhead line axis. Given an allowable radius range, the point cloud is filtered to remove points that exceed the radius range. The filtered point set is projected onto the axis for one-dimensional projection clustering to obtain several clusters. The morphological connectivity of the clustering results is judged, and the cluster with the largest length and the most complete structure is retained as the overhead line point cloud region.

[0009] Furthermore, for the j-th cluster, calculate the length of its axial projection coordinate set, set a minimum effective length threshold, and filter candidate clusters that meet the conditions; select the cluster corresponding to the maximum axial length among the candidate clusters as the main body of the overhead line; construct an undirected adjacency graph, calculate all connected subgraphs, and find the connected component with the most nodes; take the set of points corresponding to the maximum connected component as the final point cloud of the main body of the overhead line.

[0010] Furthermore, aligning the axis of the template cylinder with the main axis of the point cloud, based on the main axis direction, specifically includes: The point cloud is fitted with a circular cross-section in a plane perpendicular to the axis. The position of the center of the circle is estimated by least squares, and the spatial translation of the template is solved. The orientation alignment and the center translation are combined to generate the initial pose matrix of the overhead line.

[0011] Furthermore, the ICP registration method includes: By fixing the radius of the overhead line, the distance from each point to the cylindrical surface is optimized; the angle error between the point cloud normal and the cylinder's outer normal is calculated, and the normal deviation is incorporated into the ICP weight; the Huber kernel function is used to suppress long-distance errors; and the pose update is solved using Levenberg-Marquardt iteration.

[0012] Furthermore, the Levenberg-Marquardt iterative solution for pose update uses the following expression: Transform parameters using parameter vectors This indicates that a residual function is constructed for the i-th observation: ; Linearization in the current estimate Obtained from: ; Where J is the Jacobian matrix (partial derivatives of the residuals with respect to the parameters), updated using the Levenberg-Marquardt method: ; Where: W is the diagonal weight matrix. The damping factor, Update parameters to the identity matrix: ; Repeat the iteration until convergence.

[0013] Furthermore, solving for the optimal attachment posture specifically includes: Based on the precisely registered cylindrical model, the area where the hanger can stably clamp is calculated, including the optimal clamping height, contact normal direction, and accessible direction. With the attachment point as the origin, the overhead line normal as the X-axis, and the insertion direction as the Z-axis, a local coordinate system for attachment is constructed. Combined with the DH model of the robotic arm, the reachability of the local coordinate system for attachment is solved, and postures that exceed the joint range or have singular points are excluded. The overhead line, robotic arm, connector, and on-site environment model are input into the collision detection module, and a safe distance is set to avoid interference. The optimal connection posture and path are solved through multi-objective optimization. The robotic arm completes the automatic connection of the overhead line grounding wire according to the planned trajectory.

[0014] A vision robot recognition and attachment system suitable for overhead power lines includes: The acquisition module is used to drive the binocular camera to acquire multi-exposure depth maps and confidence maps of overhead lines under preset shooting poses, and generate complete point clouds of overhead lines. The identification module is used to extract the main axis direction and combine it with radius interval filtering and one-dimensional projection clustering to separate and identify the overhead line point cloud region from the fused point cloud and construct the cylindrical template of the overhead line. The initial positioning module is used to spatially align the axis of the pre-established overhead line cylindrical template with the axis of the measured overhead line. By fitting the cross-sectional radius and solving the least squares translation, the initial position and attitude of the overhead line in three-dimensional space are obtained. The precise positioning module is used to iteratively optimize the pose of the overhead line point cloud based on the ICP registration method, by introducing cylindrical geometric constraints, normal consistency constraints and robust kernel functions, so as to obtain the precise pose of the overhead line. The execution module is used to analyze the attachable area based on the local geometry of the overhead line, and establish a multi-objective optimization model by combining the movement range of the robotic arm, the joint range limit, and the collision detection constraint. The optimal attachment posture is solved through multi-objective optimization, and the robotic arm is driven to complete the grounding wire attachment operation.

