An ultra-high voltage transmission line unmanned aerial vehicle image recognition induction power taking method

By constructing a target detection model and solving spatiotemporal constraints, a coupled pose vector is generated, which solves the problem of identifying the coupled pose relationship between the conductor and the power collection coil in the inspection of ultra-high voltage transmission lines by UAVs, and improves the alignment reliability and attitude stability of inductive power collection.

CN123067532APending Publication Date: 2026-11-10SHANGHAI HUAGAN DIGITAL TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611397647.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-09
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

In the current UAV inspection of ultra-high voltage transmission lines, the image recognition results are difficult to stably represent the coupling posture relationship between the transmission conductor and the power collection coil, resulting in insufficient reliability and attitude maintenance of inductive power collection alignment control.

Method used

By constructing a target detection model for the conductor and the power collection point, continuous image frames are acquired, and the projection contours of the transmission conductor, the power collection coil, and the candidate regions of the power collection point are jointly identified. The spatial pose deviation between the conductor and the power collection coil is calculated in real time, and a coupled pose vector is generated. Combined with the inter-frame registration relationship and spatiotemporal constraint calculation, a coupled pose vector for inductive power collection alignment control is generated.

Benefits of technology

It improves the alignment reliability and attitude maintenance stability of UAV inductive power supply control, reduces alignment errors caused by environmental interference, ensures that the generation of power supply point candidate areas does not depend on isolated image frames, and provides smoother and more complete visual recognition input data.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123067532A_ABST
    Figure CN123067532A_ABST
Patent Text Reader

Abstract

The present application belongs to the technical field of image recognition and unmanned aerial vehicle vision control, and particularly relates to a kind of superhigh voltage transmission line unmanned aerial vehicle image recognition induction power taking method.Method collects the continuous image of unmanned aerial vehicle close to high-voltage conductor, and simultaneously identifies power transmission conductor, visible arc segment of power taking coil and power taking point by target detection model;According to the fitting of conductor segmentation boundary, the local tangent direction of conductor center axis is fitted, and the center of power taking coil is calculated according to the visible arc segment of power taking coil combined with the prior parameter of coil physical size, and the projection normal is calculated, and the spatial pose of power transmission conductor and the deviation of power taking coil are calculated combined with space-time constraint, to generate a coupled pose vector containing lateral deviation, longitudinal deviation, angle deviation and window stability.The method reduces the influence of single frame false detection and image drift on power taking alignment, and improves the attitude stability and power taking reliability in the process of unmanned aerial vehicle non-contact online energy supplement.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of image recognition and UAV visual control technology, specifically relating to a method for UAV image recognition induction power extraction for ultra-high voltage transmission lines. Background Technology

[0002] UAV inspections of ultra-high voltage transmission lines typically rely on onboard cameras to acquire images of conductors, insulators, fittings, and the surrounding environment. Image recognition models are then used to detect, segment, or identify defects in the target area. For conductor targets, conventional solutions often employ edge extraction, bounding box detection, or semantic segmentation to determine the approximate location of the conductor in the image, followed by UAV flight control system-based line-following or hovering photography. When close proximity to the conductor is required for specific tasks, existing solutions typically use the center of the conductor's bounding box, the center of the segmented region, or manually set image reference points as the approach basis, converting pixel deviations into flight attitude adjustments. While these solutions can meet the requirements for inspection photography and defect identification, the output primarily focuses on target presence, location, and image region, lacking attitude representation related to the inductive power coupling relationship.

[0003] In existing online power replenishment or inductive power harvesting control for drones, visual positioning and power harvesting control are processed separately. The vision component is responsible for identifying the location of the power transmission line, while the power harvesting control component determines whether to initiate the power harvesting process based on the relative proximity between the drone and the line or the circuit output status. In the image recognition stage, the line is usually treated as a linear target, and the power harvesting coil is often considered as a fixed reference object for the airborne mechanism. The spatial relationship between the two is often expressed through simple center deviation, distance estimation, or preset attitude conditions. In ultra-high voltage line environments, line images are easily affected by backlighting, background edges, tower obstruction, line sag, and flight vibrations. The line boundary and target center obtained in a single frame image are prone to abrupt changes. If subsequent control still relies on the coordinate error of a single image, it is difficult to reflect the coupling state between the local tangential of the line, the projected normal of the coil, and the power harvesting point.

[0004] Image recognition results cannot directly and stably characterize the coupling pose relationship between the transmission line and the power-collecting coil to meet the requirements of inductive power collection. Specifically, ordinary target detection or segmentation results can only provide the image region of the transmission line, and cannot simultaneously provide the constraint relationship between the central axis of the transmission line, the local tangential direction, the coil center, the coil projection normal, and the candidate region of the power-collecting point; single-frame recognition results are also prone to boundary drift or false detection due to environmental interference. As a result, when the UAV approaches the transmission line, it can only perform alignment based on coarse pixel deviations, and cannot determine whether the current coil attitude is within the coupling window, nor can it stably convert the visual recognition results into the coupling pose vector required for inductive power collection alignment control, thus affecting the alignment reliability and attitude maintenance stability in the non-contact power collection process. Summary of the Invention

[0005] The purpose of this invention is to provide a method for image recognition and inductive power extraction from UAVs for ultra-high voltage transmission lines, which can solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for image recognition and inductive power extraction from an UAV on an ultra-high voltage transmission line includes: acquiring continuous image frames during the process of the UAV approaching the ultra-high voltage transmission line;

[0008] A target detection model for power transmission lines and power extraction points is constructed, and the projection contours of power extraction coils and candidate regions of power extraction points in the continuous image frames are jointly identified.

[0009] The conductor's central axis and local tangential direction are fitted based on the segmentation boundary of the transmission conductor, and the coil center and projection normal are determined based on the projected profile of the power collection coil.

[0010] The central axis of the conductor, the local tangential direction, the center of the coil, the projection normal, and the candidate area of ​​the power extraction point are input into the spatiotemporal constraint solution process to calculate the alignment deviation between the spatial pose of the transmission conductor and the power extraction coil in real time, and generate a coupled pose vector for inductive power extraction alignment control.

[0011] Preferably, before the continuous image frames are input into the target detection model, an inter-frame registration relationship is established according to the change in the UAV's heading and the change in the scale of the traverse image;

[0012] In the inter-frame registration relationship, continuous regions of conductor texture, edge regions of coil projection, and weak texture regions of background are selected to form image recognition constraint samples respectively;

[0013] Based on the image recognition constraint samples, each image frame is processed for brightness equalization, edge preservation enhancement, and scale normalization, and the shooting time sequence identifier corresponding to each image frame is retained, so that the subsequent conductor center axis fitting results and power coil projection recognition results can be associated according to the same time sequence coordinates.

