Real-time analysis method and system for network tree barrier based on dynamic vision and slam

Abstract

The application discloses a kind of based on dynamic vision and SLAM's network tree barrier real-time analysis method and system.Translational and flight displacement of unmanned aerial vehicle holder are cooperatively controlled to generate dynamic vision baseline, construct bionic binocular model, and time series image is simulated as binocular image pair;After epipolar correction and stereo matching, depth point cloud is generated;Through semantic segmentation network identification extraction key target, generate semantic point cloud;Dimension reduction motion model is established by holder stabilization, and centimeter-level pose estimation is realized by filtering or optimization algorithm fusion vision inertial odometry and RTK data;Finally, the semantic point cloud is optimized, and three-dimensional reconstruction is completed based on multi-modal fusion, and the risk assessment result is output by tree line spacing calculation and safety margin analysis.The application realizes the precise, efficient, automatic inspection and risk early warning of network tree barrier.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power equipment inspection, and particularly relates to a distribution network tree barrier real-time analysis method and system based on dynamic vision and SLAM. BACKGROUND

[0002] The distribution network line corridor often passes through complex natural environments, and the growth of trees along the line can easily lead to insufficient tree-line distance, causing discharge, short circuit and other faults, which seriously threaten the safe and stable operation of the power grid. Therefore, accurate perception and risk assessment of tree barriers are one of the core tasks of power grid inspection.

[0003] Traditional tree barrier monitoring methods mainly rely on manual line inspection, laser radar scanning or manned aerial photography. Manual line inspection is low in efficiency, high in risk and strong in subjectivity; laser radar technology is high in precision but expensive in equipment, low in data collection efficiency and difficult to process in real time; traditional aerial photography is difficult to synchronously obtain accurate three-dimensional geometric information and semantic information, and cannot meet the demand of real-time early warning.

[0004] In recent years, visual detection technology based on unmanned aerial vehicles has provided a new idea for tree barrier inspection. However, the existing visual scheme still faces many technical bottlenecks: first, traditional monocular vision lacks depth information, and conventional binocular vision has a fixed baseline, and the measurement accuracy decreases sharply with the increase of distance in outdoor wide-area scenes; second, unstable flight attitude of unmanned aerial vehicles, changes in light, missing image texture and other factors make feature matching difficult, and the three-dimensional reconstruction accuracy and robustness are insufficient; third, existing methods usually perform perception, positioning and reconstruction in steps, lack tight coupling optimization of multi-source sensors (vision, IMU, GNSS), and cause large cumulative error of pose estimation, which makes it difficult to meet the centimeter-level accuracy requirement of tree barrier measurement; finally, there is a lack of full-automatic process from raw point cloud to tree barrier risk quantization analysis, making it difficult to realize real-time early warning.

[0005] Therefore, there is an urgent need for a new method for distribution network tree barrier inspection that can adapt to complex outdoor environments and realize high-precision real-time three-dimensional perception and analysis. SUMMARY

[0006] The purpose of the present application is to overcome the shortcomings of the prior art and provide a distribution network tree barrier real-time analysis method and system based on dynamic vision and SLAM, which can realize high-precision, high-efficiency and automatic tree barrier inspection.

[0007] In a first aspect, the present application provides a distribution network tree barrier real-time analysis method based on dynamic vision and SLAM, which comprises:

[0008] S1: generate a dynamic vision baseline by controlling the pan motion of the unmanned aerial vehicle gimbal and the flight displacement of the unmanned aerial vehicle, construct a bionic binocular parallax imaging model, and simulate the time sequence image sequence as a binocular vision image pair;

[0009] S2: performing epipolar rectification and stereo matching on the binocular vision image pair, generating depth point cloud data, and performing real-time obstacle detection and early warning based on the data;

[0010] S3: performing semantic segmentation on the depth point cloud data, identifying and extracting key targets such as conductors, trees and insulators, and generating point cloud data with semantic labels;

[0011] S4: establishing a dimension reduction motion model using the gimbal stabilization feature, fusing visual inertial odometry and real-time differential positioning data through multi-state constraint Kalman filtering or sliding window optimization algorithm, and realizing real-time pose estimation with centimeter-level accuracy;

[0012] S5: performing voxel filtering and down-sampling processing on the point cloud data with semantic labels, completing three-dimensional reconstruction of key targets using a multi-modal fusion algorithm, and outputting risk assessment results based on tree line spacing calculation and safety margin analysis.

[0013] Optionally, in an implementation form of the first aspect of the present application, the construction of the bionic binocular parallax imaging model in step S1 specifically includes:

[0014] An improved LightStereo neural network architecture is used for parallax calculation, multi-scale image features are extracted through a depth separable convolution network with shared weights, and a channel attention mechanism is introduced to enhance feature discriminability;

[0015] A pixel-by-pixel matching cost matrix of parallax direction is constructed, a recursive Hourglass network with a symmetric encoder-decoder structure is used for multi-scale feature fusion, and a skip connection and a recursive optimization module are used to improve the detail preservation ability;

[0016] An initial parallax map is regressed through a differentiable soft argmax operation, and a parallax confidence is evaluated in combination with an uncertainty estimation module;

[0017] A ConvGRU recurrent convolution unit with spatiotemporal consistency is introduced to establish an iterative optimization mechanism, and a multi-scale pyramid optimization strategy is used to optimize the initial parallax in multiple rounds;

[0018] A dynamic baseline is generated by coordinating the gimbal active micro-shift and the flight displacement of the unmanned aerial vehicle, an accurate binocular imaging model is established in combination with the camera intrinsic parameters and the lens distortion model, and high-precision conversion from parallax to depth information is realized.

[0019] Optionally, in an implementation form of the first aspect of the present application, the construction of the bionic binocular parallax imaging model further includes photometric error optimization based on a direct SLAM method, specifically including:

[0020] Parallax optimization is performed using the photometric error model of the Direct Sparse Odometry (DSO). The photometric error calculation formula is as follows:

[0021] ,

[0022] 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.

[0023] Introducing geometric error constraints, and combining camera pose and 3D point positions for joint optimization, the objective function is:

[0024] ,

[0025] 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;

[0026] A multi-scale pyramid optimization strategy is adopted to minimize photometric error from coarse to fine at different resolution levels:

[0027] ,

[0028] where, is the total photometric error on the image pyramid at the th level, is the pyramid level, is the level weight coefficient, is the level index variable, is the image at the th level; is the weight coefficient calculated on the th level pyramid, is the pixel intensity of the three-dimensional point projected onto the th level pyramid image of the key frame , is the pixel intensity of the three-dimensional point on the th level pyramid image of its reference frame, is the pixel intensity of the three-dimensional point projected onto the th level pyramid image of the key frame ;

[0029] A visual-inertial tightly coupled optimization framework is established combined with IMU pre-integration data, and the objective function is:

[0030] ,

[0031] where, is the total objective function of the visual-inertial tightly coupled optimization framework, is the visual error term, i.e. the above photometric or re-projection based error, is the IMU error term, the difference between the predicted state calculated based on IMU pre-integration theory and the optimized state, is the prior error term;

[0032] An adaptive weight adjustment based on deep learning is realized, which predicts the confidence weight of each pixel through a neural network. A lightweight convolutional neural network is used to learn and predict the photometric consistency confidence of the two images in the vicinity of the point , and output the weight. An adaptive weight adjustment based on deep learning is realized, which predicts the confidence weight of each pixel through a neural network:

[0033] ,

[0034] where, is a function expression of predicting weights by a convolutional neural network, is a lightweight convolutional neural network, which functions to learn and predict two images is the position information of the point is the confidence of the photometric consistency of the nearby region, and the weights are output , represents an image block input into the neural network, is the position information of the point, which can be provided as an additional input to the network.