[0015] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. Unlike existing solutions that use single-frame point cloud acquisition, this invention overcomes the problems of depth data distortion and point cloud loss caused by strong reflections on the metal surface of overhead lines by combining multi-exposure depth maps with confidence-weighted fusion and TSDF voxel-weighted average fusion, thereby improving the quality and completeness of point clouds.

[0016] 2. Unlike existing general point cloud segmentation methods, this invention is specifically designed for the slender cylindrical structure of overhead lines, combining principal axis extraction, radius screening, and projection clustering. It can still identify overhead line areas under complex background and partial occlusion conditions, thus improving the robustness of extraction.

[0017] 3. Unlike existing general template matching or standard ICP methods, this invention introduces cylindrical geometric constraints and normal consistency constraints into the ICP optimization framework, and combines robust kernel functions to suppress reflective noise interference, thereby achieving sub-millimeter-level precision registration for overhead line structures.

[0018] 4. Unlike existing posture planning methods that only consider a single constraint, this invention integrates robotic arm accessibility analysis, joint angle constraints, and workspace collision detection into a unified multi-objective optimization model, which automatically solves for the optimal attachment posture that satisfies all safety constraints. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the connection state of the connector of the present invention; Figure 3 This is a schematic diagram of the point cloud data collected by the present invention; Figure label: 1-Hanger; 2-Binocular camera movement; 3-Robotic arm; Detailed Implementation All features disclosed in this specification, or all steps in all disclosed methods or processes, may be combined in any way, except for mutually exclusive features and / or steps.

[0020] Any feature disclosed in this specification (including any appended claims and abstract) may be replaced by other equivalent or similar features, unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is merely one example of a series of equivalent or similar features.

[0021] In the description of this invention, it should be understood that the terms "upper", "lower", "left", "right", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0022] Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature.

[0023] like Figure 1 As shown, a vision robot identification and attachment method suitable for overhead power lines includes the following steps: S1: Overhead line point cloud data acquisition and fusion.

[0024] The robotic arm moves the binocular camera to a preset shooting pose to acquire multi-exposure depth maps and confidence maps of the overhead line area to overcome the problem of missing points in the point cloud caused by reflections from the metal surface of the overhead line. Multi-frame depth maps are aligned through inter-frame registration, and 3D point cloud data of the overhead line is generated using confidence-weighted fusion or TSDF (truncated signed distance function) fusion methods. The fused point cloud has low noise and sufficient compensation for occlusion areas, providing reliable input for subsequent positioning and trajectory planning. By acquiring multi-exposure depth maps and using TSDF fusion, the problem of missing depth caused by uneven illumination under a single exposure is effectively solved, improving the integrity and accuracy of the point cloud.

[0025] S2: Extraction of overhead line area.

[0026] In step S2, the point cloud of the overhead line area is successfully extracted by extracting the principal axis direction and combining it with radius interval filtering and one-dimensional projection clustering. Based on this, a cylindrical template of the overhead line is further constructed to provide a geometric reference for subsequent coarse positioning. By comparing the actual point cloud with the template cylinder, we can obtain the initial attitude estimate of the overhead line and provide initial values ​​for subsequent fine registration.

[0027] S3: Initial positioning of overhead lines.

[0028] After obtaining the point cloud of the overhead line, a cylindrical template with the same dimensions as the actual overhead line is constructed. By aligning the axis, the cylindrical axis of the template is made consistent with the main axis of the measured overhead line point cloud. The initial position and attitude of the overhead line in three-dimensional space are obtained by fitting the cross-sectional radius and solving the least squares translation. This achieves coarse positioning, and the coarse positioning result is used as the initial value for subsequent fine registration, which can significantly improve the convergence speed and stability of fine registration.

[0029] S4: Overhead line precision alignment.

[0030] The structured ICP registration method is used to optimize the pose and position of the overhead line point cloud by introducing cylindrical geometric constraints, normal consistency constraints and robust kernel functions, thereby obtaining the accurate pose of the overhead line. In steps S3 and S4, the accuracy of overhead line pose estimation is significantly improved by combining cylindrical templates, least squares translation, and the ICP algorithm, and by utilizing cylindrical geometric constraints and normal consistency constraints.

[0031] S5: Automatic attachment posture planning.