[0014] Preferably, the target detection model includes a shared feature extraction network, a fine-grained conductor segmentation branch, a conductor central axis regression branch, a power extraction coil projection recognition branch, and a power extraction point candidate region generation branch.

[0015] The shared feature extraction network extracts multi-scale features including the slender edges of the conductor, the closed contour of the coil, and the local orientation of the conductor.

[0016] The fine-grained segmentation branch of the conductor outputs the boundary mask of the transmission conductor, the conductor center axis regression branch outputs the center axis sampling point sequence based on the boundary mask of the transmission conductor, the power extraction coil projection identification branch outputs the coil projection parameters, and the power extraction point candidate region generation branch generates candidate power extraction windows based on the center axis sampling point sequence and the coil projection parameters.

[0017] Preferably, the spatiotemporal constraint solution process includes: establishing a local coordinate system for the conductor based on the conductor's central axis within a single frame image, and projecting the coil center onto the local coordinate system to obtain the lateral deviation and longitudinal deviation;

[0018] The angular deviation is obtained based on the angle between the local tangential direction and the projection normal.

[0019] The window stability is obtained based on the positional continuity of the candidate power-accessible locations in adjacent image frames;

[0020] The lateral deviation, longitudinal deviation, angular deviation, and window stability are collectively encoded into a coupled pose vector.

[0021] Preferably, the inter-frame registration relationship establishes the main direction constraint of the conductor according to the change of the center line direction of the transmission conductor in adjacent image frames, and establishes the coil motion constraint according to the center displacement of the projected contour of the power taking coil in adjacent image frames;

[0022] When the relative change between the main direction constraint of the conductor and the motion constraint of the coil exceeds the preset consistency range, the corresponding image frame is marked as a frame to be corrected.

[0023] The conductor center axis sampling point sequence and coil projection parameters in the preceding stable image frame are used to perform pose compensation for the frame to be corrected, and the compensated image features are input into the target detection model.

[0024] Preferably, the conductor center axis regression branch performs skeletonized sampling on the boundary mask of the transmission conductor to form a sequence of center axis sampling points arranged along the extension direction of the conductor;

[0025] The central axis sampling point sequence is segmented by local curvature, and the sampling point segments with continuous curvature and boundary confidence that meet the identification conditions are determined as valid conductor segments.

[0026] The candidate power-access point generation branch generates candidate power-access windows based solely on the effective conductor segment, coil projection parameters, and conductor local coordinate system, and configures a window confidence level associated with conductor straightness and coil overlap for each candidate power-access window.

[0027] Preferably, after identifying the visible local arc segment of the coil, the power-taking coil projection recognition branch combines the prior parameters of the known physical dimensions of the coil to fit the local arc segment into a constrained projection ellipse posture, and calculates the center of the projection ellipse, the major axis direction, the minor axis direction, and the arc segment confidence.

[0028] When the relative position of the projected ellipse and the central axis of the conductor in the image coordinates satisfies the preset coupling geometric relationship, the center of the calculated projected ellipse is taken as the coil center, the major axis direction and the minor axis direction are converted into the projection normal, and the projection normal is established with the local tangential direction at an angle.

[0029] Preferably, the window stability is jointly determined by the center drift of the candidate power-taking window in consecutive image frames, the window shape change, the confidence change of the conductor boundary, and the confidence change of the coil arc segment;

[0030] When generating the coupled pose vector, the lateral deviation, longitudinal deviation and angular deviation are respectively assigned weights associated with window stability, and the deviation in the low-stability image frame is constrained and updated according to the deviation change trend of the adjacent stable image frame.

[0031] Preferably, after the pose compensation of the frame to be corrected is completed, a historical feature queue is established, which includes effective conductor segments, central axis sampling point sequence, coil projection parameters and window confidence.

[0032] When a new image frame enters the target detection model, the features in the historical feature queue that match the main direction of the conductor in the new image frame are used as prior constraints to filter out false conductor responses and tower edge responses in the new image frame, to fill in the conductor boundary interruption area caused by coil occlusion, and to update the continuity identifier of the candidate power extraction window.

[0033] Preferably, the coupled pose vector is hierarchically encoded according to continuity identifier and window stability before output;

[0034] When the same candidate power extraction window maintains the same main conductor direction in consecutive image frames and the confidence of the coil arc segment meets the recognition conditions, the corresponding lateral deviation, longitudinal deviation, angular deviation, window stability and window confidence are combined into an effective coupled pose vector.

[0035] When the continuity of the candidate power-taking window is interrupted, the previously mentioned effective coupled pose vector is retained, and the deviation component of the interrupted image frame is temporally interpolated and encoded.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] 1. By employing a combined technical solution that integrates continuous image frames, a target detection model for conductors and power collection points, conductor central axis fitting, power collection coil arc segment attitude calculation, and spatiotemporal constraint calculation, the image recognition output is transformed from a simple target position into a coupled pose vector usable for power collection alignment control. The transmission conductor segmentation boundary is used to determine the central axis, which further defines the local tangential direction. The visible arc segment of the power collection coil, combined with prior constraints on the coil's physical dimensions, is used to calculate the coil center and projection normal. Candidate regions for power collection points are used to define the local coupling window of the conductor. The temporal relationship in continuous image frames is used to constrain single-frame jumps. These technical features work together to correlate lateral deviation, longitudinal deviation, angular deviation, and window stability in the same calculation process, reducing alignment errors caused by relying solely on the target box center or single-frame segmentation results. This allows the UAV's inductive power collection control to obtain visual state input corresponding to the conductor-coil coupling relationship.

