Distribution network tree barrier real-time analysis method and system based on dynamic vision and SLAM

Abstract

The invention discloses a distribution network tree barrier real-time analysis method and system based on dynamic vision and SLAM. The method comprises the following steps: generating a dynamic visual baseline by cooperatively controlling the translation and flight displacement of an unmanned aerial vehicle holder, and constructing a bionic binocular model to simulate a time sequence image into a binocular image pair; generating a depth point cloud through epipolar correction and stereo matching; key targets are recognized and extracted through a semantic segmentation network, and semantic point clouds are generated; establishing a dimensionality reduction motion model by utilizing pan-tilt stability augmentation, and fusing a visual inertial odometer and RTK data by adopting a filtering or optimization algorithm to realize centimeter-level pose estimation; and finally, performing optimization processing on the semantic point cloud, completing three-dimensional reconstruction based on multi-modal fusion, and outputting a risk assessment result through tree line spacing calculation and safety margin analysis. According to the invention, accurate, efficient and automatic routing inspection and risk early warning of the distribution network tree obstacles are realized.

Description

Technical Field

[0001] This invention belongs to the field of power equipment inspection technology, specifically relating to a method and system for real-time analysis of tree obstacles in distribution networks based on dynamic vision and SLAM. Background Technology

[0002] Power distribution network lines often traverse complex natural environments, and the growth of trees along the lines can easily lead to insufficient tree-to-line spacing, causing faults such as discharge and short circuits, seriously threatening the safe and stable operation of the power grid. Therefore, accurate detection and risk assessment of tree obstacles is one of the core tasks of power grid inspection.

[0003] Traditional tree obstruction monitoring methods mainly rely on manual line inspection, lidar scanning, or manned aerial photography. Manual line inspection is inefficient, risky, and highly subjective; lidar technology offers high precision but is expensive, has low data acquisition efficiency, and is difficult to process in real time; traditional aerial photography struggles to simultaneously acquire accurate three-dimensional geometric and semantic information, failing to meet the needs of real-time early warning.

[0004] In recent years, UAV-based visual inspection technology has provided a new approach for tree obstacle inspection. However, existing visual solutions still face many technical bottlenecks: First, traditional monocular vision lacks depth information, while conventional binocular vision has a fixed baseline, and its measurement accuracy drops sharply with increasing distance in wide-area outdoor scenarios; second, factors such as unstable UAV flight attitude, changes in lighting, and missing image textures make feature matching difficult, resulting in insufficient accuracy and robustness in 3D reconstruction; third, existing methods typically perform perception, localization, and reconstruction in separate steps, lacking tight coupling optimization of multi-source sensors (vision, IMU, GNSS), leading to large cumulative errors in pose estimation and making it difficult to meet the requirements for centimeter-level accuracy in tree obstacle measurement; finally, there is a lack of a fully automated process from raw point clouds to quantitative analysis of tree obstacle risk, making it difficult to achieve real-time early warning.

[0005] Therefore, there is an urgent need for a new method for inspecting tree obstacles in power distribution networks that can adapt to complex outdoor environments and achieve high-precision real-time three-dimensional perception and analysis. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for real-time analysis of tree obstacles in power distribution networks based on dynamic vision and SLAM, which can achieve high-precision, high-efficiency, and automated tree obstacle inspection.

[0007] In a first aspect, embodiments of this application provide a real-time analysis method for tree obstacles in power distribution networks based on dynamic vision and SLAM, the method comprising: S1: By controlling the translational motion of the UAV gimbal and the flight displacement of the UAV to generate a dynamic visual baseline, a biomimetic binocular parallax imaging model is constructed, and the time sequence of images is simulated as a pair of binocular visual images. S2: Perform epipolar correction and stereo matching on the binocular vision image pair to generate depth point cloud data, and perform real-time obstacle detection and early warning based on the data; S3: Perform semantic segmentation on the deep point cloud data, identify and extract key targets such as conductors, trees and insulators, and generate point cloud data with semantic labels; S4: Utilize the gimbal stabilization characteristics to establish a dimension-reduced motion model, and fuse visual inertial odometry and real-time differential positioning data through multi-state constrained Kalman filtering or sliding window optimization algorithms to achieve real-time pose estimation with centimeter-level accuracy. S5: Perform voxel filtering and downsampling on semantically labeled point cloud data, use a multimodal fusion algorithm to complete the 3D reconstruction of key targets, and output risk assessment results based on tree line spacing calculation and safety margin analysis.

[0008] Optionally, in one implementation of the first aspect of the present invention, the construction of the biomimetic binocular parallax imaging model in step S1 specifically includes: An improved LightStereo neural network architecture is used for disparity calculation. Multi-scale image features are extracted through a deep separable convolutional network with shared weights, and a channel attention mechanism is introduced to enhance feature discriminativeness. A pixel-wise matching cost matrix in the parallax direction is constructed, and a recursive Hourglass network with a symmetric encoder-decoder structure is used for multi-scale feature fusion. The detail preservation capability is improved by skip connections and recursive optimization modules. The initial disparity map is regressed through a differentiable soft argmax operation, and the disparity confidence is evaluated in conjunction with the uncertainty estimation module. A spatiotemporally consistent ConvGRU recurrent convolutional unit is introduced to establish an iterative optimization mechanism, and a multi-scale pyramid optimization strategy is adopted to optimize the initial disparity in multiple rounds. By actively micro-moving the gimbal and coordinating with the flight displacement of the UAV to generate a dynamic baseline, and combining camera intrinsic parameters and lens distortion models to establish an accurate binocular imaging model, a high-precision conversion from parallax to depth information is achieved.

[0009] Optionally, in one implementation of the first aspect of the present invention, the construction of the biomimetic binocular parallax imaging model further includes photometric error optimization based on the direct SLAM method, specifically including: Parallax optimization is performed using the photometric error model of the Direct Sparse Odometry (DSO). The photometric error calculation formula is as follows: , in, Indicates reference frame The cost of photometric error, and Reference frames and matching frames Exposure compensation parameters, and Reference frames and matching frames Image intensity, and The brightness deviation parameter of the frame. and For the exposure time, and These are the exposure compensation parameters; Indicates in reference frame In the middle, pixels Image intensity at that location, Indicates in the matching frame In the middle, pixels Image intensity at that location; Introducing geometric error constraints, and combining camera pose and 3D point position for joint optimization, the objective function is: , in, Let the overall objective function of the joint optimization problem be... For index variables, Refers to a point in three-dimensional space. Refers to a keyframe in which that point was observed. These are the weighting coefficients. In keyframes In the middle, three-dimensional points The image intensity value at the pixel location projected onto the image plane. To create 3D points In the reference frame, the image intensity value corresponding to this point, Here, γ is the Huber norm, and γ is the switching threshold. For depth map in gradient at, Weights for smoothing terms; L1 norm of depth map gradient; A multi-scale pyramid optimization strategy is employed to minimize photometric errors from coarse to fine at different resolution levels: , in, For the image pyramid Total photometric error on the layer, It is a pyramid hierarchy. This refers to the hierarchical weight coefficient. For hierarchical index variables, For the first Layered images; In the The weighting coefficients calculated on the layer pyramid. For three-dimensional points Projected onto keyframe The Pixel intensity on a layered pyramid image For three-dimensional points In its reference frame the Pixel intensity on a layered pyramid image For three-dimensional points Projected onto keyframe The Pixel intensity on a layered pyramid image; A vision-inertial tightly coupled optimization framework is established using IMU pre-integrated data, with the objective function being: , in, The overall objective function of the vision-inertial tightly coupled optimization framework is... This is the visual error term, namely the error based on photometry or reprojection mentioned above. The error term represents the difference between the predicted state and the optimized state calculated based on IMU pre-integration theory. This is the prior error term; Implement adaptive weight adjustment based on deep learning, predicting the confidence weight of each pixel through a neural network: , in, This is the functional expression for predicting weights using a convolutional neural network. It is a lightweight convolutional neural network whose function is to learn and predict two images. At point Calculate the confidence level of luminosity consistency in the vicinity area and output the weights. , This represents the image patch input into the neural network. The location information of the point can be provided to the network as additional input.

[0010] Optionally, in one implementation of the first aspect of the present invention, the generation of depth point cloud data in S2, and the real-time obstacle detection and early warning based on the data, specifically includes: A stereo matching framework for multimodal sensor fusion is constructed, and motion interference suppression is achieved through tight coupling of IMU, RTK and visual data; Adaptive epipolar correction technology is used to dynamically adjust the correction model based on IMU attitude data and gimbal motion parameters to eliminate image geometric distortion. A semi-global matching algorithm using hybrid cost aggregation is employed, which integrates the Census-gradient hybrid cost function and a multi-scale aggregation strategy, and optimizes the aggregation path based on an attention mechanism. A disparity optimization network combining Kalman filtering and deep learning is used to construct a motion model based on IMU data to predict disparity changes, and temporal filtering is used to improve the spatiotemporal consistency of the disparity map. Establish a dynamic baseline binocular imaging model, calculate the baseline length in real time, and achieve millimeter-level baseline accuracy control; A real-time early warning system based on deep learning is constructed, which uses a lightweight network to achieve millisecond-level obstacle detection and establishes a multi-threshold early warning mechanism and dynamic region of interest management.