[0032] After determining the precise pose of the overhead line, the system analyzes the hookable area based on the local geometry of the overhead line, including the position of the hook clamping point, the surface normal direction, and the appropriate approach direction. Based on this information, a local coordinate system for hooking is constructed, and a multi-objective optimization model is established by combining the kinematic constraints of the robotic arm, the joint range limitation, and the collision objects in the field environment. By solving the optimization model, a hooking posture that meets the requirements of safe distance, accessibility and directional consistency is generated. The robotic arm then executes the planned posture and path to realize the automatic and safe hooking of the grounding wire to the overhead line.

[0033] Furthermore, as a specific embodiment, step S1 is implemented as follows: S11: The robotic arm moves the binocular camera to the preset shooting position, such as... Figure 2 As shown, by setting different exposure times, multiple frames of depth maps and corresponding confidence maps are continuously acquired; ; in For pixels The corresponding depth value, For the camera intrinsic parameter matrix, For 3D points in camera coordinate system; low exposure suppresses strong reflective areas, and high exposure supplements dark area information, making the overall depth data more complete; S12: Using a feature-point-based inter-frame registration method, calculate the extrinsic parameters between adjacent depth frames and align all depth maps to a unified coordinate system; let the rotation and translation matrices of the i-th frame be... Then the three-dimensional point corresponding to the pixel in the i-th frame In the reference coordinate system: ; S13: Confidence-weighted fusion preprocessing, which performs pixel-by-pixel weighting on the depth map based on the confidence map, reducing the weight of reflective and noisy areas to make the fusion input more stable; In each frame of the depth map, the camera simultaneously outputs a confidence map to characterize the depth reliability of each pixel; for a pixel position x=(u,v), its confidence is defined as: ; For pixel position x=(u,v), the depth of multiple frames By confidence level Weighted fusion as fusion depth : ; Here, we can simply take the confidence level of the weights: ; S14: Point cloud fusion based on TSDF or voxel averaging. Aligned depth frames are input into the TSDF fusion module. A fused voxel is constructed using a truncation distance function, generating a continuous, low-noise, and detailed 3D point cloud of the overhead line. A commonly used TSDF update formula is: For a voxel v, its TSDF value F and weight W are updated using the k-th observation. ; ; The local TSDF value for the k-th observation is: ; Let v be the depth measured by projecting v onto the k-th frame. is the measured depth from the center of a single precursor voxel to the camera, is the truncation distance, and trunc(.) indicates truncation to [-1,1]; weights This can be taken as the overall confidence level of the observation. This is the upper limit of the weight, used to stabilize the value.

[0034] Furthermore, step S2 specifically includes: S21: Extraction of principal axis direction of point cloud; perform PCA (principal component analysis) on the fused point cloud to obtain the first principal direction vector of the point cloud, which is used as the initial direction of the axis of the overhead line; Given the fused point cloud First, calculate the centroid of the point cloud: ; in Represent the 3D coordinates of the k-th point in the point cloud; calculate the covariance matrix C: ; Perform eigenvalue decomposition on the covariance matrix to obtain pairs of eigenvalues ​​and eigenvectors. (according to (Sort) ; Take the eigenvector corresponding to the largest eigenvalue The main axis direction of the overhead line: ; 'a' is a unit vector that can be used for subsequent axis definition and projection; S22: Point cloud filtering based on radius interval, assuming the overhead line axis passes through the point Let the axial direction be 'a'. The radial distance from any point P to the axial direction is defined as: ; Given the allowable radius range [ If the filter condition is: P is retained if and only if After filtering, the point set is obtained: ; Based on the typical radius of overhead lines (e.g., 15–25 mm), calculate the distance from the point to the axis, and remove point clouds that exceed the radius range, excluding background structures such as walls, line clips, and insulators; S23: One-dimensional projection clustering along the overhead line direction, projecting each point in the selected point set onto the axial direction to obtain one-dimensional coordinates: ; We obtain the one-dimensional projection set S={ DBSCAN can be used to segment continuous segments in one dimension. Let the parameters (neighborhood radius) and minPts (minimum number of points) be used, then the clustering determination is as follows: If a point is a core point and its neighborhood contains at least minPts points, then the connected core points and their density-reachable points are merged into a cluster. Let the obtained clusters be { The original 3D point subset corresponding to each cluster is} ; S24: Extract point cloud data of local areas of overhead lines, perform morphological connectivity judgment on clustering results, retain the cluster with the largest length and the most complete structure as the point cloud region of overhead lines, and finally obtain clean and well-defined main data of overhead lines.