[0038] 2. In further processing, the update method of the recognition results is jointly defined by inter-frame registration relationship, conductor principal direction constraint, coil motion constraint, effective conductor segment screening, projection ellipse attitude calculation combined with prior constraints, window confidence, historical feature queue, and hierarchical coding. When the conductor boundary is disturbed by background lines, false responses can be screened out by historical feature queue and conductor principal direction matching; when the power-taking coil partially obstructs the conductor, the pose compensation of the obstructed conductor boundary and central axis can be performed using coil projection parameters and previous stable image frames; when the candidate power-taking window is briefly interrupted, the state continuity of the previous effective coupled pose vector can be preserved by continuity identification and temporal interpolation. As a result, the visual recognition results maintain good temporal consistency under flight jitter, scale changes, and partial occlusion conditions. The generation of candidate power-taking point regions no longer depends on isolated image frames, and subsequent inductive power-taking alignment control can obtain smoother and more complete input data. Attached Figure Description

[0039] Figure 1 A flowchart of the image recognition and inductive power extraction method for UAVs in ultra-high voltage transmission lines;

[0040] Figure 2 Flowchart of preprocessing and inter-frame registration for consecutive image frames;

[0041] Figure 3 A flowchart for branch recognition in the target detection model of conductors and power take-off points;

[0042] Figure 4 The flowchart shows the process of calculating the stability of candidate power extraction windows, updating the historical feature queue, and hierarchical encoding of coupled pose vectors. Detailed Implementation

[0043] In one embodiment, please refer to Figure 1A method for image recognition and inductive power extraction from UAVs approaching UAV transmission lines is proposed. This method is used for visual state recognition and alignment data generation during UAV approach to UAV transmission lines. Specifically, it includes acquiring continuous image frames during the UAV's approach to the UAV transmission line; constructing a target detection model of the transmission line and power extraction point; jointly identifying the transmission line, the projected contour of the power extraction coil, and candidate regions of potential power extraction points in the continuous image frames; fitting the transmission line's central axis and local tangential direction based on the transmission line's segmentation boundaries; determining the coil center and projection normal based on the power extraction coil's projected contour; and inputting the transmission line's central axis, local tangential direction, coil center, projection normal, and candidate regions of potential power extraction points into a spatiotemporal constraint solution process. The method then calculates the alignment deviation between the spatial pose of the transmission line and the power extraction coil in real time and generates... The coupled pose vector used for inductive power-taking alignment control uses continuous image frames as a unified data source, so that the wire identification results, coil identification results, and power-taking point identification results are associated and processed in the same image coordinate link. This avoids separating wire detection and power-taking mechanism alignment into two unconstrained identification tasks. The output of the target detection model does not use a single target box as the final data, but continues to form coupled geometric quantities through the wire central axis, local tangential direction, coil center, and projection normal. This forms a structured visual state that can be read by subsequent visual servo closed-loop control. The coupled pose vector includes both position deviation and attitude deviation and power-taking window stability state, enabling the UAV to determine whether the power-taking coil is located in the local coupling area of ​​the wire before non-contact inductive power taking.

[0044] In this embodiment, consecutive image frames enter the image buffer queue according to the shooting sequence. Before entering the target detection model, each image frame is assigned a time sequence identifier and an image coordinate identifier. The time sequence identifier is used to indicate the sequential relationship of the image frames during the UAV's approach to the conductor. The image coordinate identifier is used to maintain the data consistency of the conductor, the power-taking coil, and the candidate power-taking window under the same coordinate reference. Specifically, the image buffer queue performs coarse localization of the conductor region in adjacent image frames, uses the continuity of the slender edges of the conductor region to form an initial search band, and performs edge preservation enhancement within the initial search band to prevent the weak texture area of ​​the conductor from being overly smoothed due to changes in background brightness. At the same time, the closed edge information of the region where the projection contour of the power-taking coil is located is preserved so that the coil center and projection normal can be stably extracted in the subsequent recognition process. Consecutive image frames are not simply treated as independent images and processed frame by frame. Instead, the changes in conductor direction, coil projection displacement, and candidate window continuity between adjacent frames are preserved in the image buffer queue to provide traceable data input for subsequent spatiotemporal constraint solutions.

[0045] In this embodiment, the target detection model adopts a computational structure that combines shared image features with branch outputs. The input is a continuous image frame after time-series identification and scale normalization. The output includes the transmission line segmentation boundary, the sequence of sampling points along the conductor's central axis, the local tangential direction of the conductor, the projected contour of the power extraction coil, the coil center, the coil projection normal, and candidate regions for power extraction points. Shared image features are used to preserve the slender edges of the conductor, weak background texture, the closed contour of the coil, and the local orientation of the conductor in the same feature map. The conductor segmentation boundary is obtained from the fine-grained segmentation results of the conductor. The sequence of sampling points along the central axis is obtained by skeletonizing the inner side of the conductor boundary. The local tangential direction is obtained by fitting the local direction between adjacent central axis sampling points. The coil center is obtained from the geometric center of the projected contour of the power extraction coil. The projection normal is obtained by converting the major axis direction and the minor axis direction of the projected contour. The candidate regions for power extraction points are jointly determined by the local straightness of the conductor, the coil projection position, and the continuity of the conductor boundary. The outputs of the target detection model maintain a common image coordinate reference relationship, enabling conductor target recognition and power extraction point recognition to participate in the alignment deviation calculation under the same constraint system.

[0046] In this embodiment, the fitting of the conductor's central axis does not use a single edge line as the conductor's position. Instead, it is based on the conductor region formed by the conductor segmentation boundary. Multiple central axis sampling points are extracted along the conductor's extension direction. The central axis sampling points form a central axis sampling point sequence according to the arrangement order in the image coordinates. After local smoothing, the conductor's central axis is obtained. The local tangential direction is determined by the central axis direction in the neighborhood of the current central axis sampling point. When a local area is affected by background lines or conductor boundary breaks, the local tangential direction is corrected by constraining the direction of the adjacent stable sampling point segment. The local arc segment edge of the power-taking coil is obtained by fitting edge detection and prior shape constraints. The coil center is determined by combining the known physical size parameters of the coil with the geometric distribution of the local arc segment. The projection normal is determined by transforming the relationship between the tangent direction of the local arc segment and the radial constraint of the coil. Thus, the conductor's posture and the coil's posture are converted into calculable geometric quantities in the same image plane.

[0047] In this embodiment, the coupled pose vector is generated according to the conductor local coordinate system. The conductor local coordinate system takes the sampling point of the conductor's central axis within the candidate power extraction window as the origin, the local tangential direction as the longitudinal reference, and the direction perpendicular to the local tangential direction as the lateral reference. The projection of the coil center in the conductor local coordinate system is used to form the lateral and longitudinal deviations. The angle between the projection normal and the local tangential direction is used to form the angular deviation. The continuity of the candidate power extraction window in consecutive image frames is used to form the window stability. The coupled pose vector is calculated according to the following formula.