[0035] Optionally, in an implementation form of the first aspect of the present application, the generating of the depth point cloud data in S2, and the real-time obstacle detection and early warning based on the data specifically comprises:

[0036] A stereo matching framework of multi-modal sensor fusion is constructed, and motion interference suppression is realized through tight coupling of IMU, RTK and visual data;

[0037] An adaptive epipolar rectification technology is adopted, a rectification model is dynamically adjusted according to IMU attitude data and gimbal motion parameters, and image geometric distortion is eliminated;

[0038] A semi-global matching algorithm using hybrid cost aggregation is used, a Census-gradient hybrid cost function and a multi-scale aggregation strategy are fused, and an aggregation path is optimized based on an attention mechanism;

[0039] A disparity optimization network combining Kalman filtering and deep learning is adopted, a motion model is constructed by using IMU data to predict the change of disparity, and temporal filtering is used to improve the spatiotemporal consistency of the disparity map;

[0040] A dynamic baseline binocular imaging model is established, the baseline length is calculated in real time, and millimeter-level baseline precision control is realized;

[0041] A real-time early warning system based on deep learning is constructed, a lightweight network is used to realize millisecond-level obstacle detection, a multi-threshold early warning mechanism and dynamic region of interest management are established.

[0042] Optionally, in an implementation form of the first aspect of the present application, the point cloud semantic segmentation step in step S3 specifically comprises:

[0043] S3.1, a point cloud semantic segmentation network based on deep learning is used to process the depth point cloud data, the network comprises: a point cloud feature extraction module, a PointNet++ or KPConv network is used to extract multi-scale point cloud features; an attention mechanism module, an attention weight is introduced in the early warning region to enhance the key region feature representation; a semantic segmentation head, a multi-layer perceptron is used to output the semantic class probability distribution of each point;

[0044] S3.2, Focus on identifying and extracting three types of key targets: conductor: based on spatial continuity and geometric feature extraction, using RANSAC algorithm to fit conductor curve; trees: according to point cloud density distribution and vegetation characteristics, combined with depth information to distinguish different tree species; insulator: based on geometric shape features and spatial position relationship detection, using template matching method for accurate positioning;

[0045] S3.3, Post-processing optimization of segmentation results: applying conditional random field CRF to optimize semantic boundary, eliminating isolated misclassified points; based on spatial connectivity, clustering analysis is carried out to merge fragmented segmentation areas; using time sequence consistency constraint, the stability is improved by fusing multi-frame point cloud segmentation results;

[0046] S3.4, Generate point cloud data with semantic labels, including: three-dimensional coordinates, color information and semantic category labels of each point; bounding box and geometric parameters of key targets; confidence score and spatial relationship description of targets in warning area; establish target database to record historical information of identification results, provide multi-time sequence comparison and data tracing ability.

[0047] Optionally, in an implementation form of the first aspect of the application, the step S4 of realizing real-time pose estimation with centimeter-level accuracy is tight coupling SLAM pose estimation, specifically including:

[0048] A dimension reduction motion model based on three-dimensional translation is established by using the gimbal stabilization characteristics of the unmanned aerial vehicle, and the six-degree-of-freedom pose estimation problem is simplified to three-degree-of-freedom translation estimation;

[0049] The visual inertial odometer VIO and the real-time differential positioning RTK are tightly coupled and deeply fused through multi-state constraint Kalman filtering or sliding window optimization algorithm, wherein:

[0050] The RTK provides absolute position reference and suppresses the IMU cumulative drift;

[0051] The visual system provides relative pose estimation and environment perception when the RTK signal is lost;

[0052] Joint optimization is performed through re-projection error and IMU pre-integration error;

[0053] The optimization objective function of the tight coupling deep fusion is :

[0054] ,

[0055] Wherein: is the state vector to be optimized, is the IMU pre-integration residual error, is the visual re-projection residual error, is the first pre-integrated quantity of the IMU measurement, and Σ is a corresponding covariance matrix, X is a state vector to be optimized, , denotes the square of the Mahalanobis norm.

[0056] Optionally, in an implementation form of the first aspect of the application, the tightly coupled SLAM pose estimation step further comprises:

[0057] iteratively optimizing the state vector X by a sliding window optimization algorithm, the sliding window containing a plurality of key frames and corresponding IMU measurements and visual feature points;

[0058] introducing an edge strategy to handle historical states outside the window, retaining their constraint information to avoid information loss while controlling the computational complexity;

[0059] using a robust kernel function to weight the re-projection error and the IMU error, suppressing outlier interference and improving the robustness of the system in a dynamic environment;

[0060] using an online calibration technique to estimate and compensate for the extrinsic parameter deviation between the camera and the IMU in real time, ensuring the consistency of multi-sensor data.

[0061] Optionally, in an implementation form of the first aspect of the application, the using a robust kernel function to weight the re-projection error and the IMU error specifically comprises:

[0062] using a robust kernel function to weight and optimize the visual re-projection residual and the IMU pre-integrated residual, the robust kernel function including a Huber kernel function, a Cauchy kernel function or a Tukey double-weight kernel function;

[0063] the mathematical expression of the Huber kernel function is:

[0064] the mathematical expression of the Cauchy kernel function is: ,

[0065] the mathematical expression of the Tukey double-weight kernel function is:

[0066] ,

[0067] wherein, is a residual value, is a threshold parameter, also known as an inflection point, is a scale parameter, is a kernel function;

[0068] The type of kernel function is dynamically selected according to Mahalanobis distance of the residual, and the kernel function parameters are adjusted, and the abnormal residual exceeding the confidence interval is processed by nonlinear weight reduction, so that the interference of factors such as dynamic obstacles, sudden changes of light and feature mismatch on pose estimation is effectively inhibited, and the estimation accuracy and robustness of the system in a complex environment are improved.

[0069] Optionally, in an implementation form of the first aspect of the application, the S5 comprises:

[0070] An adaptive spatial voxel grid division method is adopted, and the voxel size is dynamically adjusted according to the point cloud density distribution, small size voxels are used in dense areas to retain detailed features, and large size voxels are used in sparse areas to improve processing efficiency;

[0071] The effective voxels are screened through a double threshold mechanism, first, the primary screening is performed based on the number of point clouds in the voxel, and then the secondary screening is performed in combination with the point cloud density distribution characteristics, so as to eliminate outlier voxels and retain real object voxels;

[0072] The effective voxels after screening are processed by intelligent downsampling, representative point clouds are selected based on information entropy evaluation, and a strategy combining random sampling and feature preservation is adopted, so as to ensure the integrity of the features while avoiding the grid effect;

[0073] The optimized point cloud data is input into a multi-modal fusion tree barrier analysis module, an improved ICP algorithm is used in combination with semantic constraints for accurate registration, or a deep learning network is used to realize fine three-dimensional reconstruction of key targets;

[0074] The tree barrier risk is quantitatively analyzed, the minimum spatial distance between the trees and the conductor is calculated, the safety margin factors such as thermal expansion and wind deflection are comprehensively considered, a multi-parameter risk evaluation model is established to evaluate the risk level;

[0075] Finally, a complete analysis report containing a three-dimensional spatial relationship diagram, a risk distribution diagram and a warning suggestion is generated, thereby providing decision support for tree barrier prevention and control of distribution networks.

