Power line detection method and system based on closed-form evolutionary bayesian probability map

Abstract

The application belongs to the technical field of power line detection, and discloses a power line detection method and system based on a closed evolutionary Bayesian probability map, which comprises the following steps: acquiring point cloud data and performing frame processing; extracting power line candidate points of the first frame of point cloud data and performing clustering analysis, performing power line spatial segment reconstruction and connectivity analysis based on the clustering results, and obtaining an initial power line track; constructing a probability map and extracting a region of interest of the probability map; initializing a MaxUKF tracker based on the probability map, using the initialized MaxUKF tracker to perform track prediction on the point cloud data of subsequent frames, using the region of interest to crop the point cloud data of the subsequent frames, performing local detection on the cropped point cloud data, completing the predicted track according to the local detection result, and outputting complete power lines. The application can realize a power line detection method with the functions of real-time detection, closed probability evolution, adaptive weight, non-Gaussian distribution processing and intelligent completion.

Description

Technical Field

[0001] This invention relates to the field of power line detection technology, and in particular to a power line detection method and system based on a closed evolutionary Bayesian probability map. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] With the rapid development of technologies such as drone inspection and smart grid maintenance, the demand for automated detection and monitoring of power lines is increasing. Traditional power line detection methods mainly rely on manual inspection or image-based visual inspection, which suffers from low efficiency, significant susceptibility to ambient lighting, and inability to obtain accurate three-dimensional information. LiDAR technology can provide high-precision three-dimensional point cloud data, offering a new technological approach for power line detection.

[0004] However, existing lidar power line detection methods still face many challenges in terms of real-time performance, intelligence, and adaptability. First, traditional methods typically perform global detection on each frame of point cloud, resulting in high computational costs and difficulty in meeting real-time processing requirements. Second, in complex environments, power lines are easily obscured for short periods by obstacles such as trees and construction vehicles, leading to intermittent detection. Third, existing methods lack the ability to intelligently utilize historical detection information and perform probabilistic reasoning, failing to provide a reliable completion mechanism when detection fails. Finally, traditional Kalman filtering assumes a Gaussian distribution, which cannot accurately model the nonlinearity and multimodal uncertainties in power line tracking. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a power line detection method and system based on a closed-loop evolutionary Bayesian probabilistic map. This method enables real-time detection, closed-loop probabilistic evolution, adaptive weighting, handling of non-Gaussian distributions, and intelligent completion capabilities.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides a power line detection method based on a closed evolutionary Bayesian probability map, comprising the following steps:

[0008] S1. Acquire 3D point cloud data from the LiDAR and perform frame-by-frame processing on the point cloud data;

[0009] S2. Extract candidate power line points from the first frame of point cloud data, cluster the candidate points, reconstruct and analyze the spatial segments of power lines based on the clustering of candidate power line points, merge the spatial segments of power lines, and obtain the initial power line trajectory.

[0010] S3, based on the initial power line trajectory, a closed evolutionary Bayesian inference is used to construct a probability map, and a region of interest of the probability map is extracted;

[0011] S4, based on the initial power line trajectory and the probability map, a MaxUKF tracker is initialized, the MaxUKF tracker after initialization is used to predict the trajectory of the point cloud data of the subsequent frame, the probability map is updated, the point cloud data of the subsequent frame is cropped using the region of interest, the cropped point cloud data is subjected to the local detection of S2, the predicted trajectory is completed according to the local detection result, and a complete power line is output.

[0012] As an optional implementation, a closed evolutionary Bayesian inference is used to construct a probability map, specifically:

[0013] A multi-modal enhanced observation variable is constructed, an information weight matrix based on causal adaptive weight is calculated, the multi-modal enhanced observation variable is optimally linearly combined based on the information weight matrix, a spatiotemporal coupling evolutionary dynamics is used to replace a traditional exponential decay, a nonlinear probability distribution of the observation variable is modeled, and an inter-line coupling constraint is introduced to obtain a final probability map.

[0014] As an optional implementation, an environment-adaptive prior updating mechanism is further established, and the prior distribution is dynamically adjusted according to historical observation data and regional characteristics.

[0015] As an optional implementation, the weight coefficients of the information weight matrix satisfy a normalization constraint and are dynamically adjusted according to the reliability of different observation sensors.

[0016] As an optional implementation, the inter-line coupling constraint is that the sum of the existence probabilities of all power lines at the same position should not exceed a physical upper limit, and the coupling evolution is realized through Softmax normalization.

[0017] As an optional implementation, the multi-modal enhanced observation variable includes a geometric observation component, a reflection observation component and a height observation component.

[0018] In a second aspect, the present application provides a power line detection system based on a closed evolutionary Bayesian probability map, comprising:

[0019] A data acquisition and preprocessing module is configured to acquire laser radar three-dimensional point cloud data and perform frame processing on the point cloud data.

[0020] An initial power line trajectory construction module is configured to extract power line candidate points of the first frame of point cloud data, cluster the candidate points, perform power line spatial segment reconstruction and connectivity analysis based on power line candidate point clustering, merge the power line spatial segments, and obtain an initial power line trajectory.

[0021] The probability map construction module is configured to: based on the initial power line trajectory, construct a probability map by using closed-form evolving Bayesian inference, and extract a region of interest of the probability map;

[0022] The power line detection module is configured to: initialize a MaxUKF tracker based on the initial power line trajectory and the probability map, use the initialized MaxUKF tracker to predict a trajectory of point cloud data of a subsequent frame, update the probability map, use the region of interest to crop the point cloud data of the subsequent frame, perform local detection on the cropped point cloud data, complete the predicted trajectory according to a local detection result, and output a complete power line.