[0048] ;

[0049] in, This represents the coupled pose vector of the image frame corresponding to time t. This indicates the deviation of the coil center from the lateral reference of the local coordinate system of the conductor. This indicates the deviation of the coil center from the longitudinal reference of the local coordinate system of the conductor. This represents the angular deviation between the projection normal and the local tangential direction. This indicates the stability of the candidate power extraction window. During the calculation, if the center of the coil falls to the left of the central axis of the conductor, then... Take the negative value; if it falls to the right of the conductor's central axis, then... Take a positive value; if the center of the coil is located on the forward extension side of the candidate power extraction window, then... Take a positive value, if it is located on the reverse extension side. If a negative value is taken, the relationship between the projection normal and the local tangential direction being nearly orthogonal or parallel is determined by a pre-established coupling geometric relationship. The value of is determined by the center drift and morphological changes of the candidate power-taking window in adjacent image frames, and each component of the coupled pose vector comes from the image recognition output and the spatiotemporal constraint solution results.

[0050] In a preferred embodiment, please refer to Figure 2 Before inputting consecutive image frames into the target detection model, an inter-frame registration relationship is established based on the UAV's heading change and the conductor image scale change. Within this inter-frame registration relationship, continuous texture regions of the conductor, edge regions of the coil projection, and weakly textured background regions are selected to form image recognition constraint samples. Based on these constraint samples, each image frame undergoes brightness equalization, edge preservation enhancement, and scale normalization. The capture time sequence identifier for each image frame is also retained, ensuring that subsequent conductor center axis fitting results and coil projection recognition results can be correlated according to the same temporal coordinates. Specifically, the UAV's heading change is used to determine the overall directional change of the conductor in adjacent image frames. The image scale change of the conductor is used to determine the change in the conductor pixel width caused by the change in the relative distance between the UAV and the conductor. The continuous texture region of the conductor is used to preserve the edge direction of the conductor. The edge region of the coil projection is used to preserve the closed outline of the power-taking coil. The weak texture region of the background is used to identify low-contrast interference in the sky or distant background. Brightness equalization only applies to the local brightness distribution of the image without changing the relative geometric position of the conductor and the coil. Edge preservation enhancement is limited to the conductor search zone and the neighborhood of the coil projection. The scale normalization process is constrained by the scale relationship between the conductor pixel width and the outer region of the coil projection to avoid the inability to directly compare geometric quantities in different image frames due to changes in image scale.

[0051] Furthermore, the inter-frame registration relationship can be determined jointly by the main direction of the conductor and the direction of coil motion. The main direction of the conductor is obtained from the overall fitting direction of the sampling point sequence of the conductor's central axis, and the direction of coil motion is obtained from the displacement of the coil center in adjacent image frames. When the UAV approaches the conductor, the scale of the conductor image usually changes continuously with the approach process, and the direction of movement of the coil projection center should also remain continuous with the approach trend in the visual servoing process. If the main direction of the conductor changes abruptly in a certain image frame but the attitude of the coil arc segment does not change accordingly, then the image frame may have a situation where the background lines are misidentified as conductors. If the conductor boundary is locally broken but the attitude of the coil arc segment remains continuous, then the image frame may have a situation where the coil partially occludes the conductor or the conductor boundary is disturbed and not fully extracted. Therefore, in this embodiment, the main direction constraint of the conductor and the coil motion constraint are combined into a consistency check quantity, and the consistency check quantity is used to mark the frame to be corrected and the stable image frame, so that the image data entering the target detection model has a clear temporal reference relationship.

[0052] In a preferred embodiment, please refer to Figure 3 The target detection model includes a shared feature extraction network, a fine-grained conductor segmentation branch, a conductor central axis regression branch, a power extraction coil projection recognition branch, and a power extraction point candidate region generation branch. The shared feature extraction network extracts multi-scale features including the slender edges of the conductor, the closed contour of the coil, and the local orientation of the conductor. The fine-grained conductor segmentation branch outputs a transmission conductor boundary mask. The conductor central axis regression branch outputs a central axis sampling point sequence based on the transmission conductor boundary mask. The power extraction coil projection recognition branch outputs coil projection parameters. The power extraction point candidate region generation branch generates candidate power extraction windows based on the central axis sampling point sequence and the coil projection parameters. Specifically, the shared feature extraction network preserves the gray-scale transitions of the wire edges and the coil contour edges in shallow features, and preserves the semantic differences between the local direction of the wire and the background in deep features. The fine-grained segmentation branch of the wire defines the pixel region of the wire through a boundary mask. The wire central axis regression branch generates central axis sampling points inside the boundary mask instead of directly regressing the center line from the original image. The power-taking coil projection recognition branch obtains projection parameters that can express the coil posture based on visible local arc segments combined with the prior size constraints of the coil. The candidate power-taking window generation branch jointly calculates the straightness of the wire sampling points, the boundary confidence, and the coil projection position.

[0053] In this embodiment, the training data for the target detection model can consist of inspection images of ultra-high voltage transmission lines, images of UAVs approaching conductors, and projection images of power extraction coils. The training annotations include conductor boundary masks, conductor central axis sampling points, visible arc segment annotations of coils, coil projection attitude parameters, and candidate power extraction window regions. The conductor boundary mask is used to supervise the fine-grained segmentation branch of the conductor, the conductor central axis sampling points are used to supervise the conductor central axis regression branch, the coil projection contour and ellipse parameters are used to supervise the power extraction coil projection recognition branch, and the candidate power extraction window regions are used to supervise the generation branch of candidate regions for power extraction points. During training, the target detection model applies coordinate consistency constraints to multiple outputs in the same image frame, ensuring that the conductor central axis sampling points do not deviate from the conductor region formed by the conductor boundary mask, the coil projection parameters do not deviate from the geometric constraints of the visible arc segments, and the candidate power extraction windows do not fall in areas with low confidence of the conductor boundary or sudden changes in the local curvature of the conductor. This ensures that the model output maintains the geometric correlation required for coupling the power extraction scenario from the training stage.

[0054] In this embodiment, the correlation between the conductor center axis and the coil projection parameters is constrained by geometric loss. Geometric loss is used to limit the data deviation between the conductor center axis, the coil center, and the projection normal, and can be specifically expressed by the following formula.

[0055] ;

[0056] in, Represents geometric loss, Indicates the center of the coil. Indicates the candidate power supply window. This represents the normalized distance from the coil center to the candidate power extraction window. Indicates the normal direction of the coil projection. Indicates the local tangential direction of the conductor. This represents the window confidence level of the candidate power-on window. and These represent the weights of the distance, angle, and window terms, respectively. The absolute value of the dot product of two direction vectors is represented. When the coil center is close to the candidate power extraction window and the projection normal and the local tangential direction of the conductor satisfy the preset coupling geometric relationship, the distance and angle terms are in a smaller value range. When the candidate power extraction window is generated by the conductor region with high boundary confidence, the window term is in a smaller value range. The training process uses this to constrain different output branches to generate data that can jointly participate in the coupled pose calculation.