[0076] In a second aspect, the embodiments of the present application provide a distribution network tree barrier real-time analysis system based on dynamic vision and SLAM, which is applied to the distribution network tree barrier real-time analysis method based on dynamic vision and SLAM as described in the second aspect, and the system comprises:

[0077] A dynamic binocular vision perception module is configured to generate a dynamic vision baseline by controlling the pan motion of a UAV gimbal and the flight displacement of a UAV, construct a bionic binocular parallax imaging model, and simulate a time sequence image sequence as a binocular vision image pair.

[0078] A depth information generation and early warning module is configured 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.

[0079] a point cloud semantic segmentation module for performing semantic segmentation on the depth point cloud data, identifying and extracting key targets of conductors, trees and insulators, and generating point cloud data with semantic labels;

[0080] a tightly coupled SLAM pose estimation module for establishing a reduced dimension motion model using gimbal stabilization characteristics, fusing visual inertial odometry and real-time differential positioning data through multi-state constraint Kalman filtering or sliding window optimization algorithm, and realizing real-time pose estimation with centimeter-level precision;

[0081] a point cloud optimization and analysis module for performing voxel filtering and down-sampling processing on the point cloud data with semantic labels, completing three-dimensional reconstruction of key targets using a multi-modal fusion algorithm, and outputting risk assessment results based on tree line distance calculation and safety margin analysis.

[0082] In a third aspect, an electronic device is provided, comprising:

[0083] a processor;

[0084] a memory for storing processor-executable instructions;

[0085] wherein the processor is configured to implement the dynamic vision and SLAM-based real-time analysis method for distribution network trees and obstacles when executing the instructions.

[0086] In a fourth aspect, a computer-readable storage medium is provided, which stores a program instructing a device to execute the dynamic vision and SLAM-based real-time analysis method for distribution network trees and obstacles.

[0087] A dynamic vision and SLAM-based real-time analysis method and system for distribution network trees and obstacles are disclosed. The method generates a dynamic vision baseline by cooperatively controlling the pan and flight displacement of a drone gimbal, constructs a bionic binocular parallax imaging model, and simulates a time sequence image sequence as a binocular image pair. Depth point cloud data is generated through epipolar rectification and stereo matching. Key targets such as conductors, trees and insulators are identified and extracted through a deep learning point cloud semantic segmentation network, and point clouds with semantic labels are generated. A reduced dimension motion model is established using gimbal stabilization characteristics, and multi-state constraint Kalman filtering or sliding window optimization algorithm is used to tightly couple and fuse visual inertial odometry and RTK positioning data, realizing real-time pose estimation with centimeter-level precision. Finally, voxel filtering and down-sampling optimization are performed on the semantic point cloud, three-dimensional reconstruction is completed based on a multi-modal fusion algorithm, and risk assessment results are output through tree line distance calculation and safety margin analysis. The present application realizes precise, efficient and automated inspection and risk warning for distribution network trees and obstacles.

[0088] Beneficial effects:

[0089] 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.

[0090] 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.

[0091] 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.

[0092] 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

[0093] 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.

[0094] Figure 2 The architecture diagram of the real-time tree obstacle analysis system based on dynamic vision and SLAM provided in this application.

[0095] Figure 3 A schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0096] 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.

[0097] 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.

[0098] It should be noted that in the embodiments of the present application, the first, second and the like are used only for the purpose of distinguishing description and cannot be understood as indicating or implying relative importance, nor can it be understood as indicating or implying sequence. The features limited by the first and the second can explicitly or implicitly include one or more of the features. In the description of the embodiments of the present application, the words such as example or for example are used to represent as an example, illustration or description. Any embodiment or design scheme described as an example or for example in the embodiments of the present application should not be interpreted as more preferred or more advantageous than other embodiments or design schemes. Rather, the use of example or the like is intended to present the relevant concept in a specific manner.

[0099] Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor are within the scope of protection of the present application.

[0100] Embodiment one

[0101] Figure 1 The flowchart of the real-time analysis method of the network tree barrier based on dynamic vision and SLAM provided by an embodiment of the present application is shown in the figure. Figure 1 As shown in the figure, a real-time analysis method of a network tree barrier based on dynamic vision and SLAM includes:

[0102] S1: generate a dynamic vision baseline by controlling the pan motion of the unmanned aerial vehicle gimbal and the flight displacement of the unmanned aerial vehicle, construct a bionic binocular parallax imaging model, and simulate the time sequence image sequence as a binocular vision image pair. Through dynamic vision perception, the problem of insufficient accuracy of traditional fixed baseline binocular system in long distance measurement is solved. By cooperatively controlling the pan motion of the gimbal and the flight of the unmanned aerial vehicle, the vision baseline is generated and accurately controlled, the monocular time sequence image sequence is simulated as a high-precision “bionic binocular” system, and high-quality image pair input is provided for subsequent three-dimensional reconstruction.

[0103] Among them, the binocular vision image pair refers to a group of two images of the same scene taken at the same time from different horizontal angles. It simulates the observation method of human eyes. The traditional binocular system needs two physical cameras, while the present application uses a single camera to “simulate” the effect of double cameras through motion. Specifically: the input is a video sequence (i.e. time sequence image sequence) taken by a single moving camera. For example, a series of continuous images taken while the unmanned aerial vehicle is flying and the gimbal is panning.

[0104] Simulation process: select "virtual camera" position: from the continuous video frames, select two frames of images. These two frames of images are taken at different times (t1 and t2). Calculate the relative pose: through SLAM, visual odometry or combined with IMU / RTK data, accurately calculate the relative position and attitude change of the camera between t1 and t2. Construct the imaging pair: the camera at t1 is regarded as the "left camera", and the camera at t2 is regarded as the "right camera". The relative horizontal displacement between them is regarded as the "dynamic baseline". At this time, although the two images are not taken at the same time, we already know the accurate geometric relationship between them (equivalent to knowing the positions of the two virtual cameras).

[0105] Since the two frames of images are taken at different times, there may be object movement (such as swaying leaves, moving cars) or light changes (cloud cover) in the scene, which will destroy the "simultaneity" assumption and cause matching failure. The solution of this scheme: this is where the improved LightStereo network, attention mechanism, ConvGRU spatio-temporal optimization and other technologies mentioned below come into play. These advanced deep learning algorithms are designed to be robust enough to tolerate this non-ideal situation to some extent, and still be able to perform effective matching and calculate reliable disparities.

[0106] Specifically, in this embodiment, the construction of the bionic binocular stereo disparity imaging model in step S1 specifically includes:

[0107] An improved LightStereo neural network architecture is used for disparity calculation, multi-scale image features are extracted through a depth separable convolution network with shared weights, and a channel attention mechanism is introduced to enhance feature discriminability; an improved LightStereo neural network architecture is used for disparity calculation, and a core engine for high-precision disparity calculation is constructed. This step uses the powerful feature extraction and matching capability of deep neural networks to replace traditional algorithms, aiming to directly and efficiently calculate the disparity value of each pixel from the image pair, providing basic data for subsequent depth recovery. Multi-scale image features are extracted through a depth separable convolution network with shared weights, which greatly improves the calculation efficiency while ensuring accuracy. Shared weights avoid repeated calculation of left and right image features; depth separable convolution greatly reduces the number of model parameters and computational complexity; extracting multi-scale features ensures that both large-scale texture regions and subtle detail differences can be matched, adapting to objects at different distances. The introduction of a channel attention mechanism enhances feature discriminability, improves the robustness and matching accuracy of the model in complex environments. This mechanism allows the network to autonomously learn and focus on features that are more important to the matching task, suppressing redundant or unimportant information, thereby enhancing the ability to resist interference factors such as light changes, shadows, and repetitive textures.