[0011] Optionally, in one implementation of the first aspect of the present invention, the point cloud semantic segmentation step in step S3 specifically includes: S3.1. A deep learning-based point cloud semantic segmentation network is used to process deep point cloud data. The network includes: a point cloud feature extraction module, which uses PointNet++ or KPConv network to extract multi-scale point cloud features; an attention mechanism module, which introduces attention weights in the warning area to enhance the feature representation of key areas; and a semantic segmentation head, which outputs the semantic category probability distribution of each point through a multilayer perceptron. S3.2. Focus on identifying and extracting three key targets: conductors: based on spatial continuity and geometric feature extraction, the RANSAC algorithm is used to fit the conductor curve; trees: identified based on point cloud density distribution and vegetation features, and different tree species are distinguished by combining depth information; insulators: based on geometric shape features and spatial position relationship detection, the template matching method is used for accurate positioning. S3.3 Post-processing optimization of segmentation results: Apply Conditional Random Field (CRF) to optimize semantic boundaries and eliminate isolated misclassified points; perform cluster analysis based on spatial connectivity to merge fragmented segmented regions; utilize temporal consistency constraints to fuse multi-frame point cloud segmentation results to improve stability; S3.4 Generate point cloud data with semantic labels, including: the three-dimensional coordinates, color information and semantic category labels of each point; the bounding boxes and geometric parameters of key targets; the confidence scores and spatial relationship descriptions of targets within the warning area; establish a target database to record historical information of the identification results and provide multi-time series comparison and data traceability capabilities.

[0012] Optionally, in one implementation of the first aspect of the present invention, the real-time pose estimation with centimeter-level accuracy in step S4 is a tightly coupled SLAM pose estimation, specifically including: By leveraging the stabilization characteristics of UAV gimbals, a dimensionality-reduced motion model based primarily on three-dimensional translation is established, simplifying the 6-DOF pose estimation problem into a 3-DOF translation estimation problem. Visual inertial odometry (VIO) and real-time differential positioning (RTK) are tightly coupled and deeply fused using multi-state constrained Kalman filtering or sliding window optimization algorithms, where: RTK provides an absolute position reference and suppresses IMU cumulative drift; The vision system provides relative pose estimation and environmental awareness when the RTK signal is lost; Joint optimization is performed using reprojection error and IMU pre-integration error; The optimization objective function of the tightly coupled deep fusion is: : , in: Let be the state vector to be optimized. For IMU pre-integration residuals, For visual reprojection residuals, For the first The pre-integral values ​​of each IMU measurement, Σ is the corresponding covariance matrix, and X is the state vector to be optimized. , This represents the square of the Mahalanobis norm.

[0013] Optionally, in one implementation of the first aspect of the present invention, the tightly coupled SLAM pose estimation step further includes: The state vector X is iteratively optimized using a sliding window optimization algorithm. The sliding window contains multiple keyframes and their corresponding IMU measurements and visual feature points. An edge-out strategy is introduced to handle historical states outside the window, preserving their constraint information to avoid information loss, while controlling computational complexity. A robust kernel function is used to weight the reprojection error and IMU error, suppressing out-of-point interference and improving the system's robustness in dynamic environments; By using online calibration technology to estimate and compensate for the extrinsic parameter deviation between the camera and the IMU in real time, the consistency of multi-sensor data is ensured.

[0014] Optionally, in one implementation of the first aspect of the present invention, the weighting of the reprojection error and the IMU error using a robust kernel function specifically includes: A robust kernel function is used to perform weighted optimization on the visual reprojection residual and the IMU pre-integration residual. The robust kernel function includes the Huber kernel function, the Cauchy kernel function, or the Tukey dual-weight kernel function. The mathematical expression for the Huber kernel function is:

[0015] The mathematical expression for the Cauchy kernel function is: , The mathematical expression for the Tukey double-weighted kernel function is: , in, The residual value, This is the threshold parameter, also known as the inflection point. For scale parameters, For kernel functions; The kernel function type is dynamically selected and the kernel function parameters are adjusted based on the Mahalanobis distance of the residuals. Abnormal residuals that exceed the confidence interval are subjected to nonlinear weight reduction processing, which effectively suppresses the interference of dynamic obstacles, sudden changes in illumination and feature mismatch on pose estimation, and improves the estimation accuracy and robustness of the system in complex environments.

[0016] Optionally, in one implementation of the first aspect of the present invention, step S5 includes: An adaptive spatial voxel mesh generation method is adopted, which dynamically adjusts the voxel size according to the point cloud density distribution. Small voxels are used in dense areas to preserve detailed features, while large voxels are used in sparse areas to improve processing efficiency. The effective voxels are screened through a dual threshold mechanism. First, a primary screening is performed based on the number of points in the voxel, and then a secondary screening is performed based on the point cloud density distribution characteristics to remove outlier voxels and retain real ground voxels. The selected effective voxels are intelligently downsampled, and representative point clouds are selected based on information entropy evaluation. A strategy combining random sampling and feature preservation is adopted to ensure feature integrity while avoiding grid effects. The optimized point cloud data is input into the multimodal fusion tree obstacle analysis module, and the improved ICP algorithm combined with semantic constraints is used for accurate registration, or a deep learning network is used to achieve fine 3D reconstruction of key targets. Conduct a quantitative analysis of tree obstacle risk, calculate the minimum spatial distance between trees and power lines, and comprehensively consider safety margin factors such as thermal expansion and wind deflection to establish a multi-parameter risk assessment model to evaluate the risk level. The final result is a complete analysis report containing a three-dimensional spatial relationship diagram, a risk distribution map, and early warning recommendations, providing decision support for the prevention and control of tree obstacles in the power distribution network.

[0017] Secondly, embodiments of this application provide a real-time tree obstacle analysis system for distribution networks based on dynamic vision and SLAM, applied to the real-time tree obstacle analysis method for distribution networks based on dynamic vision and SLAM as described in the second aspect, the system comprising: The dynamic binocular vision perception module is used to generate a dynamic visual baseline by controlling the translational motion of the UAV gimbal and the flight displacement of the UAV, and to construct a biomimetic binocular parallax imaging model to simulate a time sequence of images as a pair of binocular vision images. The depth information generation and early warning module is used to perform epipolar correction and stereo matching on the binocular vision image pair, generate depth point cloud data, and perform real-time obstacle detection and early warning based on the data; The point cloud semantic segmentation module is used to perform semantic segmentation on the deep point cloud data, identify and extract key targets such as wires, trees and insulators, and generate point cloud data with semantic labels. The tightly coupled SLAM pose estimation module is used to establish a dimensionality-reduced motion model by utilizing the gimbal stabilization characteristics. It integrates visual inertial odometry and real-time differential positioning data through multi-state constrained Kalman filtering or sliding window optimization algorithms to achieve real-time pose estimation with centimeter-level accuracy. The point cloud optimization and analysis module is used to perform voxel filtering and downsampling on point cloud data with semantic labels, use a multimodal fusion algorithm to complete the 3D reconstruction of key targets, and output risk assessment results based on tree line spacing calculation and safety margin analysis.

[0018] Thirdly, embodiments of this application provide an electronic device, including: processor; Memory used to store processor-executable instructions; The processor is configured to implement the real-time analysis method for tree obstacles in the distribution network based on dynamic vision and SLAM as described in the first aspect when executing the instructions.

[0019] Fourthly, embodiments of this application provide a computer-readable storage medium storing a program that instructs a device to execute the real-time analysis method for tree obstacles in a distribution network based on dynamic vision and SLAM as described in the first aspect.

[0020] This invention discloses a method and system for real-time analysis of tree obstacles in power distribution networks based on dynamic vision and SLAM. The method generates a dynamic visual baseline by collaboratively controlling the translation and flight displacement of a UAV gimbal, constructing a biomimetic binocular parallax imaging model to simulate time-series image sequences as binocular image pairs. Depth point cloud data is generated through epipolar correction and stereo matching. Key targets such as conductors, trees, and insulators are identified and extracted using a deep learning point cloud semantic segmentation network, generating point clouds with semantic labels. A dimensionality-reduced motion model is established using the gimbal's stabilization characteristics. Multi-state constrained Kalman filtering or sliding window optimization algorithms are used to tightly couple and fuse visual inertial odometry and RTK positioning data, achieving real-time pose estimation with centimeter-level accuracy. Finally, voxel filtering and downsampling optimization are applied to the semantic point cloud, and 3D reconstruction is completed based on a multimodal fusion algorithm. Risk assessment results are output through tree-line spacing calculation and safety margin analysis. This invention achieves accurate, efficient, and automated inspection and risk warning of tree obstacles in power distribution networks.

[0021] Beneficial effects:

[0022] 1. High-precision depth perception: Through a biomimetic binocular model and dynamic baseline control, long-distance, high-precision 3D information acquisition is achieved, overcoming the defect that the measurement accuracy of traditional fixed-baseline binocular systems decreases with distance.

[0023] 2. Robust pose estimation: By integrating VIO and RTK data using a tightly coupled fusion algorithm, IMU drift is effectively suppressed, and continuous pose output with centimeter-level accuracy can still be provided even in GNSS signal rejection environments.

[0024] 3. Automated semantic understanding: The point cloud segmentation network based on deep learning can automatically and accurately identify and extract key targets such as guide wires and trees, realizing the semanticization of perception results and providing a structured data foundation for risk assessment.

[0025] 4. Real-time risk assessment: Through multimodal 3D reconstruction and tree line spacing analysis, the safety margin can be quantified and risk warning reports can be generated, realizing a closed loop from raw data to risk decision-making, which significantly improves inspection efficiency and automation level. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of a method for real-time analysis of tree obstacles in a power distribution network based on dynamic vision and SLAM, provided in an embodiment of this application.

[0027] Figure 2 The architecture diagram of the real-time tree obstacle analysis system based on dynamic vision and SLAM provided in this application.