[0057] In a preferred embodiment, the spatiotemporal constraint solution process includes establishing a local coordinate system for the conductor based on the conductor's central axis within a single frame image, projecting the coil center onto the local coordinate system to obtain lateral and longitudinal deviations, obtaining angular deviations based on the angle between the local tangential direction and the projection normal, obtaining window stability based on the positional continuity of the candidate power-taking point region in adjacent image frames, and encoding the lateral deviation, longitudinal deviation, angular deviation, and window stability together into a coupled pose vector. Specifically, the conductor's local coordinate system is not a fixed image coordinate system, but is updated according to the direction of the conductor's central axis within the candidate power-taking window. When the conductor extends sag or obliquely in the image, the lateral deviation is still based on the conductor's local normal, the longitudinal deviation is still based on the conductor's extension direction, the angular deviation is calculated from the projection normal and the local tangential direction, and the window stability is obtained by comprehensively considering the position, shape, and confidence continuity of the candidate power-taking window in consecutive image frames, so that the coupled pose vector does not lose its physical correspondence due to the global tilt of the conductor in the image.

[0058] Furthermore, the lateral and longitudinal deviations can be calculated by projecting the coil center relative to the local coordinate system of the conductor, specifically using the following formula.

[0059]

[0060] ;

[0061] in, This represents the coordinates of the coil center at time t. This indicates the coordinates of a local reference point on the conductor within the candidate power extraction window. This represents the local tangential unit vector obtained from the central axis of the conductor. Indicates and The vertical local normal unit vector, when When mainly distributed along the local normal, The absolute value is relatively large and When the difference vector is close to zero, and is mainly distributed along the direction of the conductor's extension, The absolute value is relatively large and Approaching zero, subsequent alignment control can distinguish between the lateral deviation of the coil from the conductor and the longitudinal deviation of sliding along the candidate window based on the independent components in two directions.

[0062] In a preferred embodiment, the inter-frame registration relationship establishes the main direction constraint of the conductor according to the change of the centerline direction of the transmission conductor in adjacent image frames, and establishes the coil motion constraint according to the center displacement of the projected contour of the power take-up coil in adjacent image frames. When the relative change between the main direction constraint of the conductor and the coil motion constraint exceeds the preset consistency range, the corresponding image frame is marked as the frame to be corrected. The frame to be corrected is used to perform pose compensation by referencing the conductor center axis sampling point sequence and coil projection parameters in the preceding stable image frame, and the compensated image features are input into the target detection model. Specifically, the main direction constraint of the conductor is represented by the angle of the main direction of the center axis sampling point sequence in adjacent image frames, and the coil motion constraint is represented by the displacement direction of the coil center in adjacent image frames. When the main direction of the conductor is continuous in the preceding stable image frame and a short-term change in direction occurs in the frame to be corrected, the system uses the center axis sampling point sequence in the stable image frame as a reference to find the conductor response with the same direction in the frame to be corrected. When the confidence of the visible arc segment of the coil in the frame to be corrected is insufficient due to illumination interference, the system predicts the pose range of the coil arc segment in the preceding stable image frame and rematches the local edge features within the prediction range.

[0063] In this embodiment, the pose compensation of the frame to be corrected does not directly copy the recognition result of the previous stable image frame. Instead, it maps the conductor central axis sampling point sequence and coil projection parameters of the previous stable image frame as prior constraints to the frame to be corrected. The mapping process takes the common change of the conductor principal direction and the coil center displacement as a reference. If there is a local missing part of the conductor boundary in the frame to be corrected, the central axis sampling point segment in the previous stable image frame that is consistent with the current conductor principal direction is retained, and continuous curve compensation is used on both sides of the missing area to generate the central axis candidate segment. If the conductor boundary in the frame to be corrected is partially occluded by the coil, the boundary features of the visible part of the conductor are retained, and the central axis candidate segment of the occluded part is completed according to the previous conductor central axis parameters. The visible arc segment of the coil is not affected by the conductor occlusion. It only needs to be combined with the previous pose parameters for consistency verification. The compensated image features are not the final output, but are used as constrained features to be input into the target detection model again, so that the target detection model avoids deviating from the original temporal trajectory of the conductor and coil when re-identifying.

[0064] In a preferred embodiment, the conductor central axis regression branch performs skeletonized sampling on the boundary mask of the transmission conductor to form a sequence of central axis sampling points arranged along the extension direction of the conductor. The central axis sampling point sequence is segmented by local curvature. Sampling point segments with continuous curvature and boundary confidence that meet the recognition conditions are determined as valid conductor segments. The power-taking point candidate region generation branch generates candidate power-taking windows based only on the valid conductor segments, coil projection parameters, and the local coordinate system of the conductor. For each candidate power-taking window, a window confidence associated with the straightness of the conductor and the overlap of the coil is configured. Specifically, skeletonized sampling finds points with relatively balanced pixel distances to the two sides of the boundary along the extension direction of the conductor within the boundary mask. After forming the central axis sampling point sequence, the sampling point sequence is divided into multiple candidate segments according to the local curvature changes. Sampling point segments with small local curvature changes and continuous confidence on both sides of the boundary are determined as valid conductor segments. Sampling point segments with abrupt changes in local curvature, boundary breaks, or dense background responses do not participate in the generation of candidate power-taking windows. The coil projection parameters are used to determine whether the coil center can form a calculable coupling relationship with the valid conductor segment.

[0065] In this embodiment, the local curvature segmentation can be calculated by the change in adjacent directions of the central axis sampling point sequence, specifically using the following formula.

[0066] ;

[0067] in, Indicates the first Local curvature measurement of each central axis sampling point , and These represent the coordinates of adjacent central axis sampling points, Describing the vector norm, This represents a stable term used to avoid a denominator of zero. If adjacent sampling points are approximately aligned along the same straight line, the numerator will have a smaller value. If the sampling point is within a lower range and exhibits a broken line change or local false detection due to background interference, the numerator value will increase. At a higher level, local curvature metrics and conductor boundary confidence are used together to determine valid conductor segments and restrict candidate power extraction windows to be generated only in regions where the conductor central axis is continuous and the boundary identification is stable.