[0108] The pixel-by-pixel matching cost matrix of the parallax direction is constructed, a recursive Hourglass network with a symmetrical encoder-decoder structure is used for multi-scale feature fusion, and a skip connection and a recursive optimization module are used to improve the detail retention capability.

[0109] The initial disparity map is regressed through a differentiable soft argmax operation, and the disparity confidence is evaluated by combining an uncertainty estimation module; the initial disparity map is regressed through a differentiable soft argmax operation, precise and smooth regression from discrete cost space to continuous disparity value is realized. The traditional method simply selects the minimum value from the cost volume, which introduces quantization error and is not differentiable. This operation obtains sub-pixel level disparity estimation in a differentiable manner, improves accuracy, and enables end-to-end training of the entire network. The uncertainty estimation module is combined to evaluate the disparity confidence, providing a reliability score for each pixel's disparity estimation. This step identifies and quantifies the uncertainty of the matching results (e.g., in weak texture and occluded areas), and the generated confidence map can be used as a weight in subsequent processes (such as point cloud filtering and SLAM optimization), reducing the impact of unreliable measurements on overall system accuracy.

[0110] A spatiotemporal consistent ConvGRU recurrent convolutional unit is introduced to establish an iterative optimization mechanism, and a multi-scale pyramid optimization strategy is used to optimize the initial disparity. A spatiotemporal consistent ConvGRU recurrent convolutional unit is introduced to establish an iterative optimization mechanism, and temporal information is used for optimization. ConvGRU is a unit that combines convolution operation and recurrent neural network (RNN), which can process sequence data and maintain 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 "remembers" and integrates the disparity information of previous frames. This approach can improve temporal consistency, making the disparity map results of consecutive frames smoother and more stable, avoiding flickering and jumping. The occluded area is optimized, and the historical information is used to infer the disparity value of the occluded area in the current frame. Noise can be suppressed, and random noise in single-frame images can be effectively smoothed by fusing multiple frames of information.

[0111] A multi-scale pyramid optimization strategy is used to optimize the initial disparity, and the accuracy is gradually refined from coarse to fine. The algorithm repeatedly performs disparity calculation on image pyramids of different resolutions (scales). First, a rough disparity is calculated at the lowest resolution (coarsest scale), and then the result is used as the initial value to correct and refine it at higher resolutions (finer scales).

[0112] 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.

[0113] 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.

[0114] 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:

[0115] Parallax optimization is performed using the photometric error model of the Direct Sparse Odometry (DSO). The photometric error calculation formula is as follows:

[0116] ,

[0117] 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 is the exposure compensation parameter for the frame, used to compensate the overall brightness difference between different frames due to auto exposure or illumination changes. and is the exposure time, used to eliminate the image brightness scaling effect caused by different exposure times. and is the exposure compensation parameter; represents the image intensity (brightness value) at pixel point in the reference frame . is the two-dimensional coordinate of the image pixel. represents the image intensity (brightness value) at pixel point in the matching frame . is the point in the reference frame projected to the matching frame according to the camera pose and depth information.

[0118] The photometric error model of direct sparse odometry (DSO) is used for disparity optimization, bypassing feature extraction and matching, and directly using pixel brightness information for optimization. This step optimizes the disparity by minimizing the brightness difference (photometric error) between corresponding pixel blocks in the reference frame and the matching frame. It works better for areas lacking texture features (such as walls and sky), and can generate a more dense disparity map.

[0119]

[0120] ,

[0121] where, is the total objective function of the joint optimization problem, composed of the photometric error term and the smoothing term. The goal of optimization is to find the best camera pose and depth map that minimizes the function value. is the index variable, refers to a certain point in three-dimensional space, refers to a certain key frame where the point is observed, is the weight coefficient, used to assign different importance to each residual term. It can be predicted according to the residual size or through a deep learning network to reduce the influence of outliers. is the image intensity value at the pixel position where the three-dimensional point is projected onto the image plane in the key frame . is the image intensity value corresponding to the three-dimensional point in the reference frame where the point is created. Huber Norm, a robust loss function. For small errors, it adopts L2 norm (good for convergence), for large errors, it adopts L1 norm (reduce the impact of outliers), γ is the switching threshold, is the gradient of the depth map at , is the smoothness term weight; : L1 norm of the depth map gradient, used as a smoothness regularizer. The purpose is to encourage the depth map to be smooth (i.e. neighboring pixels have similar depth values), avoiding unreasonable sharp changes.

[0122] Geometric error constraints are introduced to jointly optimize camera poses and 3D point positions, ensuring the correctness and consistency of the optimization results in geometry. This step combines photometric consistency with geometric consistency (such as image gradient smoothness, 3D point and camera pose constraints) to establish a more comprehensive objective function for joint optimization, avoiding geometric distortion that may result from optimizing photometric error alone.

[0123] A multi-scale pyramid optimization strategy is adopted to minimize photometric error from coarse to fine at different resolution levels:

[0124] ,

[0125] where, is the total photometric error on the layer of the image pyramid. is the pyramid level, is the level weight coefficient, is the level index variable, is the layer image; is the weight coefficient calculated on the layer pyramid. is the pixel intensity of the three-dimensional point projected onto the layer pyramid image of the key frame . is the pixel intensity of the three-dimensional point on the layer pyramid image of its reference frame, is the pixel intensity of the three-dimensional point projected onto the layer pyramid image of the key frame .

[0126] A multi-scale pyramid optimization strategy is adopted to perform coarse-to-fine photometric error minimization at different resolution levels to improve the convergence and accuracy of the optimization process. The step first calculates a rough disparity and structure on a low-resolution image as an initial value for higher resolution layers, which is gradually refined. This strategy expands the convergence range, avoids the optimization process from falling into a local minimum, and ultimately obtains high-precision details.

[0127] A vision-inertial tightly coupled optimization framework is established in combination with IMU pre-integration data, and the objective function is:

[0128] ,

[0129] Wherein, is the total objective function of the vision-inertial tightly coupled optimization framework, is a vision error term, i.e. the above photometric or re-projection error, is an IMU error term, which is the difference between the predicted state calculated based on the IMU pre-integration theory and the optimized state, is a prior error term, which is the prior constraint information generated from the marginalization operation, used to maintain consistency with the old frame and prevent information loss; in combination with the IMU pre-integration data, a vision-inertial tightly coupled optimization framework is established, which provides high-frequency, short-time accurate motion priors for vision optimization. The IMU pre-integration provides a prediction of the relative motion between two images, and this constraint is added to the optimization framework. This makes the system more robust in image blur, fast motion or texture missing, and can estimate quantities such as scale and gravity direction that cannot be observed by pure vision systems.

[0130] Adaptive weight adjustment based on deep learning is realized, and the confidence weight of each pixel is predicted by a neural network. The function expression for predicting the weight by a convolutional neural network is, for example, a lightweight convolutional neural network, which functions to learn and predict the photometric consistency confidence of the two images in the vicinity of the point Adaptive weight adjustment based on deep learning is realized, and the confidence weight of each pixel is predicted by a neural network:

[0131] ,

[0132] Wherein, is the function expression for predicting the weight by a convolutional neural network, is a lightweight convolutional neural network, which functions to learn and predict the photometric consistency confidence of the two images in the vicinity of the point and output the weight , represents an image block input into the neural network, The position information of the points can be provided as additional input to the network.

