A power line end-to-end target tracking method

Abstract

The application discloses a power transmission line end-to-end target tracking method, and relates to the technical field of power transmission line intelligent video monitoring and target tracking. The method comprises the following steps: acquiring a power transmission line video frame and a pan-tilt state information; inputting the video frame and the pan-tilt state information into an end-to-end target tracking network to obtain tracking features; calling a historical trajectory of a power transmission line scene target and fusing the historical trajectory with the tracking features in time sequence, and completing historical trajectory association based on a power transmission line topology and pan-tilt state constraints; determining whether the target is shielded and lost according to an association result, updating the historical trajectory if no shielding and loss occur, and buffering original historical trajectories corresponding to the lost target if short-time loss occurs; and outputting tracking results of each target in a current video frame based on an update state or a buffering state of the historical trajectory. The scheme is beneficial to avoiding the defect of multi-link error accumulation in a stage-by-stage algorithm, accurately distinguishing the position relationship between the target and the line equipment, and improving the continuity and accuracy of the whole scene tracking.

Description

Technical Field

[0001] This application relates to the field of intelligent video monitoring and target tracking technology for power transmission lines, and in particular to an end-to-end target tracking method for power transmission lines. Background Technology

[0002] Transmission lines cross complex areas such as mountains and suburbs. Dangerous sources such as floating objects, construction equipment, and birds in the passage can easily cause line faults. Video monitoring based on fixed-point pan-tilt cameras is the mainstream inspection method.

[0003] Existing traditional tracking algorithms assume a constant camera viewpoint and rely on the assumption of continuous target position movement between frames. This makes them unsuitable for handling sudden changes in image scale and overall background shifts caused by gimbal rotation, pitch adjustment, and optical zoom. Furthermore, traditional solutions employ a segmented architecture for detection, feature extraction, and trajectory matching. The independent operation of multiple modules leads to a cascading error. Moreover, the algorithms only extract target appearance and coordinate information, failing to incorporate constraints related to the transmission line topology and gimbal operating status. This results in a lack of spatial safety modeling for the transmission line, leading to frequent trajectory breaks, target loss, and identity jumps in scenarios involving target occlusion, distant small targets, and complex line background interference. Consequently, tracking stability is poor, and accurate trajectory data cannot be provided for subsequent line risk assessment. Summary of the Invention

[0004] The purpose of this application is to provide an end-to-end target tracking method for power transmission lines, which aims to solve the technical problems of target loss and error accumulation in existing technologies under conditions of gimbal-based variable viewing angles and complex line scenarios.

[0005] To achieve the above objectives, this application provides an end-to-end target tracking method for transmission lines, comprising: Acquire video frames of the power transmission line and the pan-tilt-zoom (PTZ) status information synchronously with the video frames; The video frames and the PTZ status information are input into an end-to-end target tracking network. The end-to-end target tracking network then fuses the image features, PTZ status features, and transmission line topology features to obtain the tracking features. The historical trajectory of the target in the power transmission line scene is retrieved and fused with the time sequence of the tracking features. The historical trajectory association is completed based on the power transmission line topology and the gimbal state constraints. Based on the association results, determine whether the target is occluded and lost contact. If no occlusion or loss of contact occurs, update the historical trajectory. If a short-term loss of contact occurs, cache the original historical trajectory corresponding to the lost target. Based on the update status or cache status of the historical trajectory, output the tracking results of each target in the current video frame.

[0006] In one embodiment, the video frame and the PTZ status information are input into an end-to-end target tracking network. The end-to-end target tracking network then fuses image features, PTZ status features, and transmission line topology features to obtain tracking features, including: The video frame and the gimbal status information are input into the end-to-end target tracking network; wherein, the gimbal status information includes the gimbal horizontal rotation angle, gimbal pitch angle, zoom ratio, and preset position number; The gimbal status information is encoded and converted to generate gimbal status encoding features corresponding to network operations; Multi-scale feature extraction and small target enhancement processing are performed on the video frames input to the end-to-end target tracking network to obtain refined image features; Based on the refined image features, the transmission line topology modeling is completed, and the transmission line topology features are extracted; The refined image features, the gimbal status coding features, and the transmission line topology features are fused to generate fused features for target tracking, which are then used as the tracking features.

[0007] In one embodiment, the gimbal status information is encoded and converted to generate gimbal status encoding features corresponding to network operations, including: The gimbal's horizontal rotation angle and its pitch angle are subjected to sine and cosine transformations respectively to obtain angular characteristics. The preset bit number is embedded and encoded to obtain the preset bit feature; The angle feature, the zoom ratio feature, and the preset position embedding feature are concatenated to generate a gimbal status coding feature with unified dimensions.

[0008] In one embodiment, multi-scale feature extraction and small target enhancement processing are performed on video frames input to the end-to-end target tracking network to obtain refined image features, including: The video frames input to the end-to-end target tracking network are fed into the backbone feature extraction network to output multi-scale initial image features. Based on the multi-scale initial image features, high-level semantic features and low-level detail features are obtained by splitting them. Upsampling and convolution fusion are used to fuse the high-level semantic features and the low-level detail features layer by layer to obtain multi-scale fused features. Weight coefficients are generated based on convolution operations and the Sigmoid activation function. The weight coefficients and the multi-scale fusion features are then multiplied element-wise to complete the weighted correction process, resulting in refined image features.

[0009] In one embodiment, transmission line topology modeling is completed based on the refined image features, and the transmission line topology features are extracted, including: All the refined image features are input into a preset transmission line topology modeling network for calculation; The characteristics of the transmission line topology are extracted from the conductor region, tower structure, insulator string position, line corridor region, and safety distance region through the transmission line topology modeling network, and combined to form the transmission line topology features.

[0010] In one embodiment, the refined image features, the gimbal status coding features, and the transmission line topology features are fused to generate fused features for target tracking, which are used as the tracking features, including: The refined image features are subjected to scale unification and channel integration processing to obtain regularized image fusion features; The dimensionality of the gimbal status coding features is expanded by broadcasting, and the spatial size is adjusted to be the same as that of the normalized image fusion features. The normalized image fusion features, PTZ status coding features, and transmission line topology features after dimension matching are spliced ​​together, and the weight coefficients corresponding to the topology features are generated through convolution operation and Sigmoid activation function. The topological features of the transmission line are weighted and corrected using the weighting coefficients, and gating fusion processing is performed on the three types of features to generate fused features for target tracking.