[0068] In a preferred embodiment, after identifying the visible local arc segment of the coil, the power-taking coil projection recognition branch combines the known physical radius and prior parameters of the coil's loop width to fit the local arc segment into a constrained projection ellipse posture, and calculates the center of the projection ellipse, the major axis direction, the minor axis direction, and the arc segment confidence. When the relative position of the projection ellipse and the conductor's central axis in the image coordinates satisfies a preset coupling geometric relationship, the calculated center of the projection ellipse is taken as the coil center, the major axis direction and the minor axis direction are converted into the projection normal, and an angular correlation is established between the projection normal and the local tangential direction. Specifically, the visible arc segment of the coil is composed of local arc-shaped edge points in the image corresponding to the edge of the coil projection. Isolated edge points are removed during attitude fitting, and the fixed size parameters of the coil are used as strong constraints to avoid overfitting. The center of the projection ellipse is calculated by combining the geometric distribution of the arc segment with the prior radius. The major axis direction corresponds to the radial extension direction of the coil projection, and the minor axis direction corresponds to the axial extension direction of the coil projection. The confidence of the arc segment is determined by the continuity of the visible edge points, the gradient strength, and the fitting residual. When the confidence of the arc segment does not meet the recognition conditions, the coil projection parameters do not directly participate in the generation of the coupled pose vector, but instead enter the compensation process of the frame to be corrected.

[0069] In this embodiment, the transformation of the projection normal is jointly defined by the major axis and minor axis of the projection ellipse, avoiding the inference of the coil attitude solely from the center position of the ellipse. Specifically, the major axis vector and the minor axis vector are normalized, and the direction of projection attitude change is determined based on the axial length change of the coil projection ellipse in consecutive image frames. When the major axis rotates continuously relative to the local tangential direction of the conductor while the minor axis changes in a consistent manner, the projection normal is updated according to the continuous change of the ellipse axis. When the ellipse contour is partially missing, causing instability in the minor axis direction, the projection normal is jointly constrained by the major axis direction, the previous projection normal, and the currently visible contour. After establishing an angular relationship between the projection normal and the local tangential direction, the angular deviation is no longer indirectly inferred from the positional relationship between the coil center and the conductor's central axis, but is directly calculated using the attitude parameters of the coil projection contour.

[0070] In a preferred embodiment, please refer to Figure 4The stability of the window is jointly determined by the center drift of the candidate power-taking window in consecutive image frames, the change in window shape, the change in confidence of the conductor boundary, and the change in confidence of the coil arc segment. When generating the coupled pose vector, weights associated with the window stability are assigned to the lateral deviation, longitudinal deviation, and angular deviation, respectively. The deviation in the low-stability image frame is constrained and updated according to the deviation change trend of the adjacent stable image frame. Specifically, the center drift of the candidate power-taking window is used to describe whether the candidate window has a sudden position jump in the adjacent image frame. The change in window shape is used to describe whether the length and width of the candidate window remain continuous with the change of conductor scale. The change in confidence of the conductor boundary is used to describe whether the conductor segmentation boundary has a short-term break or is blocked by the coil. The change in confidence of the coil arc segment is used to describe the recognition stability of the visible edge of the power-taking coil. The window stability participates in the encoding process of the coupled pose vector, so that the deviation component has a high reference weight when the image recognition is stable, and is subject to the temporal constraints of the preceding and following stable image frames when the image recognition is unstable for a short time.

[0071] In this embodiment, the window stability can be calculated using multi-factor normalization of candidate windows, specifically using the following formula.

[0072] ;

[0073] in, This represents the normalized drift of the candidate power-taking window center relative to the previous stable image frame. This represents the change in the shape of the candidate power-taking window. This represents the change in confidence level at the conductor boundary. This indicates the change in confidence level across the coil arc segment. and They represent and The corresponding weights for each change This indicates that, in exponential operations, when the candidate power extraction windows are continuous in position and shape, the conductor boundaries are stable, and the coil arc segment identification is stable, the various changes are in a low range. When the value approaches a higher level, and there is a jump in the candidate window, a break in the conductor boundary, or insufficient confidence in the coil segment, at least one change will increase. The system reduces the direct reference strength of the current image frame deviation component and initiates constraint updates accordingly.

[0074] In a preferred embodiment, after pose compensation is completed for the frame to be corrected, a historical feature queue is established, which includes effective conductor segments, central axis sampling point sequences, coil projection parameters, and window confidence. When a new image frame enters the target detection model, features in the historical feature queue that match the main direction of the conductor in the new image frame are used as prior constraints to filter out false conductor responses and tower edge responses in the new image frame, and to complete the conductor boundary interruption area caused by coil occlusion. At the same time, the continuity identifier of the candidate power extraction window is updated. Specifically, the historical feature queue stores the conductor and coil features of the most recent stable image frames according to the shooting time sequence. The stored content includes the range of central axis sampling points of effective conductor segments, the local tangential direction of the conductor, the coil projection ellipse parameters, the position of the candidate power extraction window, and the window confidence. After a new image frame enters the recognition process, the system first matches similar conductor segments in the historical feature queue according to the main direction of the conductor, and then excludes linear responses in the new image frame that are inconsistent in direction, inconsistent in boundary width, or cannot establish a continuous relationship with the historical candidate window, so that the tower edge, background lines, and coil occlusion edges do not participate in the generation of the conductor central axis.

[0075] Furthermore, the continuity identifier is used to indicate the continuity relationship of the same candidate power extraction window in consecutive image frames. The continuity identifier is jointly determined by the conductor main direction matching result, the candidate window center drift result, the window shape matching result, and the coil projection relative position matching result. When the candidate power extraction window in the new image frame matches the window in the historical feature queue in the conductor main direction, and the candidate window center drift and shape change are within the preset consistency range, the continuity identifier follows the historical window identifier. When the candidate window in the new image frame cannot establish a matching relationship with the historical window, the continuity identifier is marked as interrupted. If the interruption only occurs in a few adjacent image frames and the conductor main direction of the preceding and following stable image frames is consistent, the system allows the candidate window state to be restored through timing interpolation. If there is a continuous interruption and it is accompanied by a significant change in the conductor main direction, the system stops referencing the previous candidate window and re-establishes the candidate power extraction window.

[0076] In a preferred embodiment, the coupled pose vector is hierarchically encoded according to continuity markers and window stability before output. When the same candidate power-taking window maintains the same main conductor direction in consecutive image frames and the coil arc confidence meets the recognition conditions, the corresponding lateral deviation, longitudinal deviation, angular deviation, window stability, and window confidence are combined into an effective coupled pose vector. When the continuity marker of the candidate power-taking window is interrupted, the previous effective coupled pose vector is retained, and the deviation components of the interrupted image frame are temporally interpolated and encoded. Specifically, the hierarchical encoding divides the coupled pose vector into a stable output state, a constrained output state, and a maintained output state. The stable output state corresponds to image frames where the continuity marker is not interrupted and the window stability is in a high range. The constrained output state corresponds to image frames where the candidate window can still be tracked, but some deviation components are affected by occlusion or boundary breakage. The maintained output state corresponds to image frames where the candidate window is interrupted for a short time and the previous effective coupled pose vector needs to be referenced. During the output process, the source markers of the lateral deviation, longitudinal deviation, and angular deviation are retained in each state.