[0133] An adaptive weight adjustment based on deep learning is implemented to predict the confidence weight of each pixel by a neural network, intelligently reducing the influence of unreliable pixels (outliers) in optimization. This step uses a lightweight CNN network to automatically determine the photometric consistency confidence of each pixel (e.g., whether the pixel is occluded, in a light abrupt change area, or on a dynamic object), and assigns a lower weight to pixels with low confidence, thereby significantly improving the robustness of the system in complex real scenes.

[0134] The functions of the above steps collectively constitute a more advanced and robust optimization system. Instead of relying on traditional feature matching, it directly optimizes pixel brightness and integrates multiple constraints such as geometry and inertia, and finally intelligently removes outliers through deep learning. This combination aims to produce a more accurate, denser, and more reliable disparity map and depth information, laying a solid data foundation for subsequent semantic segmentation and risk assessment.

[0135] S2: Perform epipolar rectification and stereo matching on the binocular vision image pair to generate depth point cloud data, and perform real-time obstacle detection and warning based on the data. Recover three-dimensional information from two-dimensional images. By performing epipolar rectification and stereo matching on the generated image pair, calculating the disparity and converting it into depth information, depth point cloud data of the scene is generated. On this basis, preliminary real-time obstacle detection and warning based on geometric distance are realized.

[0136] Specifically, in this embodiment, the generation of depth point cloud data in S2 and the real-time obstacle detection and warning based on the data specifically include:

[0137] A multi-modal sensor fusion stereo matching framework is constructed to achieve motion interference suppression through tight coupling of IMU, RTK, and vision data. The robustness of stereo matching under intense motion of the unmanned aerial vehicle is improved. By fusing high-frequency attitude data of the IMU and absolute position information of the RTK, image motion blur and geometric distortion caused by unmanned aerial vehicle shaking, wind speed, etc. are compensated and eliminated, providing stable and reliable image input for stereo matching.

[0138] An adaptive epipolar rectification technique is used to dynamically adjust the rectification model according to the IMU attitude data and gimbal motion parameters, eliminating image geometric distortion; solving the problem of traditional epipolar rectification failure caused by continuous platform motion. Traditional rectification is for fixed binocular systems, while this invention dynamically calculates and applies the epipolar rectification model based on real-time sensed platform attitude (IMU) and gimbal state, ensuring that even if the platform is moving, the matching search can be performed on the correct horizontal line, ensuring matching efficiency and accuracy.

[0139] The semi-global matching algorithm with hybrid cost aggregation is used to fuse the Census-Gradient hybrid cost function and the multi-scale aggregation strategy, and the aggregation path is optimized based on the attention mechanism. The best balance between precision and efficiency is achieved to realize robust pixel matching. Hybrid cost function: combine the Census transform which is insensitive to light and the gradient information which is sensitive to texture to enhance the discriminability of the matching cost and adapt to complex outdoor lighting changes. Semi-global matching and multi-scale aggregation: through multi-path optimization and cross-scale information fusion, while maintaining high computational efficiency, a noise-free and more consistent disparity map is obtained.

[0140] Census transform: a non-parametric local transform. It generates a bit string descriptor by comparing the relative relationship (not the absolute value) of the brightness of a pixel and its neighborhood pixels. This makes it extremely invariant to linear changes in light and gamma changes, making it very suitable for complex lighting conditions in outdoor inspection. Gradient information: captures the edge and texture features of the image. It is very sensitive to the geometric structure of objects and can provide rich detail information, but it is limited in uniform areas and sensitive to noise. Hybrid strategy: weighted fusion of the two (e.g. Cost_hybrid = a * Cost_census + β * Cost_gradient). This realizes the complementary advantages: the light robustness of Census provides stability for matching, while the detail sensitivity of gradient ensures the accuracy of matching, especially in object edge areas.

[0141] Multi-scale aggregation strategy: realize reliable matching from coarse to fine, eliminate ambiguity at low resolution, restore details at high resolution, and ensure the accuracy and integrity of matching. Cost aggregation is performed on low-resolution images (coarse scale). At this time, the image texture is simplified, and large-scale unique patterns are easier to match, which can reliably estimate the general object outline and larger disparity, but details are lost. The coarse scale matching result is 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 make fine corrections within a small disparity range. This strategy greatly reduces the computational load and avoids ambiguity and errors caused by searching the entire disparity range at the fine scale.

[0142] Optimization of aggregation path based on attention mechanism: adaptive cost aggregation is realized, and the aggregation process is intelligently guided, so that the algorithm can focus on information-rich areas and paths, and the utilization rate of effective information is improved. Traditional semi-global matching (SGM) performs cost aggregation along fixed multiple (such as 8 or 16) paths, and treats all pixels "equally". This is not good in weak texture areas or repeated texture areas, because the paths in these areas may be ambiguous. The introduction of attention mechanism, the algorithm dynamically calculates the "importance weight" of each pixel on each aggregation path. For areas with rich texture and high confidence (such as conductor edges and insulator textures), their aggregation paths will be given high weights, and these reliable information will be fully propagated. For weak texture, occlusion or possibly noisy areas, their path weights will be suppressed to prevent unreliable information from polluting the global cost calculation. This is equivalent to installing a "smart navigation" for the aggregation process, which can autonomously select the optimal information propagation path, significantly improving the matching success rate in challenging areas.

[0143] Hybrid cost aggregation semi-global matching algorithm: global optimization of disparity map, the local matching cost generated by the above steps is integrated into a globally consistent, smooth and accurate disparity map through an efficient global optimization method. The core idea of "semi-global": approximate solution to the two-dimensional global energy function minimization problem through one-dimensional optimization along multiple paths (such as 8, 16 directions), achieving a perfect balance between computational complexity and result accuracy. "Hybrid cost aggregation": refers to the integration of the above hybrid cost function, multi-scale information and attention weight into the aggregation framework of SGM. It is not simply the sum of the results of each path, but a weighted fusion, where the weight is determined by the attention mechanism.

[0144] Disparity optimization network combining Kalman filter and deep learning, predict disparity changes by constructing motion model with IMU data, and use time filtering to improve the spatio-temporal consistency of disparity map; further optimize the quality of disparity map to ensure the smoothness and stability of the output. Kalman filter: use IMU data to predict the trend of scene motion, filter and smooth the disparity map in time domain, and reduce jitter. Deep learning network: repair occlusion and error matching areas. Spatio-temporal consistency: combine the advantages of both, output a continuous and consistent disparity result in time and space, avoiding flicker and jump of frame-by-frame results.

[0145] Establish a dynamic baseline binocular imaging model to calculate the baseline length in real time and achieve millimeter-level baseline precision control; provide core parameters for high-precision triangulation. Accurately control and measure the dynamic baseline length generated by the translation of the gimbal and the flight of the UAV in real time, which is the key to converting disparity into depth information. Millimeter-level control accuracy directly ensures the centimeter-level or even millimeter-level precision of the final depth / distance measurement.

[0146] Construct a real-time early warning system based on deep learning, use a lightweight network to achieve millisecond-level obstacle detection, and establish a multi-threshold early warning mechanism and dynamic region of interest management. Realize the final value of the business level-real-time risk early warning. Lightweight network: ensure the real-time operation of the algorithm on the on-board computing platform (millisecond-level response). Multi-threshold and dynamic ROI: Set different levels of risk thresholds according to safety regulations, and dynamically focus on the area near the conductor (ROI) to realize the closed loop from "perception" to "early warning" and timely output different levels of risk warning.