[0011] In one embodiment, the historical trajectory of the target in the power transmission line scene is retrieved and fused with the time sequence of tracking features. The historical trajectory association is completed based on the power transmission line topology and the gimbal state constraints, including: Historical trajectory features of targets in the power transmission line scene are retrieved, and the historical trajectory features and the tracking features of the current frame are fused in a temporal sequence using an attention mechanism to obtain temporal fusion features; By combining the historical trajectory features, the gimbal status of historical frames, and the gimbal status of the current frame, the estimated position of the historical target in the current frame is predicted, and candidate association features of the current target and historical trajectory association features are constructed respectively. The trajectory association score is calculated based on the appearance features, location information, topological relationship parameters of the current target candidate and the gimbal status parameters corresponding to the historical trajectory. The matching probability is calculated based on the association score and the historical trajectory association is completed through pairing and allocation.

[0012] In one embodiment, the estimated position of the historical target in the current frame is predicted by combining the historical trajectory features, the historical frame gimbal state, and the current frame gimbal state, and candidate association features for the current target and historical trajectory association features are constructed respectively, including: The motion prediction function is called to solve the estimated target position corresponding to each historical trajectory based on the historical target position, historical motion state, historical frame gimbal state and current frame gimbal state corresponding to a single historical trajectory. When the gimbal states are similar, the target's estimated position is calculated based on the target's motion state. When the gimbal zooms or rotates, the target's estimated position is corrected by combining the gimbal parameters. Extract the appearance features, location encoding, category encoding, topological features, and current gimbal encoding of the current target and perform feature mapping to obtain the candidate associated features of the current target; Extract the historical trajectory memory features, estimated position codes, historical category codes, historical topology features, and historical gimbal codes corresponding to the single historical trajectory, and perform feature mapping to obtain the historical trajectory association features.

[0013] In one embodiment, a trajectory association score is calculated based on the appearance features, location information, topological relationship parameters, and gimbal status parameters corresponding to the current target candidate and the historical trajectory. A matching probability is then calculated based on the association score, and the historical trajectory association is completed through pairing and allocation. This includes: Feature similarity is calculated based on the current target candidate association features and historical trajectory association features. The position intersection-union ratio is obtained by combining the current target candidate position coordinates and the historical trajectory estimated position. The topological difference is calculated based on the current target candidate topological features and the historical trajectory topological features. The gimbal state difference is obtained by relying on the current gimbal encoding and the historical gimbal encoding. The category consistency is obtained by comparing the current target candidate category and the historical trajectory category. The feature similarity, the position intersection-union ratio, the topological difference, the gimbal state difference, and the category consistency are weighted and calculated to obtain the association score between a single candidate and the trajectory. All association scores are normalized to obtain the pairwise association probabilities, and unmatched scores are introduced to represent the target disconnection scenario; Filter matching combinations that exceed the preset association threshold, and perform one-to-one matching and allocation according to the association probability from high to low. When the same target is matched by multiple trajectories, only the pair with the best probability is retained, and the remaining unmatched trajectories are put into the waiting state.

[0014] In one embodiment, the method further includes: Based on the tracking results, a risk score corresponding to the target is calculated, and multiple risk levels are divided by combining the risk score with a preset grading threshold. The tracking results include target category, movement trajectory, spatial relationship between the target and power transmission facilities, and duration of the target's stay in the risk area; Based on the risk level obtained from the classification, when the risk score exceeds the corresponding preset classification threshold, the target corresponding hazard type is matched and an alarm message is output; The alarm information includes target information, camera position information, and trajectory data.

[0015] The above-mentioned technical solution of this application has at least the following beneficial technical effects: The technical solution of this application constructs an end-to-end tracking network that fuses three types of features: image, PTZ status, and line topology. It then combines topology and PTZ constraints to achieve historical trajectory temporal correlation. This can rely on multi-dimensional features to offset the image distortion interference caused by PTZ rotation and zoom, breaking the premise limitation of fixed perspective in traditional algorithms. This helps to avoid the defects of multi-stage algorithm error accumulation. At the same time, it establishes target space constraints based on line topology, accurately distinguishes the positional relationship between the target and line equipment, reduces trajectory interruption and identity confusion caused by occlusion and interference from small targets, improves the continuity and accuracy of tracking in all scenarios, and outputs reliable tracking data for subsequent risk classification and assessment of transmission lines. Attached Figure Description

[0016] Figure 1 This application provides a flowchart illustrating one embodiment; Figure 2 This is a general flowchart of a specific embodiment of this application; Figure 3 This is an end-to-end network flowchart of a specific embodiment of this application; Figure 4 This is an attention trajectory association diagram of a specific embodiment of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to specific embodiments and accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of this application. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0018] The embodiments described in this application are only some, not all, of the embodiments described herein. All other embodiments obtained by those skilled in the art based on the embodiments described herein without inventive effort are within the scope of protection of this application.

[0019] Transmission lines traverse complex areas such as mountains, forests, and suburbs. Hazards such as floating objects, construction equipment, and birds within the transmission lines can easily cause line faults. Video monitoring using fixed-point pan-tilt-zoom (PTZ) cameras is the mainstream inspection method. Existing traditional tracking algorithms assume a constant camera viewpoint and rely on the assumption of continuous target position movement between frames. This makes them unsuitable for the sudden changes in image scale and overall background shifts caused by PTZ rotation, pitch adjustment, and optical zoom. Furthermore, traditional solutions employ a segmented architecture for detection, feature extraction, and trajectory matching. The independent operation of multiple modules leads to a cascading error. The algorithms only extract target appearance and coordinate information, failing to incorporate constraints related to the transmission line topology and PTZ operating status. This lack of spatial safety modeling for the line results in frequent trajectory breaks, target loss, and identity jumps in scenarios involving target occlusion, distant small targets, and complex line background interference. The tracking stability is poor, and accurate trajectory data cannot be provided for subsequent line risk assessment.

[0020] To address the aforementioned technical issues, this application provides an end-to-end target tracking method for transmission lines. The technical solution of this application constructs an end-to-end tracking network that fuses three types of features: image, PTZ status, and line topology. It then combines topology and PTZ constraints to achieve historical trajectory temporal correlation. This multi-dimensional feature-based approach can offset image distortion interference caused by PTZ rotation and zoom, breaking the limitation of fixed perspective in traditional algorithms. This helps avoid the cumulative error defects of multi-stage algorithms. Simultaneously, by establishing target spatial constraints based on line topology, it accurately distinguishes the positional relationship between targets and line equipment, reducing trajectory interruptions and identity confusion caused by occlusion and interference from small targets. This improves the continuity and accuracy of tracking across all scenarios, providing reliable tracking data for subsequent risk classification and assessment of transmission lines.

[0021] In one implementation, please refer to Figure 1 The end-to-end target tracking method for this transmission line includes the following steps: Step S1: Obtain video frames of the transmission line and the corresponding pan-tilt-zoom (PTZ) status information. The specific implementation process of Step S1 is as follows: Using PTZ monitoring equipment fixedly installed on transmission towers, high points along the line corridor, around substations, or other monitoring points, obtain the video frame sequence under the transmission line environment, and obtain the PTZ status information corresponding to each video frame. The PTZ status information includes the PTZ horizontal rotation angle, pitch angle, zoom ratio, focal length, preset position number, and PTZ movement speed. The PTZ status information corresponding to the t-th frame is represented as follows: ,in, Indicates the horizontal rotation angle of the gimbal. Indicates the pitch angle of the gimbal. Indicates zoom ratio or focal length status. This indicates the preset position number of the gimbal.