[0035] S241. Cluster axial length calculation, for the j-th cluster (corresponding to the original three-dimensional point set) Calculate its axial projection coordinate set Length: ; S242. Select candidate clusters and set the minimum effective length threshold: ; Will satisfy The clusters are used as a candidate cluster set. If no clusters meet the condition, it is considered that the data collection is insufficient and the process is returned to the re-collection process. S243. Select the primary cluster: Choose the cluster with the maximum axial length from the candidate clusters. As the main body of the overhead line; ; The corresponding three-dimensional point set is ; S244. Morphological connectivity assessment: To remove isolated noise points and ensure the connectivity of the point cloud, an undirected adjacency graph is constructed. The vertex set V is The points in the middle (by index number), when two points ,satisfy ; Then add an edge (u, v) to graph E, calculate all connected subgraphs (connected components) for graph G, and find the connected component with the most nodes. ; S245, Output the final point cloud of the overhead line main body, including the maximum connected components. The corresponding set of points serves as the final point cloud of the overhead power line: ; Furthermore, step S3 specifically includes: S31: Construct a cylindrical template for the overhead line. Based on the actual dimensions (radius and length range) of the overhead line, establish a cylindrical template model to provide a geometric reference for subsequent coarse positioning.

[0036] Based on the known actual dimensions of the overhead line, establish a template cylinder: radius R, axial length range [ The template cylinder can be represented as a parametric placement process: ; in The d-axis represents the starting point of the template axis and is a unit vector in the direction of the axis. satisfy ; S32: Align the template axis with the point cloud axis. Based on the main direction a of the overhead line point cloud extracted in step S2, align the axial direction d0 of the template cylinder with the main axis of the point cloud. The rotation axis and rotation angle during the alignment process are: ; The rotation matrix is ​​obtained using Rodrigues' formula: ; in For antisymmetric matrices: ; S33: Translation solution based on cross-section fitting. Perform cross-sectional circular fitting on the point cloud in a plane perpendicular to axis a, estimate the center position of the circle by least squares, and solve for the spatial translation of the template.

[0037] S34: Generate the coarse positioning pose of the overhead line, combine the direction alignment and the center translation to generate the initial pose matrix of the overhead line, providing high-quality initial values ​​for the fine registration in step S4.

[0038] Furthermore, step S4 specifically includes: S41: Construct a structured ICP energy model, and incorporate the point-to-model distance, cylindrical geometric constraints, and normal consistency constraints into the optimization objective function.

[0039] Let the observed point cloud be Transformation parameter is rotation O(3) and translation t The transformed points are: ; The overall optimization objective (with robust core) can be written as: ; in: This represents the geometric distance from the point to the cylindrical model M. Represents a measure of normal uniformity. , >0 indicates the weights of the distance and normal terms, respectively; these values ​​are set based on experience in this project. , =0.5, () represents a robust kernel function used to suppress outliers.

[0040] S42: Cylindrical geometry constraint application. By fixing the radius of the overhead line, the distance from each point to the cylindrical surface is optimized to ensure that the registration process remains robust to the actual geometry of the overhead line.

[0041] Let the axis of the cylinder be... (where a is the unit axial vector), the radius of the overhead line is r, and the radial distance from the point to the axis is: ; The signed distance from a point to a cylindrical surface is defined as: ; The energy function uses a squared term: ; S43: Normal consistency check. Calculate the angle error between the point cloud normal and the cylinder's outer normal, and incorporate the normal deviation into the ICP weight to weaken the abnormal normal caused by the reflective area.