[0028] Figure 3 A schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0029] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0030] It should be noted that in the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.

[0031] It should be noted that in the embodiments of this application, the terms "first," "second," etc., are used only for descriptive purposes and should not be construed as indicating or implying relative importance, nor as indicating or implying order. Features defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of the embodiments of this application, terms such as "exemplary" or "for example" are used to indicate examples, illustrations, or descriptions. Any embodiments or designs described as exemplary or for example in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present related concepts in a specific manner.

[0032] Based on the embodiments described in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0033] Example 1

[0034] Figure 1 This is a schematic flowchart of a real-time tree obstacle analysis method for power distribution networks based on dynamic vision and SLAM, provided as an embodiment of this application. Figure 1 As shown, a real-time analysis method for tree obstacles in power distribution networks based on dynamic vision and SLAM includes: S1: By controlling the translational motion of the UAV gimbal and the flight displacement of the UAV to collaboratively generate a dynamic visual baseline, a biomimetic binocular parallax imaging model is constructed, simulating a temporal image sequence as a binocular visual image pair. Through dynamic visual perception, the problem of insufficient accuracy in traditional fixed-baseline binocular systems at long distances is solved. By collaboratively controlling the gimbal translation and UAV flight, a visual baseline is actively generated and precisely controlled, simulating a high-precision "biomimetic binocular" system from a monocular temporal image sequence, providing high-quality image pair input for subsequent 3D reconstruction.

[0035] In this context, a binocular vision image pair refers to a set of two images of the same scene taken at the same time but from different horizontal perspectives. It mimics the way human eyes observe. Traditional binocular systems require two physical cameras, while this invention uses a single camera to "simulate" the effect of a dual-camera system through movement. Specifically, the input is a video sequence (i.e., a sequence of images) captured by a single moving camera. For example, a series of consecutive images captured by a drone while it flies and the gimbal moves.

[0036] Simulation Process: Selecting the "Virtual Camera" Position: Two frames are selected from consecutive video frames. These two frames were captured at different times (t1 and t2). Calculating Relative Pose: Using SLAM, visual odometry, or a combination of IMU / RTK data, the relative position and pose changes of the camera between times t1 and t2 are precisely calculated. Constructing the Image Pair: The camera at time t1 is considered the "left camera," and the camera at time t2 is considered the "right camera." The relative horizontal displacement between them is used as the "dynamic baseline." At this point, although the two images were not captured at the same time, we already know their precise geometric relationship (equivalent to knowing the positions of the two virtual cameras).

[0037] Since the two images were captured at different times, there may be moving objects in the scene (such as swaying leaves or moving cars) or changes in lighting (cloud occlusion), which would violate the "simultaneity" assumption and lead to matching failure. The solution lies in the improved LightStereo network, attention mechanism, and ConvGRU spatiotemporal optimization techniques mentioned below. These advanced deep learning algorithms are designed to be robust enough to tolerate such non-ideal conditions to a certain extent, still enabling effective matching and calculating reliable disparity.

[0038] Specifically, in this embodiment, the construction of the biomimetic binocular parallax imaging model in step S1 specifically includes: An improved LightStereo neural network architecture is employed for disparity calculation. Multi-scale image features are extracted using a deep separable convolutional network with shared weights, and a channel attention mechanism is introduced to enhance feature discriminativeness. This improved LightStereo neural network architecture forms the core engine for high-precision disparity calculation. This step leverages the powerful feature extraction and matching capabilities of deep neural networks to replace traditional algorithms, aiming to directly and efficiently calculate the disparity value of each pixel from image pairs, providing foundational data for subsequent depth restoration. The deep separable convolutional network with shared weights extracts multi-scale image features, significantly improving computational efficiency while maintaining accuracy. Shared weights avoid redundant calculation of features from left and right images; the deep separable convolutional network greatly reduces the number of model parameters and computational cost; and multi-scale feature extraction ensures that it can match large-scale texture regions as well as perceive subtle detail differences, adapting to objects at different distances. The introduction of a channel attention mechanism enhances feature discriminativeness, improving the model's robustness and matching accuracy in complex environments. This mechanism enables the network to autonomously learn and focus on the feature channels that are more important to the matching task, suppressing redundant or unimportant information, thereby enhancing its resistance to interference factors such as changes in lighting, shadows, and repetitive textures.

[0039] A pixel-wise matching cost matrix in the parallax direction is constructed, and a recursive Hourglass network with a symmetric encoder-decoder structure is used for multi-scale feature fusion. The detail preservation capability is improved by skip connections and recursive optimization modules.

[0040] The initial disparity map is regressed using a differentiable soft argmax operation, and the disparity confidence is evaluated in conjunction with an uncertainty estimation module. This differentiable soft argmax operation regresses the initial disparity map, achieving accurate and smooth regression from a discrete cost space to continuous disparity values. Traditional methods simply select minimum values ​​from the cost volume, which introduces quantization errors and is not differentiable. This operation obtains sub-pixel-level disparity estimation through a differentiable approach, improving accuracy and enabling end-to-end training of the entire network. The disparity confidence is evaluated in conjunction with the uncertainty estimation module, providing a reliability score for the disparity estimation of each pixel. This step identifies and quantifies the uncertainty of the matching results (e.g., in weak textures or occluded regions), and the generated confidence map can be used as weights in subsequent processes (such as point cloud filtering and SLAM optimization), reducing the impact of unreliable measurements on the overall system accuracy.

[0041] A spatiotemporally consistent ConvGRU recurrent convolutional unit is introduced to establish an iterative optimization mechanism, employing a multi-scale pyramid optimization strategy to optimize the initial disparity in multiple rounds. The ConvGRU unit, combining convolutional operations and recurrent neural networks (RNNs), can process sequential data while preserving spatial information. When processing video sequences, the disparity calculation of the current frame not only depends on the image information of the current frame but also "memorizes" and fuses disparity information from several previous frames. This approach improves temporal consistency, making the disparity map results of consecutive frames smoother and more stable, avoiding flickering and jitter. Occlusion regions are optimized by using historical information to infer the disparity value of occluded regions in the current frame. Furthermore, noise is suppressed by fusing information from multiple frames, effectively smoothing random noise in a single frame image.

[0042] A multi-scale pyramid optimization strategy is employed to optimize the initial disparity in multiple rounds, progressively refining it from coarse to fine. The algorithm repeatedly calculates disparity on image pyramids at different resolutions (scales). First, an approximate disparity is calculated at the lowest resolution (the coarsest scale), and then this result is used as the initial value to correct and refine it at higher resolutions (more refined scales).

[0043] By coordinating active micro-movement of the gimbal with the flight displacement of the UAV to generate a dynamic baseline, and combining camera intrinsic parameters and lens distortion models, a precise binocular imaging model is established, achieving high-precision conversion from parallax to depth information. The dynamic baseline is generated through active micro-movement of the gimbal and the flight displacement of the UAV, dynamically controlling and precisely measuring the interpupillary distance of the binocular vision system. Traditional binocular cameras have a fixed baseline, while this invention actively and intelligently creates a baseline length best suited to the current observation distance by controlling the translation of the gimbal and the flight of the UAV. Various errors in the imaging process are corrected to ensure the accuracy of the geometric model. Pre-calibrated camera intrinsic parameters (focal length, principal point coordinates, etc.) and lens distortion parameters (radial distortion, tangential distortion coefficients) are used to establish precise projection and back-projection models.

[0044] This system achieves high-precision conversion from parallax to depth information, executing the core triangulation principle to complete the 2D to 3D mapping. The depth Z calculation formula based on binocular vision is: Z = f * B / d (where d is the parallax). Using the precise baseline B, precise focal length f, and optimized parallax d provided in the first two steps, the depth value Z for each pixel is calculated. This generates high-precision depth maps or 3D point cloud data, which serves as the direct data source for subsequent obstacle detection, 3D reconstruction, and risk analysis.

[0045] The construction of the biomimetic binocular disparity imaging model also includes photometric error optimization based on the direct SLAM method, further introducing photometric error optimization based on direct SLAM to improve the quality and geometric consistency of the disparity map. The specific steps include: Parallax optimization is performed using the photometric error model of the Direct Sparse Odometry (DSO). The photometric error calculation formula is as follows: , in, Indicates reference frame The Photometric Error cost measures the overall level of pixel brightness difference between a point observed in the reference frame and its projected point in the matching frame. A smaller value indicates better photometric consistency in that region between the two frames. and Reference frames and matching frames Exposure compensation parameters. Used for more precise correction of exposure-related brightness variations. and Reference frames and matching frames Image intensity, and This is a brightness deviation parameter for frames, used to compensate for overall brightness differences between different frames caused by automatic exposure or changes in lighting. and This is the exposure time, used to eliminate the image brightness scaling effect caused by different exposure times. and These are the exposure compensation parameters; Indicates in reference frame In the middle, pixels Image intensity (brightness value) at that location. These are the two-dimensional coordinates of image pixels. Indicates in the matching frame In the middle, pixels Image intensity (brightness value) at that location. Points in the reference frame Projected onto the matching frame based on camera pose and depth information. The two-dimensional coordinates after that.

[0046] Disparity optimization is performed using a photometric error model based on Direct Sparse Odometry (DSO), bypassing feature extraction and matching, and directly utilizing pixel brightness information for optimization. This step optimizes disparity by minimizing the brightness difference (photometric error) of corresponding pixel blocks between the reference frame and the matching frame. It works better for areas lacking texture features (such as walls and the sky) and can generate a denser disparity map.