[0022] By introducing gimbal status information, the subsequent end-to-end tracking network can perceive the impact of gimbal rotation, pitch changes, zoom magnification, and preset position switching on target position, target scale, and background changes, thereby improving the target tracking stability in fixed gimbal patrol monitoring scenarios.

[0023] In addition, image preprocessing is performed on the video frames to obtain the image to be tracked. Image preprocessing includes at least one of the following: image enhancement, denoising, distortion correction, scale normalization, low-light enhancement, fog suppression, backlight compensation, and image jitter reduction.

[0024] In fixed-mount pan-tilt-zoom (PTZ) monitoring scenarios, although the camera's installation position is relatively fixed, the PTZ itself undergoes horizontal rotation, pitch adjustment, zoom, and preset position rotation operations. This can lead to changes in the overall background, abrupt shifts in target position, and changes in target scale between adjacent frames. Therefore, this invention retains PTZ state information during the preprocessing stage and uses this information as a crucial input for subsequent feature fusion and trajectory association.

[0025] Step S2: Input the video frames and PTZ status information into the end-to-end target tracking network. The end-to-end target tracking network fuses image features, PTZ status features, and transmission line topology features to obtain tracking features. In one embodiment, step S2 further includes the following specific steps: Step S21: Input the video frame and gimbal status information into the end-to-end target tracking network; wherein, the gimbal status information includes the gimbal horizontal rotation angle, gimbal pitch angle, zoom ratio and preset position number; Step S22: Encode and convert the gimbal status information to generate gimbal status coding features corresponding to network operations; in one embodiment, step S22 further includes the following specific steps: perform sine and cosine transformations on the gimbal horizontal rotation angle and gimbal pitch angle respectively to obtain angle features; embed the preset position number to obtain preset position features; and concatenate the angle features, zoom magnification features and preset position embedding features to generate gimbal status coding features with unified dimensions.

[0026] Step S23: Perform multi-scale feature extraction and small target enhancement processing on the video frames input to the end-to-end target tracking network to obtain refined image features; in one embodiment, step S23 further includes the following specific steps: send the video frames input to the end-to-end target tracking network into the backbone feature extraction network to output multi-scale initial image features; based on the multi-scale initial image features, split to obtain high-level semantic features and low-level detail features, and use upsampling and convolution fusion to fuse the high-level semantic features and low-level detail features layer by layer to obtain multi-scale fused features; generate weight coefficients based on convolution operations and the Sigmoid activation function, and use the weight coefficients and multi-scale fused features element-wise multiplication to complete the weighted correction processing to obtain refined image features.

[0027] Step S24: Complete the transmission line topology modeling based on refined image features and extract the transmission line topology features; in one embodiment, step S24 further includes the following specific steps: input the refined image features into a preset transmission line topology modeling network for calculation; extract conductor region features, tower structure features, insulator string position features, line corridor region features and safety distance region features through the transmission line topology modeling network, and combine them to form transmission line topology features.

[0028] Step S25: Fuse the refined image features, PTZ status coding features, and transmission line topology features to generate fused features for target tracking, which serve as the tracking features. In one embodiment, step S25 further includes the following specific steps: performing scale unification and channel integration processing on the refined image features to obtain normalized image fusion features; performing dimensionality broadcasting expansion on the PTZ status coding features to adjust the spatial size to be the same as the normalized image fusion features; performing feature concatenation on the dimensionally matched normalized image fusion features, PTZ status coding features, and transmission line topology features, and generating weight coefficients corresponding to the topology features through convolution operations and the Sigmoid activation function; using the weight coefficients to perform weighted correction on the transmission line topology features, and performing gated fusion processing on the three types of features to generate fused features for target tracking.

[0029] The specific implementation process of step S2 is as follows: The preprocessed video frames and corresponding PTZ status information are input into the end-to-end target tracking network. The end-to-end target tracking network is used to complete transmission line image feature extraction, PTZ status encoding, transmission line topology modeling, small target enhancement, feature fusion, target candidate generation, and target trajectory feature output within a unified model. The end-to-end target tracking network includes an image input module, a PTZ status encoding module, a multi-scale feature extraction module, a small target enhancement module, a transmission line topology modeling module, a feature fusion module, a target query generation module, a tracking decoding module, and a multi-task output module. Details are as follows: (1) Gimbal status coding module Fixed-mount pan-tilt-zoom (PTZ) monitoring equipment undergoes horizontal rotation, pitch adjustment, zoom magnification, and preset position switching during operation. To enable the network to perceive the impact of PTZ movement on target position, target size, and background changes, this invention encodes PTZ state information into network-processable state features.

[0030] Specifically, no. t The gimbal status information corresponding to the frame is represented as follows: ,in, Indicates the horizontal rotation angle of the gimbal. Indicates the pitch angle of the gimbal. Indicates zoom ratio or focal length status. This indicates the preset position number of the gimbal.

[0031] To avoid discontinuities caused by angular periodicity, the gimbal angle is converted into sine and cosine forms, and combined with zoom magnification and preset bit encoding to obtain gimbal state characteristics: ,in, This represents the gimbal state coding features, and MLP stands for Multilayer Perceptron. This indicates the embedded encoding of the preset bit number.

[0032] This structure can distinguish the appearance differences of the same target under different gimbal angles, focal lengths, and preset positions, thereby improving the target tracking stability under gimbal rotation and zoom monitoring conditions.

[0033] (2) Multi-scale feature extraction module The current frame image is input into the backbone feature extraction network to extract image features at multiple scales: ,in, This represents the backbone feature extraction network. Represents the image features at the l-th scale. L Indicates the number of feature scales.

[0034] Since fixed pan-tilt-zoom (PTZ) monitoring of power transmission lines typically involves long-distance shooting, large proportions of small targets, and complex background textures, this invention further employs a multi-scale feature fusion structure to fuse high-level semantic information with low-level detail information, resulting in multi-scale enhanced features: ,in, Let Up represent the l-th scale feature after fusion, Up represent the upsampling operation, and Conv represent the convolution transformation.

[0035] This structure can simultaneously preserve the semantic information and edge detail information of the target, and is suitable for small-scale or weakly textured targets such as birds, foreign objects, plastic films, kites, balloons, crane ends, and treetops in power transmission line scenarios.

[0036] (3) Small target enhancement module To address the issues of missed detection and discontinuous tracking of small targets in long-distance monitoring of power transmission lines, this invention incorporates a small target enhancement module after multi-scale features. This module locally enhances shallow, high-resolution features and suppresses interference from slender background structures such as conductors, tree branches, and tower materials.