[0077] In this embodiment, temporal interpolation coding constrains the generation of deviation components in interrupted image frames. The interpolation calculation does not change the already confirmed coupled pose vectors in the preceding and following stable image frames. Instead, it generates transitional deviation components within the interrupted interval according to the deviation change trend of adjacent stable image frames. Specifically, the following formula can be used.

[0078] ;

[0079] in, This represents the interpolated coupled pose vector of the interrupted image frame at time t. Indicates the stable moment preceding the interruption interval. The effective coupled pose vector. Indicates a stable moment after the interruption interval. The effective coupled pose vector, t represents the position of the vector. and Interrupted image frame time between t, when t is close to When t approaches, the interpolation result references more of the preceding effective coupled pose vector. In this case, the interpolation results reference the subsequent effective coupled pose vector more often. Interpolation is only enabled when the main direction of the conductor is continuous and the interrupted interval meets the constrained output conditions, so that short-term occlusion or short-term false detection will not directly cause the coupled pose vector to jump.

[0080] In this embodiment, the output interface of the coupled pose vector is oriented towards the inductive power-taking alignment control process. The output data includes vector components, hierarchical coding status, continuity identifier, window confidence, and source marker. The vector components are used to express the position and attitude relationship between the conductor and the power-taking coil. The hierarchical coding status is used to indicate whether the current visual state can be directly referenced. The continuity identifier is used to indicate whether the candidate power-taking window continues along the same conductor segment. The window confidence is used to indicate the degree to which the candidate power-taking window is supported by the effective conductor segment and the coil projection parameters. The source marker is used to distinguish whether the deviation component comes from direct solution of the current image frame, constrained update, or temporal interpolation encoding. After reading the output data, the inductive power-taking alignment control process can adopt the corresponding attitude adjustment logic according to the different states of stable output, constrained output, and maintained output. This embodiment does not rely on a single target center point as the control input, but expresses the local geometric relationship of the conductor, the coil projection attitude, and the temporal state of the power-taking window as structured visual data.

[0081] In a preferred embodiment, the target detection model for the conductor and the power extraction point can also be configured with a consistency check process for the recognition results. After the coupled pose vector is generated, the consistency check process checks in reverse whether there are coordinate conflicts between the conductor segmentation boundary, the central axis sampling point sequence, the coil projection parameters, and the candidate power extraction window. If the lateral deviation shows that the coil center has entered the vicinity of the candidate power extraction window, but the coil projection integrity is low and the conductor boundary confidence decreases, the system marks the current image frame as a constrained output state. If the angle deviation shows that the relationship between the coil projection normal and the local tangential direction is abnormal, but the ellipse integrity in the continuous image frames remains stable, the system re-checks whether the local tangential direction is affected by the conductor missegmentation. If the candidate power extraction window has high confidence but the continuity identification is interrupted, the system prioritizes checking the conductor main direction matching results in the historical feature queue. Through the above check process, the data consistency verification of each branch output is completed before entering the coupled pose vector.

[0082] In this embodiment, the collaborative relationship between the main direction of the conductor, the coil projection parameters, and the candidate power extraction window runs through the entire recognition link. The main direction of the conductor is used to limit the extension direction of the candidate window, the coil projection parameters are used to limit the attitude state of the power extraction coil relative to the conductor, and the candidate power extraction window is used to limit whether a local area of ​​the conductor is suitable for participating in the alignment calculation. The three maintain temporal consistency in consecutive image frames through inter-frame registration relationship and historical feature queue. When the conductor is affected by backlight or weak background texture in the image, the main direction of the conductor can be constrained by the previous and subsequent stable image frames. When the coil projection contour is incomplete due to local occlusion, the coil projection parameters can be compensated by the visible edge of the projection ellipse and the historical axial relationship. When the candidate power extraction window is affected by false detection interference, the window confidence and continuity identifier can exclude it from the subsequent coupled pose vector generation process.

[0083] In this embodiment, the method can also update the recognition constraint samples of the target detection model through sample backfeeding. The sample backfeeding objects include confirmed stable image frames, frames to be calibrated, constrained output state image frames, and maintained output state image frames. Stable image frames are used to supplement positive samples of conductor boundaries, central axis sampling points, and coil arc poses. Frames to be calibrated are used to supplement interference samples of background lines, tower material edges, and coil occlusion areas of conductors. Constrained output state image frames are used to supplement constraint samples under conditions of weak conductor texture and incomplete coil contours. Maintained output state image frames are used to supplement temporal samples under conditions of short-term interruption of candidate windows. During the update, the calculation rules of the coupled pose vector are not directly changed. Instead, the feature distribution of the continuous texture region of the conductor, the projection edge region of the coil, and the weak texture region of the background is updated, so that the subsequent target detection model can generate recognition outputs that are more consistent with the coupled geometric relationship under similar image conditions.

[0084] In this embodiment, the method is applicable to the fully autonomous online power replenishment process of UAVs during long-distance inspection of ultra-high voltage transmission lines. The image recognition part organizes data around the inductive power extraction coupling relationship between the conductor and the power extraction coil. Continuous image frames provide temporal information of the conductor and coil during the UAV's approach. The target detection model provides the joint recognition results of the projection contours of the transmission conductor, the power extraction coil, and the candidate areas of the power extraction point. The conductor central axis fitting and local tangential direction calculation provide the local attitude reference of the conductor. The coil center and projection normal calculation provide the attitude reference of the power extraction coil. The spatiotemporal constraint solution process converts the above data into a coupled pose vector. Inter-frame registration, pose compensation, effective conductor segment screening, projection ellipse fitting, window stability calculation, historical feature queue, and hierarchical coding further limit the source and reliable state of the coupled pose vector, so that the UAV's inductive power extraction alignment control obtains continuous, interpretable visual input data that corresponds to the spatial relationship between the conductor and the coil.