[0042] Set point cloud The normal direction of the place is (during transformation) If the normal vector of the cylindrical model at the corresponding point is given, then the normal consistency error can be defined as: ; Outer normal of cylindrical model It can be obtained by normalizing the radial direction from the point to the axis, that is: ; S44: Robust kernel function handles outliers. The Huber kernel function is used to suppress long-range errors, so that abnormal point clouds such as local occlusion and depth jumps will not affect the overall registration results.

[0043] ; in Represents the residual. The threshold value can be set to an empirical value of 1.

[0044] S45: Iterative optimization to obtain accurate pose, using LM (Levenberg-Marquardt) iterative solution to solve pose update.

[0045] Transform parameters using parameter vectors This indicates that a residual function is constructed for the i-th observation: ; Linearization in the current estimate Obtained from: ; Where J is the Jacobian matrix (partial derivatives of the residuals with respect to the parameters), updated using the Levenberg-Marquardt method: ; Where: W is the diagonal weight matrix. The damping factor, Update parameters to the identity matrix: ; Repeat the iteration until convergence.

[0046] Furthermore, step S5 specifically includes: S51: Analysis of the hanging area of ​​overhead line. Based on the precisely registered cylindrical model, the area where the hanger can be stably clamped is calculated, including the optimal clamping height, contact normal direction and accessible direction. Definition and calculation of stable clamping area: Based on the cylindrical model after fine registration, the area on the surface of the overhead line that can be stably clamped is first calculated. This area takes into account the radius change of the overhead line, the surface smoothness and the contact area of ​​the clamp, so as to ensure that the clamp can apply force evenly and avoid connection failure due to unstable clamping. The stable clamping area is usually selected in the middle of the overhead line to ensure that the grounding wire can be evenly hung on the surface of the overhead line.

[0047] Calculation of optimal clamping height: Based on the radius and geometry of the overhead line, the optimal clamping height of the gripper is calculated. This height is the height at which the gripper can most effectively and stably clamp the overhead line. It is usually near the center height of the overhead line. By analyzing the workspace of the robotic arm and the physical dimensions of the gripper, the height that can minimize interference and provide the best clamping force is calculated.

[0048] Contact normal direction: The contact normal direction is determined by the normal vector of the overhead line surface. The calculated normal vector ensures that the contact direction of the clamp is consistent with the normal of the overhead line surface, thereby optimizing clamping stability and preventing the clamp from slipping or becoming unstable.

[0049] Accessibility calculation: Accessibility refers to the position from which the gripper can approach the surface of the overhead line from a safe and unobstructed angle. The calculation of this direction takes into account the geometry of the overhead line and the operating space of the robotic arm to ensure that the gripper can smoothly enter the attachment position without collision.

[0050] These calculations ensure that the gripper can stably and accurately grasp the overhead line, providing a precise reference for subsequent automatic splicing.

[0051] S52: Establish a local coordinate system for the connection, with the connection point as the origin, the overhead line normal as the X-axis, and the insertion direction as the Z-axis, to construct a dedicated local coordinate system for the connection.

[0052] S53: Reachability analysis of the robotic arm. Combining the DH model of the robotic arm, the reachability of the local coordinate system is solved, and postures that are outside the joint range or have singular points are excluded.

[0053] S54: Collision detection and safety distance constraints. Input the overhead line, robotic arm, connector, and on-site environment model into the collision detection module and set a safety distance (e.g., ≥30 mm) to avoid interference.

[0054] S55: Multi-objective optimization generates the attachment posture. Under the premise of satisfying reachability, orientation consistency, and safe distance, the optimal attachment posture and path are solved through multi-objective optimization, including: Approach path, hook-in insertion path, clamping / fastening action path, and withdrawal path.

[0055] S56: Perform the connection operation. The robotic arm automatically connects the overhead line grounding wire according to the planned trajectory.

[0056] Figure 2 This is a schematic diagram of the connection, showing how the connector is fixed on the overhead line; Figure 3This diagram illustrates point cloud acquisition by a camera. It shows the robotic arm body, the attachment carried at the end of the robotic arm, the binocular camera mounted on the attachment, and the effective field of view coverage of the camera in its current pose. This reflects the spatial relationship between the components and the camera's sensing range during the point cloud acquisition stage.