[0037] In one embodiment, the small target enhancement feature is represented as: ,in, This represents the enhanced l-th scale feature. represents the Sigmoid activation function, and ⊙ represents element-wise multiplication. This structure is equivalent to generating a spatial attention weight for each scale feature, making the network pay more attention to suspected target regions, thereby reducing the impact of complex background regions on target detection and tracking.

[0038] (4) Transmission line topology modeling module In transmission line scenarios, target risk is closely related to its position relative to conductors, towers, insulator strings, line corridors, and safe distance zones. Therefore, this invention incorporates a transmission line topology modeling module within an end-to-end target tracking network. This module takes multi-scale enhanced features as input and outputs transmission line topology features: ,in, Represents the topology modeling network of transmission lines. This describes the topological characteristics of a transmission line. Transmission line topological characteristics include at least one of the following: , in, Indicates the characteristics of the conductor region or conductor centerline. Indicates the structural features of the tower. Indicates the positional characteristics of the insulator string. Indicates the characteristics of the route corridor area. This represents the characteristics of the safe distance area. This structure can obtain prior structural information in the transmission line scenario, enabling subsequent trajectory association to not only depend on the target's appearance and location, but also to combine the spatial relationship between the target and the key structures of the transmission line.

[0039] (5) Fusion of image features, gimbal status features and topological features After obtaining multi-scale image features, gimbal status coding features, and transmission line topology features, this invention generates fused features for target detection and trajectory association through a feature fusion module.

[0040] Specifically, the multi-scale enhancement features are subjected to unified scale transformation and channel integration to obtain image fusion features. Then, the gimbal status characteristics are... Expanded to the same spatial size as image features and fused together with transmission line topology features: , where Ft represents the fusion feature of the current frame.

[0041] Feature fusion can be achieved using a gated fusion method: , , in, The adaptive weights of the topological features are represented by Broadcast(et), which means expanding the PTZ state coding to the same spatial size as the image features, and [] represents feature stitching. This structure allows the network to adaptively enhance relevant topological features when the target is located near conductors, towers, line channels, or safe distance areas; and it can use topological constraints to reduce false detections and false associations when the background structure is similar to the target appearance.

[0042] (6) Generation of target query and trajectory query To achieve end-to-end target tracking, this invention simultaneously sets up target detection queries and historical trajectory queries in the network. Target detection queries are used to discover newly appearing targets in the current frame, while historical trajectory queries are used to continue the trajectories of targets that existed in the previous time step. The target query set is represented as follows: ,in, This indicates a target detection query. This indicates a trajectory query generated by the historical trajectory memory module.

[0043] Regarding the historical trajectory Its trajectory query can be generated by combining historical appearance features, historical location, historical topology, and historical gimbal status: ,in, Indicates the characteristics of historical trajectory memory. ψ represents the historical target location, and ψ represents the location encoding function. Indicates historical topological relationship characteristics, This represents the gimbal status characteristics corresponding to the historical frame.

[0044] This structure allows historical trajectories to directly participate in the current frame decoding process in the form of queries, thereby achieving integrated processing of target detection and trajectory continuation.

[0045] (7) Tracking decoding and multi-task output target query set and fusion features Input to the tracking and decoding module to obtain the target decoding features: Where Decoder represents the tracking decoder, This represents the target decoding features. The tracking decoder establishes a correspondence between the target query and image fusion features through an attention mechanism, enabling the network to locate existing targets, newly added targets, and suspected risk targets in the current frame image.

[0046] The multi-task output module outputs based on the target decoding features: ,in, Indicates the target category. Indicates the target location. Indicates the target confidence level. This represents the embedded features of the target appearance. Indicates the target occlusion state. This structure represents the topological relationship between the target and the transmission line structure. This structure enables the end-to-end target tracking network to complete image feature extraction, PTZ state perception, transmission line topology modeling, small target enhancement, target candidate generation, and trajectory feature output within a unified model, providing a foundation for subsequent attention trajectory association. After obtaining the fused features of the current frame, historical trajectory information is read from the historical trajectory memory module. Historical trajectory information includes the trajectory number, target category, target location, target confidence, appearance features, motion state, occlusion state, historical PTZ state, and historical topological relationship of the historical target. Since fixed-installation PTZ monitoring equipment may cycle according to preset positions, or zoom in or manually control rotation after detecting abnormal targets, there may be changes in viewpoint, scale, and background between adjacent frames or adjacent monitoring segments. To improve the continuity of target tracking, this invention designs a temporal feature fusion module to fuse the features of the current frame with the historical trajectory features. During the fusion process, the network adjusts the contribution weight of historical trajectory features to the current frame based on the similarity between the current frame PTZ state and the historical frame PTZ state. When the previous frame and the historical frame are in the same or similar preset position, similar focal length, or similar gimbal attitude, increase the weight of the corresponding historical trajectory features; when the gimbal undergoes a large angle rotation, pitch change, or zoom change, reduce the influence of irrelevant historical features.

[0047] In one implementation, the temporally fused features are represented as follows: ,in, Indicates the fusion features of the current frame. Represents historical trajectory memory features, and Attn represents the attention fusion operation. This represents the tracking features of the current frame after fusing historical trajectory information. This structure can enhance the target recognition and tracking capabilities of the current frame by utilizing historical trajectory information, and is especially suitable for scenarios such as short-term target occlusion, gimbal zoom, small target blur, image shaking, and complex background interference.

[0048] Step S3: Retrieve the historical trajectory of the target in the transmission line scene and fuse it with the time series of tracking features, and complete the historical trajectory association based on the transmission line topology and PTZ state constraints; in one embodiment, step S3 further includes the following specific steps: Step S31: Retrieve the historical trajectory features of the target in the power transmission line scene, and use an attention mechanism to perform temporal fusion of the historical trajectory features and the tracking features of the current frame to obtain temporal fusion features; Step S32: Combine historical trajectory features, historical frame gimbal status, and current frame gimbal status to predict the estimated position of the historical target in the current frame, and construct current target candidate association features and historical trajectory association features respectively; in one embodiment, step S32 further includes the following specific steps: call the motion prediction function, and solve the target estimated position corresponding to each historical trajectory based on the historical target position, historical motion state, historical frame gimbal status, and current frame gimbal status corresponding to a single historical trajectory; wherein, when the gimbal status is similar, the target estimated position is estimated based on the target motion state, and the target estimated position is corrected by combining the gimbal parameters when the gimbal zooms or rotates; extract the appearance features, position code, category code, topological features, and current gimbal code of the current target and perform feature mapping to obtain the current target candidate association features; extract the historical trajectory memory features, estimated position code, historical category code, historical topological features, and historical gimbal code corresponding to a single historical trajectory and perform feature mapping to obtain the historical trajectory association features.