[0047] Introducing geometric error constraints, and combining camera pose and 3D point position for joint optimization, the objective function is: , in, The overall objective function for the joint optimization problem consists of a photometric error term and a smoothing term. The goal of the optimization is to find the optimal camera pose and depth map that minimizes the value of this function. For index variables, Refers to a point in three-dimensional space. Refers to a keyframe in which that point was observed. These are weighting coefficients used to assign different levels of importance to each residual term. The impact of outliers can be reduced by using residual size or prediction via a deep learning network. In keyframes In the middle, three-dimensional points The image intensity value at the pixel location projected onto the image plane. To create 3D points The image intensity value corresponding to this point in the reference frame. γ is the Huber Norm, a robust loss function. The L2 norm is used for smaller errors (facilitating convergence), while the L1 norm is used for larger errors (reducing the impact of outliers). γ is the switching threshold. For depth map in gradient at, Weights for smoothing terms; The L1 norm of the depth map gradient is used as a smoothness term. The purpose is to encourage smooth depth maps (i.e., adjacent pixel depth values ​​are similar) and avoid unreasonable drastic changes.

[0048] By introducing geometric error constraints and combining camera pose and 3D point positions for joint optimization, the geometric correctness and consistency of the optimization results are ensured. This step combines photometric consistency with geometric consistency (such as image gradient smoothness and constraints on 3D points and camera pose) to establish a more comprehensive objective function for joint optimization, avoiding geometric distortion that may be caused by optimizing only photometric errors.

[0049] A multi-scale pyramid optimization strategy is employed to minimize photometric errors from coarse to fine at different resolution levels: , in, For the image pyramid Total photometric error on the layer. It is a pyramid hierarchy. This refers to the hierarchical weight coefficient. For hierarchical index variables, For the first Layered images; In the The weighting coefficients calculated on the layered pyramid. For three-dimensional points Projected onto keyframe The Pixel intensity on a layered pyramid image. For three-dimensional points In its reference frame the Pixel intensity on a layered pyramid image For three-dimensional points Projected onto keyframe The Pixel intensity on a layered pyramid image.

[0050] A multi-scale pyramid optimization strategy is employed to minimize photometric errors from coarse to fine at different resolution levels, improving the convergence and accuracy of the optimization process. This step first calculates a coarse disparity and structure on the low-resolution image as initial values ​​for higher-resolution layers, and then gradually refines the results. This strategy expands the convergence range, avoids getting trapped in local minima during the optimization process, and ultimately yields high-precision details.

[0051] A vision-inertial tightly coupled optimization framework is established using IMU pre-integrated data, with the objective function being: , in, The overall objective function of the vision-inertial tightly coupled optimization framework is... This is the visual error term, namely the error based on photometry or reprojection mentioned above. The error term represents the difference between the predicted state and the optimized state calculated based on IMU pre-integration theory. The prior error term, derived from prior constraint information generated by the marginalization operation, is used to maintain consistency with older frames and prevent information loss. A tightly coupled visual-inertial optimization box is established using IMU pre-integration data, providing high-frequency, short-term, and accurate motion priors for visual optimization. IMU pre-integration provides predictions of relative motion between two frames, and this constraint is incorporated into the optimization framework. This makes the system more robust to image blur, rapid motion, or texture loss, and enables the estimation of quantities that are unobservable by purely visual systems, such as scale and gravity direction.

[0052] Implement adaptive weight adjustment based on deep learning, predicting the confidence weight of each pixel through a neural network: , in, This is the functional expression for predicting weights using a convolutional neural network. It is a lightweight convolutional neural network whose function is to learn and predict two images. At point Calculate the confidence level of luminosity consistency in the vicinity area and output the weights. , This represents image patches input into the neural network, typically represented by points. and its projection point The local image region centered on the image The location information of the point can be provided to the network as additional input.

[0053] This system implements deep learning-based adaptive weight adjustment, predicting the confidence weight of each pixel through a neural network to intelligently reduce the influence of unreliable pixels (outsides) in the optimization process. This step utilizes a lightweight CNN network to automatically determine the photometric consistency confidence of each pixel (e.g., whether the pixel is occluded, located in a region of sudden illumination change, or on a dynamic object), and assigns lower weights to pixels with low confidence, thereby significantly improving the system's robustness in complex real-world scenarios.

[0054] The aforementioned steps collectively constitute a more advanced and robust optimization system. It no longer relies on traditional feature matching but directly optimizes pixel brightness, incorporating multiple constraints such as geometry and inertia, and finally intelligently removes outliers through deep learning. This combined approach aims to produce more accurate, denser, and more reliable disparity maps and depth information, laying a solid data foundation for subsequent semantic segmentation and risk assessment.

[0055] S2: Epipolar correction and stereo matching are performed on the binocular vision image pairs to generate depth point cloud data, and real-time obstacle detection and early warning are performed based on this data. Three-dimensional information is recovered from two-dimensional images through 3D information generation and preliminary early warning. Epipolar correction and stereo matching are performed on the generated image pairs to calculate disparity and convert it into depth information, generating depth point cloud data for the scene. Based on this, preliminary real-time obstacle detection and early warning based on geometric distance are achieved.

[0056] Specifically, in this embodiment, the generation of depth point cloud data in step S2, and the real-time obstacle detection and early warning based on this data, specifically includes: A multimodal sensor fusion-based stereo matching framework is constructed, achieving motion interference suppression through tight coupling of IMU, RTK, and visual data. This enhances the robustness of stereo matching under intense UAV movement. By fusing high-frequency attitude data from the IMU and absolute position information from the RTK, image motion blur and geometric distortion caused by UAV jitter, wind speed, etc., are compensated for and eliminated, providing stable and reliable image input for stereo matching.

[0057] This invention employs adaptive epipolar correction technology, dynamically adjusting the correction model based on IMU attitude data and gimbal motion parameters to eliminate image geometric distortion. It also addresses the failure of traditional epipolar correction due to continuous platform movement. While traditional correction is designed for fixed binocular systems, this invention dynamically calculates and applies the epipolar correction model based on real-time perceived platform attitude (IMU) and gimbal status. This ensures that matching searches can be performed on the correct horizontal line even when the platform is moving, guaranteeing matching efficiency and accuracy.

[0058] A semi-global matching algorithm employing hybrid cost aggregation fuses a Census-gradient hybrid cost function and a multi-scale aggregation strategy, optimizing the aggregation path based on an attention mechanism. This achieves an optimal balance between accuracy and efficiency, resulting in robust pixel matching. The hybrid cost function combines illumination-insensitive Census transform with texture-sensitive gradient information, enhancing the discriminative power of the matching cost and adapting to complex outdoor lighting variations. Semi-global matching and multi-scale aggregation, through multi-path optimization and cross-scale information fusion, maintains high computational efficiency while obtaining a disparity map with less noise and stronger consistency.

[0059] Census transform: A non-parametric local transform. It generates a bit string descriptor by comparing the relative (rather than absolute) brightness of a pixel with its neighboring pixels. This makes it highly invariant to linear and gamma changes in illumination, making it ideal for complex lighting conditions in outdoor inspections. Gradient information: Captures edge and texture features of an image. It is highly sensitive to the geometry of objects, providing rich detail information, but its effectiveness is limited in uniform regions and it is sensitive to noise. Hybrid strategy: Weighted fusion of the two (e.g., Cost_hybrid=α*Cost_census+β*Cost_gradient). This achieves complementary advantages: the illumination robustness of Census provides stability for matching, while the detail sensitivity of gradients ensures matching accuracy, especially in object edge regions.

[0060] Multi-scale aggregation strategy: Achieve reliable matching from coarse to fine, eliminating ambiguity at low resolution and restoring details at high resolution, ensuring the accuracy and completeness of the matching. Cost aggregation is performed on the low-resolution (coarse-scale) image. At this point, image texture is simplified, and a wide range of unique patterns are easier to match, reliably estimating the approximate object outline and large disparity, but losing details. The coarse-scale matching results are upsampled and passed to the high-resolution (fine-scale) image as a guide for its disparity search range. At the fine scale, the algorithm only needs to perform fine corrections within a very small disparity range. This strategy greatly reduces the computational cost and avoids the ambiguity and errors caused by searching the entire disparity range at the fine scale.

[0061] Attention-based path optimization: This approach enables adaptive cost aggregation, intelligently guiding the aggregation process and allowing the algorithm to focus on information-rich regions and paths, thus improving the utilization of effective information. Traditional semi-global matching (SGM) aggregates costs along a fixed number of paths (e.g., 8 or 16), treating all pixels "equally." This is ineffective in areas with weak or repetitive textures, as the paths in these areas may be ambiguous. The introduction of the attention mechanism allows the algorithm to dynamically calculate the "importance weight" of each pixel on every aggregation path. For regions with rich texture and high confidence (such as wire edges or insulator textures), their aggregation paths are assigned high weights, ensuring the full propagation of reliable information. For regions with weak texture, occlusion, or potential noise, their path weights are suppressed to prevent unreliable information from contaminating the global cost calculation. This is equivalent to installing "intelligent navigation" in the aggregation process, enabling it to autonomously select the optimal information propagation path, significantly improving the matching success rate in challenging regions.

[0062] The semi-global matching algorithm of hybrid cost aggregation: Global optimization of the disparity map integrates the local matching costs generated in the preceding steps into a globally consistent, smooth, and accurate disparity map through an efficient global optimization method. The core idea of ​​"semi-global" is to approximate the two-dimensional global energy function minimization problem through one-dimensional optimization along multiple paths (such as directions 8 and 16), achieving a perfect balance between computational complexity and result accuracy. "Hybrid cost aggregation" refers to integrating the aforementioned hybrid cost function, multi-scale information, and attention weights into the SGM aggregation framework. It does not simply add the results of each path but performs a weighted fusion, where the weights are determined by the attention mechanism.