[0147] S3: Perform semantic segmentation on the depth point cloud data, identify and extract key targets such as conductors, trees, and insulators, and generate point cloud data with semantic labels. Through semantic understanding of the environment, semantic information is given to the original point cloud. Use a deep learning network to segment the depth point cloud, automatically identify, classify, and extract key targets such as "conductor", "tree", "insulator", etc., and convert unordered point cloud data into structured semantic maps with clear class labels, laying the foundation for subsequent accurate analysis and risk assessment.

[0148] Specifically, in this embodiment, the point cloud semantic segmentation step in step S3 specifically includes:

[0149] S3.1, use a point cloud semantic segmentation network based on deep learning to process the depth point cloud data, the network includes: a point cloud feature extraction module, use PointNet++ or KPConv network to extract multi-scale point cloud features; an attention mechanism module, introduce attention weight in the early warning area to enhance the feature representation of the key area; a semantic segmentation head, output the semantic class probability distribution of each point through a multi-layer perceptron.

[0150] Extract and learn the deep features of the point cloud to complete the preliminary point-by-point classification. Point cloud feature extraction module: as the network backbone, efficiently capture local and global geometric features of point cloud. PointNet++ is good at processing point clouds of different densities, while KPConv (convolution kernel point convolution) can provide more detailed convolutional feature extraction, both of which provide rich feature representation for subsequent classification. Attention mechanism module: enhance the feature response of the key area and suppress background interference. By calculating the attention weight, the network pays more "attention" to the point cloud in the early warning area (Region of Interest, ROI) near the conductor and the edge of the tree crown, improving the classification accuracy of key targets. Semantic segmentation head: performs the final classification decision. It receives deep features fused with attention weights and outputs a probability distribution belonging to each semantic class (such as conductor, tree, insulator, background, etc.) for each point through an MLP.

[0151] S3.2, Key targets are identified and extracted: Conductors: Based on spatial continuity and geometric features extraction, RANSAC algorithm is used to fit conductor curves; Trees: According to point cloud density distribution and vegetation features recognition, combined with depth information to distinguish different tree species; Insulators: Based on geometric shape features and spatial position relationship detection, template matching method is used for accurate positioning; For the geometric characteristics of specific target objects, special algorithms are used for accurate extraction and model fitting.

[0152] Conductors: Use their spatial continuity and linear geometric features. RANSAC (Random Sample Consensus) algorithm is used to fit the three-dimensional spatial curve model of the conductor, so as to accurately restore its trend and position, rather than a bunch of scattered points. Trees: Use the characteristics of sparse spatial distribution and complex structure of vegetation point cloud (high density variation, irregular shape). Combined with depth information, the morphological characteristics of different tree species (such as trees vs. shrubs) can be further distinguished. Insulators: Use its regular and repeated geometric shape (such as disc shape) and fixed spatial position relationship with conductors. Shape-based template matching method is used to achieve sub-point cloud level accurate positioning.

[0153] S3.3, Post-processing optimization of segmentation results: Apply Conditional Random Field (CRF) to optimize semantic boundaries and eliminate isolated misclassified points; Based on spatial connectivity, perform clustering analysis and merge fragmented segmentation areas; Use temporal consistency constraints to fuse multi-frame point cloud segmentation results to improve stability; Optimize the preliminary segmentation results to eliminate noise and ambiguity, and improve the smoothness, consistency and stability of the results.

[0154] Conditional Random Field (CRF) optimization: Consider the spatial proximity and feature similarity between points, smooth the semantic boundaries, and correct isolated misclassified points that are inconsistent with the surrounding points. Clustering analysis based on spatial connectivity: Merge scattered points belonging to the same instance (such as a tree's point cloud) into a complete object, eliminating fragmented segmentation results. Temporal consistency constraints: Fuse the segmentation results of consecutive frames to stabilize the classification results of the current frame using temporal information, reducing the jitter and uncertainty of single-frame segmentation.

[0155] S3.4, Generate point cloud data with semantic labels, including: three-dimensional coordinates, color information and semantic class labels of each point; bounding box and geometric parameters of key targets; confidence score and spatial relationship description of targets in the warning area; Establish a target database to record historical information of the identification results, providing multi-time sequence comparison and data tracing capabilities. Output structured and information-rich semantic maps, and provide direct input for risk assessment and data analysis.

[0156] It not only contains the basic attributes of each point (coordinates, color, category), but also outputs high-level semantic information: the bounding box and geometric parameters of key targets, which provide the basis for subsequent safety distance calculation (e.g., calculating the minimum distance from the outer envelope box of the tree point cloud to the conductor point cloud). Confidence score and spatial relationship: identify the reliability of each recognition result (e.g., "there is a 90% chance that this is a conductor") and describe the relative position between targets (e.g., "the insulator is connected below the conductor"). Target database and historical records: achieve traceability and time sequence comparison of data, analyze the growth rate of trees, evaluate the effect of cutting, and provide long-term data support for operation and maintenance decisions.

[0157] S4: Establish a reduced dimension motion model using the gimbal stabilization feature, fuse visual inertial odometry and real-time differential positioning data through multi-state constraint Kalman filtering or sliding window optimization algorithm, and realize real-time pose estimation with centimeter-level accuracy. High-precision self-positioning provides a real-time pose (position and attitude) reference with centimeter-level accuracy for the entire system. Simplify the motion model using the gimbal stabilization feature, and use a tightly coupled algorithm to deeply integrate visual (VIO), inertial (IMU), and satellite (RTK) data, complement each other's advantages, effectively suppress the errors and drift of any single sensor, and ensure that the perception results are mapped to the accurate global coordinate system.

[0158] Specifically, in this embodiment, the real-time pose estimation with centimeter-level accuracy in step S4 is a tightly coupled SLAM pose estimation, specifically including:

[0159] A reduced dimension motion model is established using the gimbal stabilization feature of the unmanned aerial vehicle, which simplifies the 6-degree-of-freedom pose estimation problem to a 3-degree-of-freedom translation estimation. A reduced dimension motion model is established using the gimbal stabilization feature of the unmanned aerial vehicle, which simplifies the complexity of state estimation, improves the calculation efficiency and numerical stability. The stabilization function of the gimbal actively compensates for the rotational jitter of the unmanned aerial vehicle, so that the image sequence captured by the camera primarily experiences translation motion. This allows the 6-degree-of-freedom pose estimation problem (3 translations + 3 rotations) to be approximately simplified to a 3-dimensional translation estimation problem, significantly reducing the complexity of the optimization problem.

[0160] The visual inertial odometry VIO and real-time differential positioning RTK are tightly coupled and deeply integrated through multi-state constraint Kalman filtering or sliding window optimization algorithm, wherein:

[0161] RTK provides an absolute position reference and suppresses IMU cumulative drift;

[0162] The vision system provides relative pose estimation and environment perception when the RTK signal is lost;

[0163] Joint optimization is performed using reprojection error and IMU pre-integration error;

[0164] The optimization objective function of the tightly coupled deep fusion is: :

[0165] ,

[0166] 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. , 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.