[0049] Step S33: Calculate the trajectory association score based on the appearance features, location information, topology parameters, and gimbal status parameters corresponding to the current target candidate and the historical trajectory. Calculate the matching probability based on the association score and complete the historical trajectory association through pairing and allocation. In one embodiment, step S33 further includes the following specific steps: calculating feature similarity based on the current target candidate association features and historical trajectory association features; obtaining the position intersection-union ratio by combining the current target candidate position coordinates and the historical trajectory estimated position; calculating the topological difference based on the current target candidate topological features and the historical trajectory topological features; obtaining the gimbal state difference based on the current gimbal encoding and the historical gimbal encoding; comparing the current target candidate category and the historical trajectory category to obtain category consistency; performing weighted calculations on feature similarity, position intersection-union ratio, topological difference, gimbal state difference, and category consistency to obtain the association score of a single candidate and trajectory; normalizing all association scores to obtain the pairwise association probability, and introducing unmatched scores to represent the target disconnection scenario; filtering matching combinations that exceed the preset association threshold, and performing one-to-one matching allocation according to the association probability from high to low; when the same target is matched by multiple trajectories, only the pair with the best probability is retained, and the remaining unmatched trajectories are transferred to the waiting allocation state.

[0050] Step S4: Determine whether the target is occluded and lost based on the association results. If no occlusion or loss of connection occurs, update the historical trajectory. If a short-term loss of connection occurs, cache the original historical trajectory corresponding to the lost target.

[0051] Step S5: Based on the update status or cache status of historical trajectories, output the tracking results of each target in the current video frame.

[0052] The specific implementation process of steps S3 to S5 is as follows: The temporally fused tracking features of the current frame are input into the attention trajectory association module. The network then performs association matching between the current frame target candidates and historical trajectories to determine the target trajectory number and update the trajectory status. The trajectory association of this invention comprehensively considers the target appearance, target motion, gimbal status, and transmission line topology, and calculates the association probability between the current target and historical trajectories through an attention mechanism.

[0053] (1) Construct the current target candidate set and the historical trajectory set. Based on the output of the end-to-end target tracking network, the target candidate set for the current frame is obtained: , among which, the The current target candidates are represented as follows: ,in, Indicates the target category. Indicates the target location. Indicates the target confidence level. This represents the embedded features of the target appearance. Indicates an occluded state. This represents the topological relationship features of the target. The historical trajectory set is represented as: , among which, the Each historical trajectory includes trajectory number, historical location, historical category, historical appearance memory features, historical motion state, historical gimbal state, and historical topological relationship.

[0054] (2) Predicting trajectory position based on gimbal status and historical motion status In fixed-mount pan-tilt-zoom (PTZ) monitoring scenarios, changes in the target's position in the image are caused not only by the target's own movement but also by changes in the PTZ's horizontal rotation, pitch adjustment, and zoom. Therefore, this invention predicts the possible position of the historical target in the current frame based on historical trajectory states and PTZ state changes before trajectory association. For the first... The historical trajectory, and its predicted position in the current frame, is represented as follows: ,in, Indicates the predicted location. Indicates the location of the historical target. Indicates the state of historical movement. and These represent the gimbal status in historical frames and the current frame, respectively. This represents the motion prediction function. When the gimbal is in the same or similar preset position, the motion prediction function mainly predicts the target position based on the target's historical motion state; when the gimbal rotates or zooms, the motion prediction function combines the changes in the gimbal state to correct the predicted position and scale of the target in the image.

[0055] (3) Constructing trajectory association features To perform attention association, this invention constructs current target candidate features and historical trajectory features respectively. Current target candidate... The association features can be represented as: Historical trajectory The association features can be represented as: , in, Indicates the current target candidate association features, Indicates the correlation features of historical trajectories, Indicates the characteristics of historical trajectory memory. The positional encoding function is represented by Emb, and the category encoding function is represented by Emb. Indicates the current gimbal status code. This represents the gimbal status code corresponding to the historical trajectory. Through the trajectory association process, this structure simultaneously possesses the capabilities of appearance discrimination, location discrimination, category constraint, gimbal status perception, and transmission line topology constraint.

[0056] (4) Calculate the attention association score For each historical trajectory and each current target candidate The attention trajectory association module calculates the association score between the two. In one implementation, the association score is expressed as: ,in, Representing historical trajectory i With current goal j The correlation score; WQ and WK represent the learnable parameter matrix; d Indicates the feature dimension; This represents the intersection-union ratio (IUU) between the predicted location and the current target location. Indicates differences in the topology of transmission lines; Indicates differences in gimbal status; This represents the category consistency function, which takes a larger value when the historical trajectory category is consistent with the current target category. The weighting coefficients are as follows: The first term in the above formula represents the depth feature similarity between the historical trajectory and the current target; the second term represents the consistency of motion position; the third term constrains the spatial relationship continuity between the target and the transmission line structure; the fourth term constrains the consistency of the gimbal's viewing angle or focal length; and the fifth term enhances the matching probability between targets of the same category. Through this comprehensive association score, this invention can avoid false associations caused by relying solely on target box distance or a single appearance feature. For example, when background structures such as conductors, tree branches, and tower materials are similar in appearance to the target, topological relationship constraints can reduce false matching; when the gimbal zooms or switches preset positions, gimbal state constraints can reduce cross-viewpoint erroneous associations.

[0057] (5) Calculate the trajectory association probability After obtaining the correlation score, the historical trajectory is obtained through attention normalization. i The probability of association with each current target candidate: ,in, Representing historical trajectory i With current goal j The probability of association, Representing historical trajectory i The unmatch score is the score for when no target is matched in the current frame. The unmatch score indicates when the target is occluded, leaves the frame, is temporarily lost, or becomes invisible due to gimbal switching. By introducing the unmatch score, the network can avoid forced incorrect matching when the target is temporarily invisible. When the maximum association probability corresponding to a certain historical trajectory satisfies the following formula, the historical trajectory is associated with the corresponding current target, and the current target inherits the trajectory number of that historical trajectory. The formula is: ,in, This indicates the correlation threshold.

[0058] (6) One-to-one trajectory assignment and conflict handling Since the same current target should not correspond to multiple historical trajectories simultaneously, this invention performs one-to-one trajectory allocation after calculating the association probability. In one implementation, the system establishes matching relationships between historical trajectories and the current target sequentially based on the association probability from high to low; when multiple historical trajectories simultaneously match the same current target, the matching relationship with the highest association probability is retained, and the remaining historical trajectories either reselect a suboptimal target or enter an unmatched state. Alternatively, a globally optimal allocation can be performed based on the association score matrix to maximize the overall association score between all historical trajectories and the current target. This approach avoids the problem of multiple trajectories competing for the same target when multiple targets converge at close range, construction machinery overlaps, bird flocks are active, or foreign objects are mixed with background structures.