[0063] A disparity optimization network combining Kalman filtering and deep learning is employed. It constructs a motion model using IMU data to predict disparity changes and uses temporal filtering to improve the spatiotemporal consistency of the disparity map. This further optimizes the disparity map quality, ensuring smooth and stable output. Kalman filtering: Utilizes IMU data to predict scene motion trends, performing temporal filtering and smoothing on the disparity map to reduce jitter. Deep learning network: Repairs occlusion and mismatched regions. Spatiotemporal consistency: Combining the advantages of both methods, the output provides continuous and consistent disparity results in both time and space, avoiding flickering and jumps in frame-by-frame results.

[0064] A dynamic baseline binocular imaging model was established to calculate the baseline length in real time, achieving millimeter-level baseline accuracy control; this provides core parameters for high-precision triangulation. Precise control and real-time measurement of the dynamic baseline length generated by the gimbal translation and UAV flight are crucial for converting parallax into depth information. Millimeter-level control accuracy directly ensures centimeter-level or even millimeter-level accuracy in the final depth / distance measurement.

[0065] Construct a real-time early warning system based on deep learning, employing a lightweight network to achieve millisecond-level obstacle detection, and establishing a multi-threshold early warning mechanism and dynamic region of interest (ROI) management. The ultimate business value is realized through real-time risk early warning. Lightweight network: Ensures real-time algorithm execution (millisecond-level response) on the onboard computing platform. Multi-threshold and dynamic ROI: Set different risk thresholds according to safety regulations and dynamically focus on the region of interest (ROI) near the conductor, achieving a closed loop from "perception" to "early warning," and promptly outputting risk alerts of different levels.

[0066] S3: Perform semantic segmentation on the deep point cloud data, identify and extract key targets such as conductors, trees, and insulators, and generate point cloud data with semantic labels. By understanding the semantics of the environment, semantic information is assigned to the original point cloud. A deep learning network is used to segment the deep point cloud, automatically identifying, classifying, and extracting key targets such as "conductors," "trees," and "insulators," transforming the disordered point cloud data into a structured semantic map with clear category labels, laying the foundation for subsequent accurate analysis and risk assessment.

[0067] Specifically, in this embodiment, the point cloud semantic segmentation step S3 specifically includes: S3.1. A deep learning-based point cloud semantic segmentation network is used to process deep point cloud data. The network includes: a point cloud feature extraction module, which uses PointNet++ or KPConv network to extract multi-scale point cloud features; an attention mechanism module, which introduces attention weights in the warning area to enhance the feature representation of key areas; and a semantic segmentation head, which outputs the semantic category probability distribution of each point through a multilayer perceptron.

[0068] The network extracts and learns deep features from the point cloud to perform preliminary point-by-point classification. The point cloud feature extraction module, as the network backbone, efficiently captures local and global geometric features of the point cloud. PointNet++ excels at handling point clouds of varying densities, while KPConv (Kernel-Point Convolution) provides more refined convolutional feature extraction; both provide rich feature representations for subsequent classification. The attention mechanism module enhances the feature response of key regions and suppresses background interference. By calculating attention weights, the network pays more attention to point clouds in warning regions (ROIs) such as near conductors and tree canopy edges, improving the classification accuracy of key targets. The semantic segmentation head performs the final classification decision. It receives deep features incorporating attention weights and outputs a probability distribution for each point belonging to various semantic categories (such as conductors, trees, insulators, background, etc.) using an MLP.

[0069] S3.2. Focus on identifying and extracting three key types of targets: conductors: based on spatial continuity and geometric feature extraction, the RANSAC algorithm is used to fit the conductor curve; trees: based on point cloud density distribution and vegetation features, different tree species are distinguished by combining depth information; insulators: based on geometric shape features and spatial position relationship detection, the template matching method is used for accurate positioning; for the geometric characteristics of specific targets, special algorithms are used for accurate extraction and model fitting.

[0070] Conductors: Leveraging their spatial continuity and linear geometric features, algorithms such as RANSAC (Random Sample Consensus) are used to fit a 3D spatial curve model of the conductor, accurately reconstructing its direction and position, rather than a jumble of scattered points. Trees: Utilizing the sparse spatial distribution and complex structure (high density variation, irregular shapes) of vegetation point clouds, combined with depth information, the morphological characteristics of different tree species can be further distinguished (e.g., trees vs. shrubs). Insulators: Utilizing their regular, repeating geometric shapes (e.g., disc shape) and fixed spatial relationship with conductors, shape-based template matching methods are used to achieve precise localization at the sub-point cloud level.

[0071] S3.3 Post-processing optimization of segmentation results: Apply Conditional Random Field (CRF) to optimize semantic boundaries and eliminate isolated misclassified points; perform cluster analysis based on spatial connectivity to merge fragmented segmented regions; utilize temporal consistency constraints to fuse multi-frame point cloud segmentation results to improve stability; optimize preliminary segmentation results to eliminate noise and ambiguity, and improve the smoothness, consistency and stability of the results.

[0072] Conditional Random Field (CRF) optimization: Considers spatial proximity and feature similarity between points, smooths semantic boundaries, and corrects isolated misclassified points that are inconsistent with the categories of surrounding points. Spatial connectivity-based clustering analysis: Merges scattered points belonging to the same instance (such as the point cloud of a tree) into a complete object, eliminating fragmented segmentation results. Temporal consistency constraints: Fuses segmentation results from multiple consecutive frames, utilizing temporal information to stabilize the classification result of the current frame, reducing jitter and uncertainty in single-frame segmentation.

[0073] S3.4 Generate point cloud data with semantic labels, including: 3D coordinates, color information, and semantic category labels for each point; bounding boxes and geometric parameters of key targets; confidence scores and spatial relationship descriptions of targets within the warning area; establish a target database to record historical information of identification results, providing multi-time series comparison and data traceability capabilities. Output a structured, information-rich semantic map, providing direct input for risk assessment and data analysis.

[0074] It not only includes the basic attributes of each point (coordinates, color, category), but also outputs high-level semantic information: bounding boxes and geometric parameters of key targets, providing a basis for subsequent safe distance calculations (e.g., calculating the minimum distance from the tree point cloud's outer envelope to the conductor point cloud). Confidence scores and spatial relationships: indicating the reliability of each identification result (e.g., "90% certain this is a conductor") and describing the relative positions between targets (e.g., "the insulator is connected below the conductor"). Target database and historical records: enabling data traceability and time-series comparison, allowing analysis of tree growth rates, assessment of logging effectiveness, and providing long-term data support for operational decisions.

[0075] S4: Utilizing the gimbal stabilization capabilities, a reduced-dimensional motion model is established. By fusing visual inertial odometry (VIO) and real-time differential positioning (RTK) data through multi-state constrained Kalman filtering or sliding window optimization algorithms, centimeter-level accuracy in real-time pose estimation is achieved. High-precision self-localization provides a centimeter-level accurate real-time pose (position and attitude) reference for the entire system. The motion model is simplified by leveraging the gimbal stabilization capabilities, and a tightly coupled algorithm is used to deeply fuse visual (VIO), inertial (IMU), and satellite (RTK) data. This complementary approach effectively suppresses errors and drift from any single sensor, ensuring that the perception results are mapped to a precise global coordinate system.

[0076] Specifically, in this embodiment, the real-time pose estimation with centimeter-level accuracy in step S4 is tightly coupled SLAM pose estimation, which specifically includes: By leveraging the stabilization capabilities of a UAV gimbal, a dimensionality-reduced motion model primarily based on 3D translation is established, simplifying the 6-DOF pose estimation problem into a 3-DOF translation estimation problem. This simplifies the complexity of state estimation, improves computational efficiency, and enhances numerical stability. The gimbal's stabilization function actively compensates for the UAV's rotational jitter, ensuring that the image sequence captured by the camera primarily undergoes translational motion. This allows the 6-DOF pose estimation problem (3 translations + 3 rotations) to be approximately simplified into a problem primarily based on 3D translation estimation, significantly reducing the complexity of the optimization problem.

[0077] Visual inertial odometry (VIO) and real-time differential positioning (RTK) are tightly coupled and deeply fused using multi-state constrained Kalman filtering or sliding window optimization algorithms, where: RTK provides an absolute position reference and suppresses IMU cumulative drift; The vision system provides relative pose estimation and environmental awareness when the RTK signal is lost; Joint optimization is performed using reprojection error and IMU pre-integration error; The optimization objective function of the tightly coupled deep fusion is: : , in: Let the state vector be the one to be optimized. For IMU pre-integration residuals, For visual reprojection residuals, For the first The pre-integral values ​​of each IMU measurement, Σ is the corresponding covariance matrix, and X is the state vector to be optimized. , Typically represented as the square of the Mahalanobis norm, also known as the covariance-weighted square norm. The optimization objective function of the tightly coupled deep fusion achieves optimal estimation by minimizing the Mahalanobis norm (i.e., the covariance-weighted square norm). The Mahalanobis norm is a "smart" weighting method. It assigns lower weights to measurements with high uncertainty (high noise) and higher weights to measurements with low uncertainty (low noise). The system automatically places greater trust in more reliable sensor data. For example, absolute position constraints have high weights when the RTK signal is good; visual constraints have high weights when the image is clear and stable. This uncertainty-based automatic weighting is key to achieving centimeter-level accuracy because it optimally fuses all available information. The optimization objective function of the tightly coupled deep fusion is used to find the system state (pose, velocity, bias, etc.) that is most likely to explain all sensor observation data (visual and IMU) within a unified mathematical framework. The objective function applies the constraints of these three (visual geometric constraints, IMU kinematic constraints, and RTK absolute position constraints) to the state variable X to be optimized.