[0057] This invention uses a robotic arm to automate the grounding wire connection process, replacing traditional manual high-altitude operations and avoiding the risks of falls and electric shocks faced by workers. The robotic arm performs the handling, alignment, and connection of the grounding wire, reducing the physical burden on personnel and minimizing the probability of operational errors caused by fatigue.

[0058] To address challenges such as surface reflection, complex backgrounds, and missing point clouds on overhead power lines, this invention employs multi-exposure fusion and cylindrical geometry prior extraction techniques, combined with a structured ICP registration algorithm, to locate overhead power lines, thereby improving positioning accuracy and algorithm stability under complex sensing conditions. Based on positioning, the attitude planning module integrates constraints such as robotic arm accessibility, joint limits, and collision detection to solve for the trajectory, automatically generating a safe attachment posture and reducing the risk of equipment collisions or attachment failures due to improper planning.

[0059] This solution reduces auxiliary operations such as scaffolding erection and high-altitude climbing, shortens the time for a single splicing operation, and improves the reliability of splicing operations and overall maintenance efficiency.

[0060] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A visual robot recognition and attachment method suitable for overhead power lines, characterized in that, Includes the following steps: S1: The robotic arm drives a binocular camera to acquire multi-exposure depth maps and confidence maps of overhead lines under preset shooting poses. Through confidence, inter-frame registration, and TSDF voxel weighted average fusion, a complete point cloud of the overhead lines is generated. S2: Extract the principal axis direction and combine it with radius interval filtering and one-dimensional projection clustering to separate and identify the overhead line point cloud region from the fused point cloud, and construct the cylindrical template of the overhead line; S3: Based on the pre-established overhead line cylindrical template, the template axis is spatially aligned with the measured overhead line axis. The initial position and orientation of the overhead line in three-dimensional space are obtained by fitting the cross-sectional radius and solving the least squares translation problem. S4: Based on the ICP registration method, cylindrical geometric constraints, normal consistency constraints, and robust kernel functions are introduced to iteratively optimize the pose of the overhead line point cloud, thereby obtaining the accurate pose of the overhead line; S5: Based on the local geometric structure of the overhead line, analyze the available hanging area, and establish a multi-objective optimization model by combining the movement range of the robotic arm, the joint range limitation, and the collision detection constraint. Solve the optimal hanging posture through multi-objective optimization, and drive the robotic arm to complete the grounding wire hanging operation.

2. The visual robot recognition and attachment method for overhead power lines according to claim 1, characterized in that, The acquisition of multi-exposure depth maps and confidence maps for overhead power lines specifically includes: The robotic arm moves the binocular camera to a preset shooting pose and continuously acquires multiple frames of depth maps and corresponding confidence maps by setting different exposure times. Using a feature-point-based inter-frame registration method, the extrinsic parameters between adjacent depth frames are calculated, and all depth maps are aligned to a unified coordinate system. The depth maps are then weighted pixel by pixel based on the confidence map, reducing the weight of reflective and noisy areas. The aligned depth frames are then input into the TSDF fusion module, and a fusion voxel is constructed using a truncation distance function to generate a 3D point cloud of the overhead line.

3. The visual robot recognition and attachment method for overhead power lines according to claim 2, characterized in that, TSDF fusion uses the following expression for updating: ； ； The local TSDF value for the k-th observation is: ； Let v be the depth measured by projecting v onto the k-th frame. Let be the measured depth from the center of a single precursor voxel to the camera, be the truncation distance, trunc(.) indicates truncation to [-1,1], and the weights are... This can be taken as the overall confidence level of the observation. This is the upper limit of the weight, used to stabilize the value.

4. The visual robot recognition and attachment method for overhead power lines according to claim 1, characterized in that, Separating and identifying overhead line point cloud regions specifically includes: Principal component analysis is performed on the fused point cloud to obtain the first principal direction vector of the point cloud, which is used as the initial direction of the overhead line axis. Given an allowable radius range, the point cloud is filtered to remove points that exceed the radius range. The filtered point set is projected onto the axis for one-dimensional projection clustering to obtain several clusters. The morphological connectivity of the clustering results is judged, and the cluster with the largest length and the most complete structure is retained as the overhead line point cloud region.