[0059] (7) New target generation For target candidates in the current frame that have not been successfully associated with any historical trajectory, if their target confidence is greater than a preset new target threshold, they are treated as new targets and a new trajectory number is generated. That is, when a target candidate... satisfy: If a data point has not been assigned to any historical trajectory, the system initializes a new trajectory for it. ,in, This represents the threshold for generating new targets. This indicates the newly generated trajectory number. This mechanism applies to situations such as construction machinery entering the monitoring screen, birds flying into the tower area, foreign objects entering the line passage, and the emergence of new smoke or fire spots.

[0060] (8) Temporary storage and occlusion handling of unmatched trajectories For historical trajectories that fail to be successfully associated with the current target, the system does not immediately delete them. Instead, it determines whether they have entered a temporarily lost state by combining the obstruction status, historical location, PTZ status, and transmission line topology. A historical trajectory can be marked as temporarily lost when any of the following conditions are met: the predicted target location is still within the line corridor area or safety monitoring area; the target may be obstructed by conductors, towers, insulator strings, trees, or mountains; the PTZ has rotated, zoomed, or switched preset positions, causing the target to be temporarily invisible; or the historical trajectory has a high risk level and needs to be retained in the tracking memory. Trajectories marked as temporarily lost retain their trajectory number, historical appearance characteristics, historical location, historical PTZ status, and historical topology for trajectory recovery when the target reappears.

[0061] (9) Track memory update When the trajectory of history i With current goal j After successful association, the system updates the target location, appearance memory features, motion state, topology, and gimbal status of the trajectory. The trajectory appearance memory features are updated as follows: ,in, This represents the updated trajectory memory features. This represents the trajectory memory features before the update. This indicates the current target appearance embedding features. This represents the historical memory retention coefficient, and Norm indicates the normalization process. The target's motion state can be updated based on the current target position and historical positions. ,in, This represents the updated motion state. Simultaneously, it incorporates the topological features of the current target. and current gimbal status characteristics Write to the trajectory memory module for trajectory association during the next frame or subsequent preset position retracement.

[0062] (10) Target Re-identification and Trajectory Recovery For trajectories already marked as temporarily lost, when a new target candidate appears in subsequent video frames, the system recalculates the association score between the target candidate and the temporarily lost trajectory to determine whether they belong to the same target. If the target meets the association conditions in terms of appearance features, predicted location, PTZ status, and transmission line topology, its original trajectory number is restored. At that time, the system considers the current target With temporarily lost track If the target belongs to the same target, its trajectory number will be restored. Through the above-mentioned target re-identification and trajectory restoration mechanism, the system can re-associate the target after it is briefly obscured by conductors, tower materials, or trees; at the same time, it can also cope with the problem of the target being temporarily invisible due to factors such as pan-tilt zoom and preset position switching. As a result, the occurrence of trajectory interruption and duplicate numbering can be reduced, and the continuity of target tracking can be improved in complex monitoring scenarios of power transmission lines.

[0063] (11) End-to-end correlation training To enable collaborative optimization of object detection, feature extraction, and trajectory association within a unified network, this invention introduces a trajectory association loss during the training phase. In one implementation, the trajectory association loss is expressed as: ,in, Indicates a true association tag; when historical trajectory i With current goal j When belonging to the same goal otherwise The overall training loss of the network can include object classification loss, bounding box regression loss, topological feature loss, occlusion state loss, and trajectory association loss. ,in, These are the weight coefficients for the corresponding loss terms. Through the end-to-end training method described above, the network can simultaneously learn target detection capabilities, transmission line topology awareness capabilities, occlusion judgment capabilities, and trajectory association capabilities, reducing error accumulation in traditional staged target detection and trajectory matching methods.

[0064] (12) Perform occlusion detection and trajectory recovery In power transmission line environments, targets may be obstructed by conductors, tower materials, insulator strings, trees, mountains, or other objects, rendering them undetectable for short periods. To address this, this invention incorporates an obstruction state prediction unit within the end-to-end target tracking network to determine the target's current visibility status. Visibility status includes no obstruction, partial obstruction, severe obstruction, and temporarily invisible states. When a historical trajectory fails to match the currently detected target in several consecutive frames, but its predicted location indicates it remains within a reasonable line corridor area, safety monitoring area, or historical activity area, the system does not immediately delete the trajectory but marks it as temporarily lost. In this case, the system retains the historical appearance features, historical location, historical PTZ status, historical topology, and trajectory number corresponding to the trajectory for re-matching and trajectory continuation in subsequent frames. In subsequent frames, if a reappearing target meets the matching conditions for the historical trajectory in terms of appearance features, movement trend, PTZ status, and topology, the system re-associates the target to the original historical trajectory, restores its original trajectory number, and continues to update the target's trajectory status. The aforementioned occlusion recovery mechanism is applicable to scenarios such as birds reappearing in the frame after crossing power lines, construction machinery being briefly obscured by trees or towers, and foreign objects briefly leaving the visible area due to wind and then re-entering the frame. By retaining trajectory information when the target is temporarily invisible and completing trajectory recovery after the target reappears, the continuity and stability of target tracking in complex environments such as power transmission lines can be improved.

[0065] (13) Output target tracking results The end-to-end target tracking network outputs target tracking results for each frame. These results include at least one of the following: target category, target location, target confidence level, target trajectory number, target occlusion status, spatial relationship between the target and the transmission line structure, and target risk level. The spatial relationship between the target and the transmission line structure includes the distance from the target to the conductor, the distance from the target to the tower, the distance from the target to the insulator string, whether the target has entered the line corridor area, whether the target has entered the safe distance area, and whether the target is close to crossarms or other critical equipment areas. By outputting the target trajectory number, the system can continuously record the status of the same target in consecutive video frames or multiple pan-tilt-zoom (PTZ) preset position monitoring segments, providing a basis for subsequent risk assessment.

[0066] Furthermore, to improve the timeliness and reliability of risk assessment and avoid false alarms or omissions in line warnings caused by momentary misjudgments, in one embodiment, the following steps are included after step S5: calculating the risk score corresponding to the target based on the tracking results, and dividing the risk level into multiple risk levels by combining the risk score with a preset classification threshold; wherein, the tracking results include the target category, movement trajectory, spatial relationship between the target and the transmission facility, and the duration of the target's stay in the risk area; according to the risk level obtained, when the risk score exceeds the corresponding preset classification threshold, matching the hidden danger type corresponding to the target and outputting alarm information; wherein, the alarm information includes target information, machine position information, and trajectory data.