[0078] The core of this step is to achieve deep fusion of Visual Inertial Odometry (VIO) and Real-Time Differential Positioning (RTK) through tight coupling using multi-state constrained Kalman filtering or sliding window optimization algorithms. This allows for the complementary advantages of multiple sensor sources, providing absolutely accurate and relatively smooth pose output. The role of RTK is to provide an absolute position reference (centimeter-level accuracy) and effectively suppress cumulative drift caused by IMU accelerometer bias, confining pose error within a finite range and preventing its unlimited growth. The role of the vision system is to provide high-frequency relative pose changes and rich environmental feature information. When the RTK signal is briefly lost (e.g., under bridges, near trees), VIO can seamlessly take over, ensuring the continuity of pose estimation. The meaning of "tight coupling": Unlike simply selecting or switching data, tight coupling uses visual observations, IMU measurements, and RTK observations within a unified optimization framework to estimate the same set of state variables, thereby achieving the deepest level of information fusion.

[0079] Joint optimization using reprojection error and IMU pre-integration error establishes the mathematical foundation for fusion optimization, providing dual constraints for state estimation. Reprojection error: A visual constraint. It measures the difference between the position of 3D map points projected back to the image plane and the actual observed feature point positions, ensuring the estimated pose is consistent with the observed environmental features. IMU pre-integration error: A kinematic constraint. It measures the difference between the pose change predicted by IMU measurements and the optimized pose change, providing smooth and high-frequency motion priors. Joint optimization: By minimizing the weighted sum of these two errors, the obtained pose estimate conforms to both environmental geometry and the kinematics model, resulting in more accurate and reliable results.

[0080] The optimization objective function for tightly coupled fusion defines the specific implementation of "joint optimization" in mathematical form. This objective function explicitly states that it aims to simultaneously minimize the weighted sum of squares of the IMU pre-integration residual and the visual reprojection residual. The weights are determined by their respective covariance matrices; sensor data with high uncertainty naturally has lower weights, reflecting the algorithm's intelligent consideration of sensor noise characteristics.

[0081] The steps described above construct the "brain" and "navigation system" of the entire system. Through an innovative dimensionality reduction model and deep, tightly coupled fusion, it seamlessly integrates the continuous relative pose of the VIO, the high-frequency motion data of the IMU, and the precise absolute position of the RTK, ultimately outputting centimeter-level accuracy, absolute reference, and continuously reliable pose information. This high-precision spatiotemporal reference is the fundamental prerequisite for the success of all subsequent processing (such as point cloud stitching, 3D reconstruction, and distance measurement).

[0082] The tightly coupled SLAM pose estimation step also includes: The state vector X is iteratively optimized using a sliding window optimization algorithm. The sliding window contains multiple keyframes and their corresponding IMU measurements and visual feature points. This algorithm achieves an optimal balance between computational complexity and optimization accuracy. The sliding window mechanism only retains the most recent series of keyframes and their states (pose, velocity, bias, map points) in the optimization window. When a new frame arrives, the old frame is removed from the window. This prevents the optimization problem from growing indefinitely over time, keeping the computational load constant and meeting real-time requirements. Furthermore, compared to filtering methods that only consider the current frame, optimizing multiple frames achieves higher accuracy.

[0083] An edge-mining strategy is introduced to handle historical states outside the window, preserving their constraint information to avoid information loss while controlling computational complexity. This strategy retains valuable constraint information contained in older frames removed from the window, preventing "information loss." When a frame is removed from the sliding window, it is not simply discarded. The edge-mining strategy transforms the frame's state into a priori probability constraint on the remaining states within the window, which is then incorporated into subsequent optimization problems. This approach controls the scale of the optimization problem while preserving the constraint of historical observations on the current state, effectively suppressing accumulated errors over long periods—a key technique for ensuring algorithm consistency.

[0084] A robust kernel is used to weight reprojection errors and IMU errors, suppressing outlier interference and improving the system's robustness in dynamic environments. This robust kernel further enhances the system's robustness in dynamic and complex environments, mitigating the destructive impact of anomalous observations. Standard optimization defaults to using least squares (L2 norm), which is highly sensitive to outliers. Robust kernels (such as Huber and Cauchy) reduce the weight of larger residual terms to weaken or even eliminate the impact of anomalous observations caused by dynamic objects (such as moving vehicles or swaying branches), erroneous data associations, and transient occlusion, resulting in more stable and reliable state estimation results.

[0085] Online calibration technology estimates and compensates for extrinsic parameter deviations between the camera and IMU in real time, ensuring the consistency of multi-sensor data. This ensures the accuracy of the geometric basis for multi-sensor data fusion. The rigid transformation (extrinsic parameter) between the camera and IMU is the geometric basis of tightly coupled fusion. This parameter can change slightly due to factors such as vibration and temperature differences (extrinsic parameter deviation). Online calibration technology incorporates the extrinsic parameter as a state variable into the optimization process for real-time estimation, automatically compensating for this deviation. This ensures that the visual projection and IMU integration are geometrically strictly consistent, avoiding a sharp decline in fusion performance due to inaccurate extrinsic parameters.

[0086] The specific steps of using robust kernel functions to weight reprojection errors and IMU errors include: using robust kernel functions to perform weighted optimization on the visual reprojection residuals and IMU pre-integration residuals, transforming the standard least squares optimization problem into a more robust optimization problem. This step is the declaration of the core idea, clearly stating the application of robust kernel processing to the two key residuals (visual geometric error and IMU kinematic error), laying the foundation for subsequent specific kernel function selection. The robust kernel functions include the Huber kernel function, Cauchy kernel function, or Tukey dual-weight kernel function; providing multiple mathematical tools to handle different outlier scenarios. Different kernel functions have different suppression characteristics: Huber kernel: maintains quadratic growth for small errors (efficient optimization), and converts large errors to linear growth (outlier suppression), a common choice balancing efficiency and robustness. Cauchy kernel: provides smoother and more gradual suppression of large errors, less prone to gradient abrupt changes, and more stable numerical behavior. Tukey kernel: completely discards errors exceeding the threshold (weights reduced to zero), the most aggressive strategy for handling significant outliers.

[0087] The mathematical expression for the Huber kernel function is: , The mathematical expression for the Cauchy kernel function is: , The mathematical expression for the Tukey double-weighted kernel function is: , in, The residual value, This is the threshold parameter, also known as the inflection point. For scale parameters, For kernel functions; The kernel function type is dynamically selected and the kernel function parameters are adjusted based on the Mahalanobis distance of the residuals. Abnormal residuals that exceed the confidence interval are subjected to nonlinear weight reduction processing, which effectively suppresses the interference of dynamic obstacles, sudden changes in illumination and feature mismatch on pose estimation, and improves the estimation accuracy and robustness of the system in complex environments.

[0088] Adaptive intelligent weighting enables the system to automatically adjust its suppression strategy based on the severity of errors. Mahalanobis distance: a distance metric that considers data correlation and measures the degree of anomaly of a residual relative to its expected distribution (described by the covariance matrix) more accurately than Euclidean distance. Dynamic selection and adjustment: the system no longer uses a fixed kernel function or parameter. For example, the Huber kernel might be used for mild anomalies (small Mahalanobis distance), while the Tukey kernel might be used for severe anomalies (maximum Mahalanobis distance); simultaneously, the threshold δ or scaling parameter c can be dynamically adjusted according to the residual distribution. This gives the system the ability to perceive error reliability and respond intelligently.

[0089] Nonlinear weighting is applied to outlier residuals exceeding the confidence interval. This core action directly weakens the impact of anomalous observations on the final solution. "Nonlinear" means the weights don't decrease uniformly, but rather drop sharply according to the kernel function curve, ensuring that a small number of bad data points don't dominate the optimization result. This effectively suppresses interference from dynamic obstacles, sudden changes in illumination, and feature mismatches on pose estimation. This section points out typical sources of anomalies in robust kernel function processing: Dynamic obstacles: Feature points on moving objects such as vehicles and pedestrians can generate incorrect constraints that violate the static world assumption. Sudden changes in illumination: These cause drastic changes in visual appearance, leading to feature matching errors or photometric errors. Feature mismatches: The most common source of error in visual odometry. By suppressing these interferences, the overall goal of improving the estimation accuracy and robustness of the system in complex environments is ultimately achieved.

[0090] The above steps introduce an "intelligent arbitrator" into the SLAM system. This arbitrator (robust kernel mechanism) no longer treats all sensor data equally, but dynamically reviews the reliability of each observation (residual) based on Mahalanobis distance as "evidence." For outliers that "lie," it applies corresponding "penalties" (deweighting) according to the "severity of the lie" (outlier degree), thereby ensuring that the final decision (pose estimation) is not misled and is more accurate and reliable. This is a key technological guarantee for achieving high-precision operation of the system in real-world complex scenarios.

[0091] S5: Voxel filtering and downsampling are performed on semantically labeled point cloud data. A multimodal fusion algorithm is used to complete the 3D reconstruction of key targets, and risk assessment results are output based on tree line spacing calculation and safety margin analysis. Through refined reconstruction and risk decision-making, the final risk assessment result is generated. The semantic point cloud is optimized (denoised and compressed), and a refined 3D reconstruction of key targets is completed based on a multimodal algorithm. Finally, tree line spacing is calculated based on the accurate 3D model, and quantitative risk analysis is performed after comprehensively considering safety margins. A decision report that can guide operation and maintenance is output, achieving a closed loop from perception to analysis.