[0167] The visual-inertial odometry (VIO) and real-time kinematic (RTK) are tightly coupled and deeply fused through multi-state constraint Kalman filtering or sliding window optimization algorithm to realize the complementary advantages of multi-source sensors and provide absolute accurate and relatively smooth pose output. This is the core of the step. The role of RTK: providing an absolute position reference (centimeter-level accuracy) and effectively suppressing the cumulative drift caused by the IMU accelerometer zero offset, the pose error is constrained within a limited range to prevent its unlimited growth. The role of the vision system: providing high-frequency relative pose changes and rich environmental feature information. When the RTK signal is temporarily lost (such as under bridges, near trees), VIO can seamlessly take over to ensure the continuity of pose estimation. The connotation of "tight coupling": unlike simply selecting or switching data, tight coupling uses visual observations, IMU measurements, and RTK observations together in a unified optimization framework to estimate the same set of state variables, thereby achieving the deepest level of information fusion.

[0168] Through joint optimization of re-projection error and IMU pre-integration error, the mathematical basis of fusion optimization is constructed to provide double constraints for state estimation. Re-projection error: visual constraint. It measures the difference between the position of the 3D map point projected back to the image plane and the actual observed feature point position, ensuring that the estimated pose is consistent with the observed environmental features. IMU pre-integration error: kinematic constraint. It measures the difference between the predicted pose change 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 resulting pose estimate not only conforms to the environmental geometric features but also conforms to the motion dynamics model, making the result more accurate and reliable.

[0169] The optimization objective function of the tight coupling fusion defines the specific implementation method of "joint optimization" in mathematical formula form. This objective function explicitly indicates that the weighted sum of the IMU pre-integration residual and the visual re-projection residual is to be minimized. The weight is determined by the respective covariance matrix, and the weight of the sensor data with greater uncertainty naturally decreases, reflecting the intelligent consideration of the algorithm for sensor noise characteristics.

[0170] The above steps build the "brain" and "navigation system" of the entire system. It seamlessly integrates the continuous relative pose of VIO, high-frequency motion data of IMU, and accurate absolute position of RTK through innovative dimension reduction models and deep tight coupling fusion, finally outputting a centimeter-level accurate, absolute reference, continuous and reliable pose information. This high-precision space-time reference is the fundamental prerequisite for all subsequent processing (such as point cloud stitching, three-dimensional reconstruction, distance measurement) to be successful.

[0171] The tight coupling SLAM pose estimation step further comprises:

[0172] The state vector X is iteratively optimized by a sliding window optimization algorithm, and the sliding window contains multiple key frames and their corresponding IMU measurement values and visual feature points; the state vector X is iteratively optimized by the sliding window optimization algorithm, and the best balance between calculation complexity and optimization accuracy is achieved. The sliding window mechanism only retains a series of recent key frames and their states (pose, velocity, bias, map points) into the optimization window. When a new frame arrives, the old frame is moved out of the window. This avoids the size of the optimization problem from growing indefinitely over time, maintaining the calculation amount at a constant level and meeting the real-time requirement. At the same time, compared with the filtering method considering only the current frame, optimizing multiple frames can obtain higher accuracy.

[0173] An edge strategy is introduced to process the historical state outside the window, to retain the constraint information to avoid information loss, and to control the calculation complexity; the edge strategy is introduced to process the historical state outside the window, to retain the valuable constraint information contained in the old frame moved out of the window, and to prevent "information loss". When a frame is moved out of the sliding window, it is not simply discarded. The edge strategy converts the state of the frame into a prior probability constraint on the remaining states in the window, and adds it to the subsequent optimization problem. In this way, the size of the optimization problem is controlled, and the constraint of the historical observation on the current state is retained, effectively suppressing the cumulative error of long-term operation, which is a key technology to ensure the consistency of the algorithm.

[0174] Robust kernel is used to weight the reprojection error and IMU error, to suppress outliers and improve the robustness of the system in dynamic environment; the robust kernel is used to weight the reprojection error and IMU error, to improve the robustness of the system in dynamic and complex environment, and to suppress the destructive influence of abnormal observations. The standard optimization uses least square (L2 norm) by default, which is very sensitive to outliers. The robust kernel (such as Huber, Cauchy) reduces the weight of large residual terms to weaken or even eliminate the influence of abnormal observations caused by dynamic objects (such as moving vehicles, swaying branches), incorrect data association, instantaneous occlusion, etc., so that the state estimation result is more stable and reliable.

[0175] The online calibration technology is used to estimate and compensate the extrinsic parameter deviation between the camera and the IMU in real time, so as to ensure the consistency of the multi-sensor data. The online calibration technology is used to estimate and compensate the extrinsic parameter deviation between the camera and the IMU in real time, so as to ensure the consistency of the multi-sensor data. The rigid transformation (extrinsic parameter) between the camera and the IMU is the geometric basis of the tightly coupled fusion. The parameter will change slightly (extrinsic parameter deviation) due to factors such as vibration and temperature difference. The online calibration technology adds the extrinsic parameter as a state variable into the optimization process for real-time estimation, and automatically compensates for the deviation, so as to ensure that the visual projection and the IMU integration are strictly consistent in geometry, and the fusion performance is prevented from being sharply reduced due to the inaccuracy of the extrinsic parameter.

[0176] The step of weighting the reprojection error and the IMU error by using a robust kernel function specifically includes: weighting and optimizing the visual reprojection residual and the IMU pre-integration residual by using a robust kernel function, and converting a standard least squares optimization problem into a more robust optimization problem. This step is the declaration of the core idea, and clearly applies the robust kernel to the two key residuals (visual geometric error and IMU kinematic error), and lays a foundation for subsequent specific kernel function selection. The robust kernel function includes a Huber kernel function, a Cauchy kernel function or a Tukey double-weight kernel function; and multiple mathematical tools are provided to cope with different abnormal value scenarios. Different kernel functions have different suppression characteristics: the Huber kernel: maintains quadratic growth (efficient optimization) for small errors, and converts to linear growth (suppresses abnormal values) for large errors, and is a commonly used choice to balance efficiency and robustness. The Cauchy kernel: the suppression of large errors is more smooth and gradual, and is not easy to produce gradient mutation, and the numerical behavior is more stable. The Tukey kernel: completely discards the error exceeding the threshold (the weight is reduced to zero), and is the most aggressive strategy for handling significant abnormal values.

[0177] The mathematical expression of the Huber kernel function is: ,

[0178] The mathematical expression of the Cauchy kernel function is: ,

[0179] The mathematical expression of the Tukey double-weight kernel function is:

[0180] ,

[0181] wherein, is a residual value, is a threshold parameter, also referred to as an inflection point, is a scale parameter, is a kernel function.

[0182] According to the Mahalanobis distance of the residual, the type of kernel function is dynamically selected and the parameters of the kernel function are adjusted, and the abnormal residual exceeding the confidence interval is subjected to nonlinear weight reduction processing, effectively suppressing the interference of dynamic obstacles, sudden changes in light, and feature mismatching on pose estimation, and improving the estimation accuracy and robustness of the system in complex environments.

[0183] Adaptive intelligent weighting is achieved, so that the system can automatically adjust the suppression strategy according to the severity of the error. Mahalanobis distance: a distance measure that takes into account the correlation of data, which can more accurately measure the abnormality of a residual relative to its expected distribution (described by the covariance matrix). Dynamic selection and adjustment: the system no longer fixedly uses a kernel function or parameter. For example, for mild abnormalities (small Mahalanobis distance), Huber kernel may be selected, and for severe abnormalities (extreme Mahalanobis distance), Tukey kernel may be selected; at the same time, the threshold δ or the scale parameter c can also be dynamically adjusted according to the residual distribution. This enables the system to perceive the reliability of the error and respond intelligently.