[0067] The specific implementation process of risk warning is as follows: (1) Risk assessment of transmission lines based on tracking results After obtaining continuous target tracking results, a risk assessment is performed based on the target category, target trajectory, spatial relationship between the target and the transmission line structure, and target duration. The risk assessment considers at least the following factors: target category, distance between the target and the conductor or critical equipment, whether the target entered the line corridor area, whether the target entered the safe distance area, whether the target continuously approached the conductor, the target's dwell time within the risk area, whether the target reappeared after being obstructed, or whether it repeatedly approached or showed a continuous movement trend. The risk score is expressed as: ,in, Indicates risk score, Indicates the target category. Indicates the distance between the target and the guide wire or critical equipment. This indicates the speed at which the target approaches the conductor. This indicates the time the target spends within the risk zone. This indicates the degree of overlap between the target and the line corridor area or safe distance area. When the risk score reaches a preset threshold, the system outputs the risk level and alarm information. The risk level can include normal, attention, warning, and alarm. When the target has not entered the line corridor and is far from the conductor, it is judged as normal; when the target enters the vicinity of the line corridor but has not yet approached the safe distance area, it is judged as attention; when the target continues to approach the conductor or stays near the safe distance area, it is judged as warning; when the target enters the safe distance area, or there is a significant risk of external damage, foreign objects, tree obstruction, wildfire, bird damage, etc., it is judged as alarm.

[0068] (2) Output alarm information and related data When a safety risk to the transmission line is determined to exist, the system outputs alarm information. The alarm information includes at least one of the following: target category, target trajectory number, target location, risk level, risk cause, alarm time, PTZ preset position number, corresponding video clip, and target trajectory data. In one implementation, when the target category is construction machinery, cranes, pump trucks, or engineering vehicles, if they enter the transmission line corridor area and continuously approach the conductor, or if their booms or tall components enter the safe distance zone, an external damage risk alarm is output, according to the following rules: When the target category is trees, tree crowns, or branches, if the distance between them and the conductor is less than a preset tree obstacle safety threshold, or if they show a trend of approaching the conductor within a continuous monitoring period, a tree obstacle risk alarm is output. When the target category is kites, plastic film, balloons, ribbons, or other foreign objects, if they are located near the conductor, suspended in the conductor area, or repeatedly approach the conductor under wind force, a foreign object hazard alarm is output.

[0069] When the target category is birds, a bird hazard risk warning is issued if they frequently linger near tower crossarms, insulator strings, or conductors. When the target category is bird nests or suspected nesting materials, a bird nest hazard warning is issued if they are located near tower crossarms, insulator strings, or conductors. When the target category is smoke, fire spots, or wildfire areas, a wildfire risk warning is issued if the smoke or fire area overlaps with the line corridor area, or if the smoke spreads towards the transmission line. Through the above methods, this invention achieves closed-loop processing from fixed pan-tilt-zoom video acquisition, end-to-end target tracking, continuous trajectory maintenance to transmission line risk assessment and alarm output.

[0070] In a specific embodiment of this application, please refer to Figures 2 to 4 , Figure 2 This is a schematic diagram of the overall process of end-to-end target tracking and risk assessment for transmission lines in this embodiment. After the step is started, the overall process sequentially completes the input of multi-source data from video and PTZ, image preprocessing, multi-dimensional feature extraction and fusion, and multi-parameter trajectory association guided by attention mechanism. After the branch line is identified as having occlusion or loss of connection, the trajectory cache is restored or the regular trajectory is updated. After the full-dimensional target tracking result is output, the line risk is quantitatively scored based on the target parameters. Alarm information is selectively output based on the comparison result of the risk score and the preset threshold. Finally, the process is cyclically determined whether to continue frame-by-frame detection until the process ends. Figure 3 This is a schematic diagram of the end-to-end feature extraction and temporal fusion refinement process in this embodiment. It includes four sequential execution steps: input of video frames and gimbal attitude raw data, image denoising and distortion correction preprocessing, multi-source feature extraction that fuses image appearance features and gimbal topology coding, and temporal feature aggregation processing that combines historical trajectory memory. This provides standardized feature data support for subsequent trajectory association operations. Figure 4This is a detailed flowchart of the attention trajectory association and trajectory lifecycle management process in this embodiment. Multi-scale feature extraction, associated feature construction, and attention association score calculation are completed around the candidate target. After probability normalization and optimal allocation, target trajectory matching is achieved. If the matching is successful, the historical trajectory parameters are inherited and the target status is updated. If the matching fails, the candidate target and the historical trajectory are distinguished. The candidate target is determined to create a new trajectory or be directly discarded based on the confidence threshold. The lost historical trajectory is marked as lost based on the occlusion temporary storage condition. Then, the trajectory ID is restored or invalid trajectories are cleared after timeout based on the matching results of subsequent frames. Finally, the complete association results and the updated trajectory set are summarized and output.

[0071] This application aims to protect an end-to-end target tracking method for transmission lines. The technical solution constructs an end-to-end tracking network that integrates three types of features: image, PTZ status, and line topology. It then combines topology and PTZ constraints to achieve historical trajectory temporal correlation. This method can rely on multi-dimensional features to offset the image distortion interference caused by PTZ rotation and zoom, breaking the limitation of fixed perspective in traditional algorithms. It helps to avoid the defects of multi-stage algorithm error accumulation. At the same time, it establishes target spatial constraints based on line topology, accurately distinguishes the positional relationship between the target and line equipment, reduces trajectory interruption and identity confusion caused by occlusion and small target interference, improves the continuity and accuracy of tracking in all scenarios, and outputs reliable tracking data for subsequent risk classification and assessment of transmission lines.

[0072] It should be understood that the specific embodiments described above are merely illustrative or explanatory of the principles of this application and do not constitute a limitation thereof. Therefore, any modifications, equivalent substitutions, improvements, etc., made without departing from the spirit and scope of this application should be included within the protection scope of this application. Furthermore, the appended claims are intended to cover all variations and modifications falling within the scope and boundaries of the appended claims, or equivalent forms of such scope and boundaries.

Claims

1. A method for end-to-end target tracking of transmission lines, characterized in that, include: Acquire video frames of the power transmission line and the pan-tilt-zoom (PTZ) status information synchronously with the video frames; The video frames and the PTZ status information are input into an end-to-end target tracking network. The end-to-end target tracking network then fuses the image features, PTZ status features, and transmission line topology features to obtain the tracking features. The historical trajectory of the target in the power transmission line scene is retrieved and fused with the time sequence of the tracking features. The historical trajectory association is completed based on the power transmission line topology and the gimbal state constraints. Based on the association results, determine whether the target is occluded and lost contact. If no occlusion or loss of contact occurs, update the historical trajectory. If a short-term loss of contact occurs, cache the original historical trajectory corresponding to the lost target. Based on the update status or cache status of the historical trajectory, output the tracking results of each target in the current video frame.