Claims

1. A method for image recognition and inductive power extraction from UAVs for ultra-high voltage transmission lines, characterized in that, include: Acquire continuous image frames during the process of a drone approaching an ultra-high voltage power transmission line; A target detection model for power transmission lines and power extraction points is constructed, and the projection contours of power extraction coils and candidate regions of power extraction points in the continuous image frames are jointly identified. The conductor's central axis and local tangential direction are fitted based on the segmentation boundary of the transmission conductor, and the coil center and projection normal are determined based on the projected profile of the power collection coil. The central axis of the conductor, the local tangential direction, the center of the coil, the projection normal, and the candidate area of ​​the power extraction point are input into the spatiotemporal constraint solution process to calculate the alignment deviation between the spatial pose of the transmission conductor and the power extraction coil in real time, and generate a coupled pose vector for inductive power extraction alignment control.

2. The method for image recognition and inductive power extraction from UAVs for ultra-high voltage transmission lines according to claim 1, characterized in that, Before being input into the target detection model, the continuous image frames establish inter-frame registration relationships based on the UAV heading change and the guide image scale change. In the inter-frame registration relationship, continuous regions of conductor texture, edge regions of coil projection, and weak texture regions of background are selected to form image recognition constraint samples respectively; Based on the image recognition constraint samples, each image frame is processed for brightness equalization, edge preservation enhancement, and scale normalization, and the shooting time sequence identifier corresponding to each image frame is retained, so that the subsequent conductor center axis fitting results and power coil projection recognition results can be associated according to the same time sequence coordinates.

3. The method for image recognition and inductive power extraction from an UAV for ultra-high voltage transmission lines according to claim 2, characterized in that, The target detection model includes a shared feature extraction network, a fine-grained conductor segmentation branch, a conductor central axis regression branch, a power extraction coil projection recognition branch, and a power extraction point candidate region generation branch. The shared feature extraction network extracts multi-scale features including the slender edges of the conductor, the closed contour of the coil, and the local orientation of the conductor. The fine-grained segmentation branch of the conductor outputs the boundary mask of the transmission conductor, the conductor center axis regression branch outputs the center axis sampling point sequence based on the boundary mask of the transmission conductor, the power extraction coil projection identification branch outputs the coil projection parameters, and the power extraction point candidate region generation branch generates candidate power extraction windows based on the center axis sampling point sequence and the coil projection parameters.

4. The method for image recognition and inductive power extraction from an UAV for ultra-high voltage transmission lines according to claim 3, characterized in that, The spatiotemporal constraint solution process includes: establishing a local coordinate system for the conductor based on the conductor's central axis within a single frame image, and projecting the coil center onto the local coordinate system to obtain the lateral and longitudinal deviations; The angular deviation is obtained based on the angle between the local tangential direction and the projection normal. The window stability is obtained based on the positional continuity of the candidate power-accessible locations in adjacent image frames; The lateral deviation, longitudinal deviation, angular deviation, and window stability are collectively encoded into a coupled pose vector.

5. The method for image recognition and inductive power extraction from an UAV for ultra-high voltage transmission lines according to claim 3, characterized in that, The inter-frame registration relationship establishes the main direction constraint of the conductor according to the change of the center line direction of the transmission conductor in adjacent image frames, and establishes the coil motion constraint according to the center displacement of the projected contour of the power taking coil in adjacent image frames. When the relative change between the main direction constraint of the conductor and the motion constraint of the coil exceeds the preset consistency range, the corresponding image frame is marked as a frame to be corrected. The conductor center axis sampling point sequence and coil projection parameters in the preceding stable image frame are used to perform pose compensation for the frame to be corrected, and the compensated image features are input into the target detection model.

6. The method for image recognition and inductive power extraction from an UAV for ultra-high voltage transmission lines according to claim 3, characterized in that, The conductor center axis regression branch performs skeletonized sampling on the boundary mask of the transmission conductor to form a sequence of center axis sampling points arranged along the extension direction of the conductor. The central axis sampling point sequence is segmented by local curvature, and the sampling point segments with continuous curvature and boundary confidence that meet the identification conditions are determined as valid conductor segments. The candidate power-access point generation branch generates candidate power-access windows based solely on the effective conductor segment, coil projection parameters, and conductor local coordinate system, and configures a window confidence level associated with conductor straightness and coil overlap for each candidate power-access window.

7. The method for image recognition and inductive power extraction from an UAV for ultra-high voltage transmission lines according to claim 3, characterized in that, After identifying the visible local arc segment of the coil, the power-taking coil projection recognition branch combines the prior parameters of the coil's known physical dimensions to fit the local arc segment into a constrained projection ellipse posture, and calculates the center of the projection ellipse, the direction of the major axis, the direction of the minor axis, and the confidence level of the arc segment. When the relative position of the projected ellipse and the central axis of the conductor in the image coordinates satisfies the preset coupling geometric relationship, the center of the calculated projected ellipse is taken as the coil center, the major axis direction and the minor axis direction are converted into the projection normal, and the projection normal is established with the local tangential direction at an angle.

8. The method for image recognition and inductive power extraction from an UAV for ultra-high voltage transmission lines according to claim 4, characterized in that, The stability of the window is determined by the center drift of the candidate power-taking window in consecutive image frames, the change in window shape, the change in confidence of the conductor boundary, and the change in confidence of the coil arc segment. When generating the coupled pose vector, the lateral deviation, longitudinal deviation and angular deviation are respectively assigned weights associated with window stability, and the deviation in the low-stability image frame is constrained and updated according to the deviation change trend of the adjacent stable image frame.

9. A method for image recognition and inductive power extraction from an UAV for ultra-high voltage transmission lines according to claim 5, characterized in that, After the pose compensation of the frame to be corrected is completed, a historical feature queue is established, which includes effective conductor segments, central axis sampling point sequence, coil projection parameters and window confidence. When a new image frame enters the target detection model, the features in the historical feature queue that match the main direction of the conductor in the new image frame are used as prior constraints to filter out false conductor responses and tower edge responses in the new image frame, to fill in the conductor boundary interruption area caused by coil occlusion, and to update the continuity identifier of the candidate power extraction window.

10. A method for image recognition and inductive power extraction from an UAV for ultra-high voltage transmission lines according to claim 8, characterized in that, The coupled pose vector is hierarchically encoded according to continuity identifier and window stability before output; When the same candidate power extraction window maintains the same main conductor direction in consecutive image frames and the confidence of the coil arc segment meets the recognition conditions, the corresponding lateral deviation, longitudinal deviation, angular deviation, window stability and window confidence are combined into an effective coupled pose vector. When the continuity of the candidate power-taking window is interrupted, the previously mentioned effective coupled pose vector is retained, and the deviation component of the interrupted image frame is temporally interpolated and encoded.