[0092] Specifically, in this embodiment, S5 includes: An adaptive spatial voxel grid partitioning method is adopted, dynamically adjusting the voxel size according to the point cloud density distribution. Small voxels are used in dense areas to preserve detailed features, while large voxels are used in sparse areas to improve processing efficiency. A dual threshold mechanism is used to screen effective voxels: first, a primary screening is performed based on the number of points within a voxel, and then a secondary screening is performed based on the point cloud density distribution characteristics, eliminating outliers and retaining true ground objects. The screened effective voxels undergo intelligent downsampling, selecting representative point clouds based on information entropy assessment. A strategy combining random sampling and feature preservation is adopted to ensure feature integrity while avoiding raster effects. The optimized point cloud data is input into a multimodal fusion tree barrier analysis module, where an improved ICP algorithm combined with semantic constraints is used for accurate registration, or a deep learning network is used to achieve fine 3D reconstruction of key targets. Tree barrier risk quantification analysis is performed, calculating the minimum spatial distance between trees and power lines, and comprehensively considering safety margin factors such as thermal expansion and wind deflection to establish a multi-parameter risk assessment model to determine the risk level. Finally, a complete analysis report including a 3D spatial relationship diagram, a risk distribution map, and early warning suggestions is generated, providing decision support for power distribution network tree barrier prevention and control.

[0093] These five steps form a complete "perception-localization-understanding-decision" technology loop: S1-S2 are responsible for perceiving the three-dimensional world, S4 provides a high-precision positioning benchmark for the whole process, S3 is responsible for understanding the semantics of the environment, and finally S5 is used for risk assessment and decision-making.

[0094] Example 2

[0095] like Figure 2 As shown, this application provides an architecture diagram of a real-time tree obstacle analysis system based on dynamic vision and SLAM for distribution networks. It is applied to the real-time tree obstacle analysis system based on dynamic vision and SLAM for distribution networks as described in Embodiment 1. It includes a dynamic binocular vision perception module 11, a depth information generation and early warning module 12, a point cloud semantic segmentation module 13, a tightly coupled SLAM pose estimation module 14, and a point cloud optimization and analysis module 15.

[0096] The dynamic binocular vision perception module 11 is used to generate a dynamic visual baseline by controlling the translational motion of the UAV gimbal and the flight displacement of the UAV, constructing a biomimetic binocular parallax imaging model, and simulating a time-series image sequence as a binocular vision image pair.

[0097] The depth information generation and early warning module 12 is used to perform epipolar correction and stereo matching on the binocular vision image pair, generate depth point cloud data, and perform real-time obstacle detection and early warning based on the data.

[0098] The point cloud semantic segmentation module 13 is used to perform semantic segmentation on the deep point cloud data, identify and extract key targets such as wires, trees and insulators, and generate point cloud data with semantic labels.

[0099] The tightly coupled SLAM pose estimation module 14 is used to establish a dimensionality-reduced motion model by utilizing the gimbal stabilization characteristics. It integrates visual inertial odometry and real-time differential positioning data through multi-state constrained Kalman filtering or sliding window optimization algorithms to achieve real-time pose estimation with centimeter-level accuracy.

[0100] The point cloud optimization and analysis module 15 is used to perform voxel filtering and downsampling on point cloud data with semantic labels, use a multimodal fusion algorithm to complete the 3D reconstruction of key targets, and output risk assessment results based on tree line spacing calculation and safety margin analysis.

[0101] Figure 3 This is an electronic device provided in one embodiment of this application. For example... Figure 3 As shown, the electronic device includes at least the following components: processor 101 and memory 100, communication interface 103, and bus 102.

[0102] In this embodiment of the application, memory 100 is used to store executable instructions of processor 101, which, when configured to execute instructions, implements the method as described in the first aspect.

[0103] In embodiments of this application, a computer-readable storage medium includes instructions that instruct a device to perform the method as described in the first aspect. For example, the instructions instruct the device to perform... Figure 1 The method is shown in the process steps.

[0104] In one embodiment of this application, the program operating in the electronic device may be a program that controls a central processing unit (CPU) or similar device to achieve the functions of the above-described embodiments of the present invention (a program that enables the computer to function). Information processed by these systems is then temporarily stored in random access memory (RAM) during processing, and subsequently stored in various ROMs such as read-only memory (FlashROM) and hard disk drives (HDDs), and read, corrected, and written by the CPU as needed.

[0105] It should be noted that a portion of the electronic device described above can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.

[0106] It should be noted that the computer mentioned here refers to a computer built into an electronic device, employing hardware including an operating system and peripheral devices. Furthermore, computer-readable recording media refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage systems such as hard drives built into the computer.

[0107] Furthermore, computer-readable recording media can include: media that dynamically stores programs for short periods of time, such as communication lines used when transmitting programs via networks like the Internet or communication lines like telephone lines; and media that store programs for fixed periods of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining them with programs already recorded in the computer.

[0108] Furthermore, the electronic device in the above embodiments can also be implemented as an assembly (system group) composed of multiple systems. Each system constituting the system group can possess some or all of the functions or functional blocks of the electronic device in the above embodiments. As a system group, it is sufficient to have all the functions or functional blocks of the electronic device.

[0109] Those skilled in the art should recognize that the above embodiments are only used to illustrate this application and are not intended to limit this application. Any appropriate changes and variations made to the above embodiments within the essential spirit and scope of this application fall within the scope of protection claimed in this application.

Claims

1. A real-time analysis method for tree obstacles in power distribution networks based on dynamic vision and SLAM, characterized in that, The method includes: S1: By controlling the translational motion of the UAV gimbal and the flight displacement of the UAV to generate a dynamic visual baseline, a biomimetic binocular parallax imaging model is constructed, and the time sequence of images is simulated as a pair of binocular visual images. S2: Perform epipolar correction and stereo matching on the binocular vision image pair to generate depth point cloud data, and perform real-time obstacle detection and early warning based on the data; S3: Perform semantic segmentation on the deep point cloud data, identify and extract key targets such as conductors, trees and insulators, and generate point cloud data with semantic labels; S4: Utilize the gimbal stabilization characteristics to establish a dimension-reduced motion model, and fuse visual inertial odometry and real-time differential positioning data through multi-state constrained Kalman filtering or sliding window optimization algorithms to achieve real-time pose estimation with centimeter-level accuracy. S5: Perform voxel filtering and downsampling on semantically labeled point cloud data, use a multimodal fusion algorithm to complete the 3D reconstruction of key targets, and output risk assessment results based on tree line spacing calculation and safety margin analysis.

2. The method for real-time analysis of tree obstacles in power distribution networks based on dynamic vision and SLAM according to claim 1, characterized in that, The construction of the biomimetic binocular parallax imaging model in step S1 specifically includes: An improved LightStereo neural network architecture is used for disparity calculation. Multi-scale image features are extracted through a deep separable convolutional network with shared weights, and a channel attention mechanism is introduced to enhance feature discriminativeness. A pixel-wise matching cost matrix in the parallax direction is constructed, and a recursive Hourglass network with a symmetric encoder-decoder structure is used for multi-scale feature fusion. The detail preservation capability is improved by skip connections and recursive optimization modules. The initial disparity map is regressed through a differentiable soft argmax operation, and the disparity confidence is evaluated in conjunction with the uncertainty estimation module. A spatiotemporally consistent ConvGRU recurrent convolutional unit is introduced to establish an iterative optimization mechanism, and a multi-scale pyramid optimization strategy is adopted to optimize the initial disparity in multiple rounds. By actively micro-moving the gimbal and coordinating with the flight displacement of the UAV to generate a dynamic baseline, and combining camera intrinsic parameters and lens distortion models to establish an accurate binocular imaging model, a high-precision conversion from parallax to depth information is achieved.

3. The method for real-time analysis of tree obstacles in power distribution networks based on dynamic vision and SLAM according to claim 2, characterized in that, The construction of the biomimetic binocular parallax imaging model also includes photometric error optimization based on the direct SLAM method, specifically including: Parallax optimization is performed using the photometric error model of the Direct Sparse Odometry (DSO). The photometric error calculation formula is as follows: ， in, Indicates reference frame The cost of photometric error, and Reference frames and matching frames Exposure compensation parameters, and Reference frames and matching frames Image intensity, and The brightness deviation parameter for the frame. and Exposure time; Indicates in reference frame In the middle, pixels Image intensity at that location, Indicates in the matching frame In the middle, pixels Image intensity at that location; Introducing geometric error constraints, and combining camera pose and 3D point positions for joint optimization, the objective function is: ， in, Let the overall objective function of the joint optimization problem be... For index variables, Refers to a point in three-dimensional space. Refers to a keyframe in which that point was observed. These are the weighting coefficients. In keyframes In the middle, three-dimensional points The image intensity value at the pixel location projected onto the image plane. To create 3D points In the reference frame, the image intensity value corresponding to this point, Here, γ is the Huber norm, and γ is the switching threshold. For depth map in gradient at, Weights for smoothing terms; L1 norm of depth map gradient; A multi-scale pyramid optimization strategy is employed to minimize photometric errors from coarse to fine at different resolution levels: ， in, For the image pyramid Total photometric error on the layer, It is a pyramid hierarchy. This refers to the hierarchical weight coefficient. For hierarchical index variables, For the first Layered images; In the The weighting coefficients calculated on the layer pyramid. For three-dimensional points Projected onto keyframe The Pixel intensity on a layered pyramid image For three-dimensional points In its reference frame the Pixel intensity on a layered pyramid image For three-dimensional points Projected onto keyframe The Pixel intensity on a layered pyramid image; A vision-inertial tightly coupled optimization framework is established using IMU pre-integrated data, with the objective function being: ， in, The overall objective function of the vision-inertial tightly coupled optimization framework is... For visual error terms, The error term represents the difference between the predicted state and the optimized state calculated based on IMU pre-integration theory. This is the prior error term; Implement adaptive weight adjustment based on deep learning, predicting the confidence weight of each pixel through a neural network: ， in, This is the functional expression for predicting weights using a convolutional neural network. It is a lightweight convolutional neural network whose function is to learn and predict two images. At point Calculate the confidence level of luminosity consistency in the vicinity area and output the weights. , This represents the image patch input into the neural network. The location information of the point can be provided to the network as additional input.