2. The end-to-end target tracking method for transmission lines according to claim 1, characterized in that, The video frames and the PTZ status information are input into an end-to-end target tracking network. The end-to-end target tracking network then fuses image features, PTZ status features, and transmission line topology features to obtain tracking features, including: The video frame and the gimbal status information are input into the end-to-end target tracking network; wherein, the gimbal status information includes the gimbal horizontal rotation angle, gimbal pitch angle, zoom ratio, and preset position number; The gimbal status information is encoded and converted to generate gimbal status encoding features corresponding to network operations; Multi-scale feature extraction and small target enhancement processing are performed on the video frames input to the end-to-end target tracking network to obtain refined image features; Based on the refined image features, the transmission line topology modeling is completed, and the transmission line topology features are extracted; The refined image features, the gimbal status coding features, and the transmission line topology features are fused to generate fused features for target tracking, which are then used as the tracking features.

3. The end-to-end target tracking method for transmission lines according to claim 2, characterized in that, The gimbal status information is encoded and converted to generate gimbal status encoding features corresponding to network operations, including: The gimbal's horizontal rotation angle and its pitch angle are subjected to sine and cosine transformations respectively to obtain angular characteristics. The preset bit number is embedded and encoded to obtain the preset bit feature; The angle feature, the zoom ratio feature, and the preset position embedding feature are concatenated to generate a gimbal status coding feature with unified dimensions.

4. The end-to-end target tracking method for transmission lines according to claim 2, characterized in that, Multi-scale feature extraction and small target enhancement processing are performed on the video frames input to the end-to-end target tracking network to obtain refined image features, including: The video frames input to the end-to-end target tracking network are fed into the backbone feature extraction network to output multi-scale initial image features. Based on the multi-scale initial image features, high-level semantic features and low-level detail features are obtained by splitting them. Upsampling and convolution fusion are used to fuse the high-level semantic features and the low-level detail features layer by layer to obtain multi-scale fused features. Weight coefficients are generated based on convolution operations and the Sigmoid activation function. The weight coefficients and the multi-scale fusion features are then multiplied element-wise to complete the weighted correction process, resulting in refined image features.

5. The end-to-end target tracking method for transmission lines according to claim 2, characterized in that, Based on the refined image features, transmission line topology modeling is completed, and transmission line topology features are extracted, including: All the refined image features are input into a preset transmission line topology modeling network for calculation; The characteristics of the transmission line topology are extracted from the conductor region, tower structure, insulator string position, line corridor region, and safety distance region through the transmission line topology modeling network, and combined to form the transmission line topology features.

6. The end-to-end target tracking method for transmission lines according to claim 2, characterized in that, The refined image features, the gimbal status coding features, and the transmission line topology features are fused to generate fused features for target tracking, which are used as the tracking features, including: The refined image features are subjected to scale unification and channel integration processing to obtain regularized image fusion features; The dimensionality broadcasting expansion of the gimbal status coding features is performed, and the spatial size is adjusted to be the same as the normalized image fusion features. The normalized image fusion features, PTZ status coding features, and transmission line topology features after dimension matching are spliced ​​together, and the weight coefficients corresponding to the topology features are generated through convolution operation and Sigmoid activation function. The topological features of the transmission line are weighted and corrected using the weighting coefficients, and gating fusion processing is performed on the three types of features to generate fused features for target tracking.

7. The end-to-end target tracking method for transmission lines according to claim 1, characterized in that, Retrieve the historical trajectory of the target in the power transmission line scene and fuse it with the time series of tracking features. Based on the power transmission line topology and the gimbal state constraints, complete the association of the historical trajectory, including: Historical trajectory features of targets in the power transmission line scene are retrieved, and the historical trajectory features and the tracking features of the current frame are fused in a temporal manner through an attention mechanism to obtain temporal fusion features; By combining the historical trajectory features, the gimbal status of historical frames, and the gimbal status of the current frame, the estimated position of the historical target in the current frame is predicted, and candidate association features of the current target and historical trajectory association features are constructed respectively. The trajectory association score is calculated based on the appearance features, location information, topological relationship parameters of the current target candidate and the gimbal status parameters corresponding to the historical trajectory. The matching probability is calculated based on the association score and the historical trajectory association is completed through pairing and allocation.

8. The end-to-end target tracking method for transmission lines according to claim 7, characterized in that, Combining the historical trajectory features, the gimbal status of historical frames, and the gimbal status of the current frame, the estimated position of the historical target in the current frame is predicted. Candidate association features for the current target and historical trajectory association features are constructed respectively, including: The motion prediction function is called to solve the estimated target position corresponding to each historical trajectory based on the historical target position, historical motion state, historical frame gimbal state and current frame gimbal state corresponding to a single historical trajectory. When the gimbal states are similar, the target's estimated position is calculated based on the target's motion state. When the gimbal zooms or rotates, the target's estimated position is corrected by combining the gimbal parameters. Extract the appearance features, location encoding, category encoding, topological features, and current gimbal encoding of the current target and perform feature mapping to obtain the candidate associated features of the current target; Extract the historical trajectory memory features, estimated position codes, historical category codes, historical topology features, and historical gimbal codes corresponding to the single historical trajectory, and perform feature mapping to obtain the historical trajectory association features.

9. The end-to-end target tracking method for transmission lines according to claim 8, characterized in that, Based on the appearance features, location information, topological relationship parameters, and gimbal status parameters corresponding to the current target candidate, a trajectory association score is calculated. Based on the association score, a matching probability is calculated, and the historical trajectory association is completed through pairing and allocation, including: Feature similarity is calculated based on the current target candidate association features and historical trajectory association features. The position intersection-union ratio is obtained by combining the current target candidate position coordinates and the historical trajectory estimated position. The topological difference is calculated based on the current target candidate topological features and the historical trajectory topological features. The gimbal state difference is obtained by relying on the current gimbal encoding and the historical gimbal encoding. The category consistency is obtained by comparing the current target candidate category and the historical trajectory category. The feature similarity, the position intersection-union ratio, the topological difference, the gimbal state difference, and the category consistency are weighted and calculated to obtain the association score between a single candidate and the trajectory. All association scores are normalized to obtain the pairwise association probabilities, and unmatched scores are introduced to represent the target disconnection scenario; Filter matching combinations that exceed the preset association threshold, and perform one-to-one matching and allocation according to the association probability from high to low. When the same target is matched by multiple trajectories, only the pair with the best probability is retained, and the remaining unmatched trajectories are put into the waiting state.

10. The end-to-end target tracking method for transmission lines according to any one of claims 1 to 9, characterized in that, Also includes: Based on the tracking results, a risk score corresponding to the target is calculated, and multiple risk levels are divided by combining the risk score with a preset grading threshold. The tracking results include target category, movement trajectory, spatial relationship between the target and power transmission facilities, and duration of the target's stay in the risk area; Based on the risk level obtained from the classification, when the risk score exceeds the corresponding preset classification threshold, the target corresponding hazard type is matched and an alarm message is output; The alarm information includes target information, camera position information, and trajectory data.

Images