[0184] Nonlinear weight reduction processing is performed on abnormal residuals exceeding the confidence interval, which is the core execution action, directly weakening the influence of abnormal observations on the final solution. "Nonlinear" means that the weight is not uniformly reduced, but sharply decreases according to the kernel function curve, ensuring that a small number of bad data does not dominate the optimization result. This effectively suppresses the interference of dynamic obstacles, sudden changes in light, and feature mismatching on pose estimation. This part points out the typical sources of abnormality for robust kernel processing: dynamic obstacles: feature points on moving objects such as vehicles and pedestrians will produce false constraints that violate the static world assumption. Light sudden change: causes a dramatic change in visual appearance, resulting in feature matching errors or photometric error failures. Feature mismatching: the most common source of error in visual odometry. By suppressing these disturbances, the overall goal of improving the estimation accuracy and robustness of the system in complex environments is ultimately achieved.

[0185] The above steps introduce a "smart arbitrator" to 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 the "evidence" of Mahalanobis distance. For "lying" outliers, it will impose appropriate "punishment" (weight reduction) according to the "severity of the lie" (abnormality), so as to ensure that the final decision (pose estimation) is not misled, and is more accurate and reliable. This is a key technical guarantee for realizing high-precision operation of the system in actual complex scenarios.

[0186] S5: Perform voxel filtering and downsampling processing on the point cloud data with semantic labels, complete the three-dimensional reconstruction of key targets using a multi-modal fusion algorithm, and output risk assessment results based on tree line spacing calculation and safety margin analysis. Through refined reconstruction and risk decision-making, the final risk assessment results are generated. The semantic point cloud is optimized (denoising, compression), and the multi-modal algorithm is used to complete the refined three-dimensional reconstruction of the key target. Finally, based on the accurate three-dimensional model, the tree line spacing is calculated, and after considering the safety margin, the quantitative risk analysis is performed, and the decision report that can be used to guide the operation and maintenance is output, realizing the closed loop from perception to analysis.

[0187] Specifically, in the present embodiment, S5 specifically includes:

[0188] An adaptive spatial voxel grid division method is used to dynamically adjust the voxel size according to the point cloud density distribution, small size voxels are used in dense areas to retain detailed features, and large size voxels are used in sparse areas to improve processing efficiency; through a double threshold mechanism, effective voxels are screened, first based on the number of point clouds in the voxel for primary screening, and then combined with the point cloud density distribution characteristics for secondary screening, to remove outliers and retain real object voxels; the effective voxels after screening are subjected to intelligent downsampling processing, representative point clouds are selected based on information entropy evaluation, and a strategy combining random sampling and feature preservation is used to ensure the integrity of the features while avoiding grid effects; the optimized point cloud data is input into the multi-modal fusion tree barrier analysis module, an improved ICP algorithm is used in combination with semantic constraints for accurate registration, or a deep learning network is used to realize refined three-dimensional reconstruction of key targets; tree barrier risk quantification analysis is performed, the minimum spatial distance between trees and conductors is calculated, safety margin factors such as thermal expansion and wind deflection are considered, a multi-parameter risk assessment model is established to evaluate the risk level; and finally a complete analysis report including a three-dimensional spatial relationship diagram, a risk distribution diagram and a warning suggestion is generated to provide decision support for distribution network tree barrier prevention and control.

[0189] The five steps form a complete "perception-positioning-understanding-decision" technical closed loop: S1-S2 is responsible for perceiving the three-dimensional world, S4 provides a high-precision positioning reference for the entire process, S3 is responsible for understanding the environment semantics, and finally S5 performs risk assessment and decision-making.

[0190] Embodiment Two

[0191] As shown in Figure 2 The present application provides a distribution network tree barrier real-time analysis system architecture based on dynamic vision and SLAM, which is applied to the distribution network tree barrier real-time analysis system based on dynamic vision and SLAM as described in Embodiment One, and 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.

[0192] A dynamic binocular vision perception module 11 is configured to generate a dynamic visual baseline by controlling gimbal panning motion of the UAV to cooperate with flight displacement of the UAV, construct a bionic binocular parallax imaging model, and simulate a time-series image sequence as a binocular vision image pair.

[0193] A depth information generation and early warning module 12 is configured 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.

[0194] A point cloud semantic segmentation module 13 is configured to perform semantic segmentation on the depth point cloud data, identify and extract key targets such as conductors, trees, and insulators, and generate point cloud data with semantic labels.

[0195] A tightly coupled SLAM pose estimation module 14 is configured to establish a reduced dimension motion model using gimbal stabilization characteristics, fuse visual inertial odometry and real-time differential positioning data through multi-state constraint Kalman filtering or sliding window optimization algorithm, and realize real-time pose estimation with centimeter-level precision.

[0196] A point cloud optimization and analysis module 15 is configured to perform voxel filtering and down-sampling processing on the point cloud data with semantic labels, complete three-dimensional reconstruction of key targets using a multi-modal fusion algorithm, and output risk assessment results based on tree line spacing calculation and safety margin analysis.

[0197] Figure 3 The electronic device provided in an embodiment of the present application. As shown in Figure 3 The electronic device at least includes the following parts: a processor 101 and a memory 100, a communication interface 103, and a bus 102.

[0198] In the embodiments of the present application, the memory 100 is configured to store processor 101 executable instructions, and the processor 101 is configured to implement the method of the first aspect when executing the instructions.

[0199] In the embodiments of the present application, a computer readable storage medium includes instructions, and the instructions instruct the device to execute the method of the first aspect. For example, the instructions instruct the device to execute the method shown in the flow steps of Figure 1

[0200] ​The program that works in the electronic device according to the embodiment of the present application can be a program that controls a central processing unit (CPU) or the like to realize the functions of the above-described embodiments according to one aspect of the present application (a program that causes a computer to function). Then, the information processed by these systems is temporarily stored in a random access memory (RAM) when it is processed, and then stored in various ROMs such as a read only memory (Flash ROM) and a hard disk drive (HDD), and read out, corrected, and written by the CPU as necessary.

[0201] Note that a part of the electronic device according to the above-described embodiments can also be realized by a computer. In this case, a program for realizing the control function can be recorded in a computer-readable recording medium, and realized by reading the program recorded in the recording medium into a computer and executing it.

[0202] Note that the computer referred to here means a computer built in the electronic device, and a computer including an OS, a peripheral device, and the like. Further, the computer-readable recording medium means a removable medium such as a floppy disk, a magneto-optical disk, a ROM, a CD-ROM, and the like, a storage system such as a hard disk built in the computer.

[0203] Further, the computer-readable recording medium can include a medium that dynamically stores a program for a short time, such as a communication line in the case of transmitting a program via a network such as the Internet or a communication line such as a telephone line, and a medium that stores a program for a fixed time, such as a volatile memory in the computer that is a server or a client in this case. Further, the above-described program can be a program for realizing a part of the above-described functions, and can also be a program that can realize the above-described functions by being combined with a program already recorded in a computer.

[0204] Further, the electronic device according to the above-described embodiments can also be realized as an aggregate (system group) constituted by a plurality of systems. Each system that constitutes the system group can have a part or all of each function or each functional block of the electronic device according to the above-described embodiments. As the system group, all of each function or each functional block of the electronic device can be included.

[0205] Those skilled in the art will recognize that the above embodiments are merely illustrative of the present application and should not be taken as limiting. Variations and modifications of the embodiments set forth herein can be made by those skilled in the art without departing from the spirit of the present application, and do fall within the scope of the present 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 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 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... 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