5. A visual robot recognition and attachment method for overhead power lines according to claim 4, characterized in that, For the j-th cluster, calculate the length of its axial projection coordinate set, set a minimum effective length threshold, and filter candidate clusters that meet the conditions; select the cluster with the maximum axial length as the main body of the overhead line from the candidate clusters; construct an undirected adjacency graph, calculate all connected subgraphs, and find the connected component with the most nodes; take the set of points corresponding to the maximum connected component as the final point cloud of the main body of the overhead line.

6. The visual robot recognition and attachment method for overhead power lines according to claim 1, characterized in that, Aligning the axis of the template cylinder with the main axis of the point cloud, based on the main axis direction of the point cloud, specifically includes: The point cloud is fitted with a circular cross-section in a plane perpendicular to the axis. The position of the center of the circle is estimated by least squares, and the spatial translation of the template is solved. The orientation alignment and the center translation are combined to generate the initial pose matrix of the overhead line.

7. A visual robot recognition and attachment method for overhead power lines according to claim 1, characterized in that, The ICP registration method includes: By fixing the radius of the overhead line, the distance from each point to the cylindrical surface is optimized; the angle error between the point cloud normal and the cylinder's outer normal is calculated, and the normal deviation is incorporated into the ICP weight; the Huber kernel function is used to suppress long-distance errors; and the pose update is solved using Levenberg-Marquardt iteration.

8. A visual robot recognition and attachment method for overhead power lines according to claim 7, characterized in that, The Levenberg-Marquardt iterative solution for pose update uses the following expression: Transform parameters using parameter vectors This indicates that a residual function is constructed for the i-th observation: ； Linearization in the current estimate Obtained from: ； Where J is the Jacobian matrix (partial derivatives of the residuals with respect to the parameters), updated using the Levenberg-Marquardt method: ； Where: W is the diagonal weight matrix. The damping factor, Update parameters to the identity matrix: ； Repeat the iteration until convergence.

9. A visual robot recognition and attachment method for overhead power lines according to claim 1, characterized in that, Solving for the optimal attachment posture specifically includes: Based on the precisely registered cylindrical model, the area where the hanger can stably clamp is calculated, including the optimal clamping height, contact normal direction, and accessible direction. With the attachment point as the origin, the overhead line normal as the X-axis, and the insertion direction as the Z-axis, a local coordinate system for attachment is constructed. Combined with the DH model of the robotic arm, the reachability of the local coordinate system for attachment is solved, and postures that exceed the joint range or have singular points are excluded. The overhead line, robotic arm, connector, and on-site environment model are input into the collision detection module, and a safe distance is set to avoid interference. The optimal connection posture and path are solved through multi-objective optimization. The robotic arm completes the automatic connection of the overhead line grounding wire according to the planned trajectory.

10. A visual robot recognition and attachment system suitable for overhead power lines, characterized in that, include: The acquisition module is used to drive the binocular camera to acquire multi-exposure depth maps and confidence maps of overhead lines under preset shooting poses, and generate complete point clouds of overhead lines. The identification module is used to extract the main axis direction and combine it with radius interval filtering and one-dimensional projection clustering to separate and identify the overhead line point cloud region from the fused point cloud and construct the cylindrical template of the overhead line. The initial positioning module is used to spatially align the axis of the pre-established overhead line cylindrical template with the axis of the measured overhead line. By fitting the cross-sectional radius and solving the least squares translation, the initial position and attitude of the overhead line in three-dimensional space are obtained. The precise positioning module is used to iteratively optimize the pose of the overhead line point cloud based on the ICP registration method, by introducing cylindrical geometric constraints, normal consistency constraints and robust kernel functions, so as to obtain the precise pose of the overhead line. The execution module is used to analyze the attachable area based on the local geometry of the overhead line, and establish a multi-objective optimization model by combining the movement range of the robotic arm, the joint range limit, and the collision detection constraint. The optimal attachment posture is solved through multi-objective optimization, and the robotic arm is driven to complete the grounding wire attachment operation.

Images