4. The method for real-time analysis of tree obstacles in power distribution networks based on dynamic vision and SLAM according to claim 1, characterized in that, The generation of depth point cloud data in S2, and the real-time obstacle detection and early warning based on this data, specifically include: A stereo matching framework for multimodal sensor fusion is constructed, and motion interference suppression is achieved through tight coupling of IMU, RTK and visual data; Adaptive epipolar correction technology is used to dynamically adjust the correction model based on IMU attitude data and gimbal motion parameters to eliminate image geometric distortion. A semi-global matching algorithm using hybrid cost aggregation is employed, which integrates the Census-gradient hybrid cost function and a multi-scale aggregation strategy, and optimizes the aggregation path based on an attention mechanism. A disparity optimization network combining Kalman filtering and deep learning is used to construct a motion model based on IMU data to predict disparity changes, and temporal filtering is used to improve the spatiotemporal consistency of the disparity map. Establish a dynamic baseline binocular imaging model, calculate the baseline length in real time, and achieve millimeter-level baseline accuracy control; A real-time early warning system based on deep learning is constructed, which uses a lightweight network to achieve millisecond-level obstacle detection and establishes a multi-threshold early warning mechanism and dynamic region of interest management.

5. The method for real-time analysis of tree obstacles in power distribution networks based on dynamic vision and SLAM according to claim 1, characterized in that, The point cloud semantic segmentation step in step S3 specifically includes: S3.

1. A deep learning-based point cloud semantic segmentation network is used to process deep point cloud data. The network includes: a point cloud feature extraction module, which uses PointNet++ or KPConv network to extract multi-scale point cloud features; an attention mechanism module, which introduces attention weights in the warning area to enhance the feature representation of key areas; and a semantic segmentation head, which outputs the semantic category probability distribution of each point through a multilayer perceptron. S3.

2. Focus on identifying and extracting three key targets: conductors: based on spatial continuity and geometric feature extraction, the RANSAC algorithm is used to fit the conductor curve; trees: identified based on point cloud density distribution and vegetation features, and different tree species are distinguished by combining depth information; insulators: based on geometric shape features and spatial position relationship detection, the template matching method is used for accurate positioning. S3.3 Post-processing optimization of segmentation results: Apply Conditional Random Field (CRF) to optimize semantic boundaries and eliminate isolated misclassified points; perform cluster analysis based on spatial connectivity to merge fragmented segmented regions; utilize temporal consistency constraints to fuse multi-frame point cloud segmentation results to improve stability; S3.4 Generate point cloud data with semantic labels, including: the three-dimensional coordinates, color information and semantic category labels of each point; the bounding boxes and geometric parameters of key targets; the confidence scores and spatial relationship descriptions of targets within the warning area; establish a target database to record historical information of the identification results and provide multi-time series comparison and data traceability capabilities.

6. The method for real-time analysis of tree obstacles in power distribution networks based on dynamic vision and SLAM according to claim 1, characterized in that, The real-time pose estimation with centimeter-level accuracy achieved in step S4 is a tightly coupled SLAM pose estimation, specifically including: By leveraging the stabilization characteristics of UAV gimbals, a dimensionality-reduced motion model based primarily on three-dimensional translation is established, simplifying the 6-DOF pose estimation problem into a 3-DOF translation estimation problem. Visual inertial odometry (VIO) and real-time differential positioning (RTK) are tightly coupled and deeply fused using multi-state constrained Kalman filtering or sliding window optimization algorithms, where: RTK provides an absolute position reference and suppresses IMU cumulative drift; The vision system provides relative pose estimation and environmental awareness when the RTK signal is lost; Joint optimization is performed using reprojection error and IMU pre-integration error; The optimization objective function of the tightly coupled deep fusion is: : ， in: Let be the state vector to be optimized. For IMU pre-integration residuals, For visual reprojection residuals, For the first The pre-integral values ​​of each IMU measurement, Σ is the corresponding covariance matrix, and X is the state vector to be optimized. , This represents the square of the Mahalanobis norm.

7. The method for real-time analysis of tree obstacles in power distribution networks based on dynamic vision and SLAM according to claim 6, characterized in that, The tightly coupled SLAM pose estimation step also includes: The state vector X is iteratively optimized using a sliding window optimization algorithm. The sliding window contains multiple keyframes and their corresponding IMU measurements and visual feature points. An edge-out strategy is introduced to handle historical states outside the window, preserving their constraint information to avoid information loss, while controlling computational complexity. A robust kernel function is used to weight the reprojection error and IMU error, suppressing out-of-point interference and improving the system's robustness in dynamic environments; By using online calibration technology to estimate and compensate for the extrinsic parameter deviation between the camera and the IMU in real time, the consistency of multi-sensor data is ensured.

8. The method for real-time analysis of tree obstacles in power distribution networks based on dynamic vision and SLAM according to claim 7, characterized in that, The specific steps of using a robust kernel function to weight the reprojection error and the IMU error include: A robust kernel function is used to perform weighted optimization on the visual reprojection residual and the IMU pre-integration residual. The robust kernel function includes the Huber kernel function, the Cauchy kernel function, or the Tukey dual-weight kernel function. The mathematical expression for the Huber kernel function is: ; The mathematical expression for the Cauchy kernel function is: , The mathematical expression for the Tukey double-weighted kernel function is: , in, The residual value, This is the threshold parameter, also known as the inflection point. For scale parameters, For kernel functions; The kernel function type is dynamically selected and the kernel function parameters are adjusted based on the Mahalanobis distance of the residuals. Abnormal residuals that exceed the confidence interval are subjected to nonlinear weight reduction processing, which effectively suppresses the interference of dynamic obstacles, sudden changes in illumination and feature mismatch on pose estimation, and improves the estimation accuracy and robustness of the system in complex environments.

9. A method for real-time analysis of tree obstacles in power distribution networks based on dynamic vision and SLAM according to claim 1, characterized in that, S5 specifically includes: An adaptive spatial voxel mesh generation method is adopted, which dynamically adjusts the voxel size according to the point cloud density distribution. Small voxels are used in dense areas to preserve detailed features, while large voxels are used in sparse areas to improve processing efficiency. The effective voxels are screened through a dual threshold mechanism. First, a primary screening is performed based on the number of points in the voxel, and then a secondary screening is performed based on the point cloud density distribution characteristics to remove outlier voxels and retain real ground voxels. The selected effective voxels are intelligently downsampled, and representative point clouds are selected based on information entropy evaluation. A strategy combining random sampling and feature preservation is adopted to ensure feature integrity while avoiding grid effects. The optimized point cloud data is input into the multimodal fusion tree obstacle analysis module, and the improved ICP algorithm combined with semantic constraints is used for accurate registration, or a deep learning network is used to achieve fine 3D reconstruction of key targets. Conduct a quantitative analysis of tree obstacle risk, calculate the minimum spatial distance between trees and power lines, and comprehensively consider safety margin factors such as thermal expansion and wind deflection to establish a multi-parameter risk assessment model to evaluate the risk level. The final result is a complete analysis report containing a three-dimensional spatial relationship diagram, a risk distribution map, and early warning recommendations, providing decision support for the prevention and control of tree obstacles in the power distribution network.

10. A real-time tree obstacle analysis system for power distribution networks based on dynamic vision and SLAM, applied to the real-time tree obstacle analysis method for power distribution networks based on dynamic vision and SLAM as described in any one of claims 1 to 9, characterized in that, The system includes: The dynamic binocular vision perception module is used to generate a dynamic visual baseline by controlling the translational motion of the UAV gimbal and the flight displacement of the UAV, and to construct a biomimetic binocular parallax imaging model to simulate a time sequence of images as a pair of binocular vision images. The depth information generation and early warning module is used to perform epipolar correction and stereo matching on the binocular vision image pair, generate depth point cloud data, and perform real-time obstacle detection and early warning based on the data; The point cloud semantic segmentation module is used to perform semantic segmentation on the deep point cloud data, identify and extract key targets such as wires, trees and insulators, and generate point cloud data with semantic labels. The tightly coupled SLAM pose estimation module is used to establish a dimensionality-reduced motion model by utilizing the gimbal stabilization characteristics. It integrates visual inertial odometry and real-time differential positioning data through multi-state constrained Kalman filtering or sliding window optimization algorithms to achieve real-time pose estimation with centimeter-level accuracy. The point cloud optimization and analysis module is used to perform voxel filtering and downsampling on point cloud data with semantic labels, use a multimodal fusion algorithm to complete the 3D reconstruction of key targets, and output risk assessment results based on tree line spacing calculation and safety margin analysis.

Images