[0023] In a third aspect, the present application provides an electronic device, comprising a memory and a processor, and computer instructions stored in the memory and running on the processor, when the computer instructions are run by the processor, the method of the first aspect is completed.

[0024] In a fourth aspect, the present application provides a computer readable storage medium for storing computer instructions, when the computer instructions are executed by a processor, the method of the first aspect is completed.

[0025] In a fifth aspect, the present application provides a computer program product comprising a computer program, when the computer program is executed by a processor, the method of the first aspect is completed.

[0026] Compared with the prior art, the present application has the following beneficial effects:

[0027] The present application proposes a power line detection method based on closed-form evolving Bayesian probability map, adopts closed-form evolving Bayesian inference, realizes optimal fusion of multi-source information through optimal enhanced observation variable construction and theoretically rigorous conditional probability calculation, and the information utilization efficiency is improved by more than 40% compared with traditional methods. The non-Gaussian tracker based on MaxUKF accurately processes the nonlinear and multi-modal uncertainty in power line tracking through Gaussian modeling and maximum point prediction, and the tracking accuracy is improved by 25-35%. The closed-form time evolution dynamics model replaces the traditional exponential decay, and through adaptive decay rate and equilibrium state modeling, the prediction accuracy and environmental adaptability of the probability map are improved. The deep fusion of the probability map and the MaxUKF tracker realizes the theoretically optimal state estimation, and the completion success rate under the shielding condition is more than 90%, and the completion accuracy error is controlled within 0.1m. The method significantly improves the processing efficiency and intelligent level while ensuring the detection accuracy, and provides a theoretically advanced and engineering practical technical solution for real-time monitoring of power lines.

[0028] The advantages of the additional aspects of the present application will be partially given in the following description, partially will become apparent from the following description, or will be understood by the practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0029] The accompanying drawings, which form a part of this specification, are included to provide a further understanding of the application, and are incorporated in and constitute a part of this specification. The embodiments of the application, and their

[0030] Figure 1 Flow chart of the power line detection method based on closed-form evolutionary Bayesian probability map of the present application;

[0031] Figure 2 Effect diagram of the multi-geometric feature power line extraction of the present application, wherein, Figure 2 (a) is the original point cloud diagram, Figure 2 (b) is the candidate point extraction result diagram, Figure 2 (c) is the final clustering result diagram;

[0032] Figure 3 Schematic diagram of the power line segment reconstruction and connection of the present application;

[0033] Figure 4 Schematic diagram of the closed-form evolutionary probability map construction and intelligent ROI extraction of the present application, wherein, Figure 4 (a) is the closed-form evolutionary probability map construction effect diagram, Figure 4 (b) is the intelligent ROI extraction effect diagram;

[0034] Figure 5 MaxUKF non-Gaussian tracking and intelligent completion effect diagram of the present application. DETAILED DESCRIPTION

[0035] The present application will be further described below in conjunction with the accompanying drawings and embodiments.

[0036] It should be noted that the following detailed description is merely exemplary in nature and is intended to provide further description of the application. 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.

[0037] It is also to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, steps, operations, elements, components, and / or groups thereof.

[0038] In the case of no conflict, the embodiments in the present application and the features in the embodiments can be combined with each other.

[0039] Embodiment 1

[0040] As Figure 1 shown, the embodiment provides a power line detection method based on a closed evolutionary Bayesian probability map, comprising the following steps:

[0041] S1, acquiring laser radar three-dimensional point cloud data, and performing frame processing on the point cloud data;

[0042] S2, extracting power line candidate points of the first frame of point cloud data, clustering the candidate points, performing power line spatial segment reconstruction and connectivity analysis based on the power line candidate point clustering, merging the power line spatial segments, and obtaining an initial power line track;

[0043] S3, based on the initial power line track, adopting a closed evolutionary Bayesian inference to construct a probability map, and extracting a region of interest of the probability map;

[0044] S4, initializing a MaxUKF tracker based on the initial power line track and the probability map, using the initialized MaxUKF tracker to perform track prediction on the point cloud data of the subsequent frames, updating the probability map, and using the region of interest to crop the point cloud data of the subsequent frames, performing local detection of the cropped point cloud data in S2, completing the predicted track according to the local detection result, and outputting a complete power line.

[0045] The specific scheme of the present application is as follows:

[0046] S1, acquiring laser radar three-dimensional point cloud data, and performing frame processing on the point cloud data;

[0047] (1) receiving original three-dimensional point cloud data output by a laser radar sensor, containing three-dimensional coordinates (x, y, z) and reflection intensity I information of each point. A voxel grid downsampling method is adopted to perform downsampling processing on the point cloud, and the voxel size is set to 0.3m. Voxel grid downsampling can reduce the original point cloud from an average of 50000 points to about 15000 points while maintaining the geometric features of the point cloud, and the data compression rate reaches 70%.

[0048] (2) according to the spatial range in which the power line is usually distributed, the point cloud is spatially cropped, and the point cloud data within the range of ±35m in width, ±35m in length and 0-35m in height centered on the sensor is reserved, the distant and ground point clouds are filtered out, and the spatial region where the power line may exist is focused.

[0049] (3) Height filtering based on the z-axis coordinate information of the point cloud, setting the height threshold range to 0.2m-3.5m, removing too low ground points and too high noise points, and further reducing the search space of the power line detection.

[0050] S2, extracting power line candidate points of the first frame of point cloud data, clustering the candidate points, performing power line spatial segment reconstruction and connectivity analysis based on the power line candidate point clustering, merging the power line spatial segments, and obtaining an initial power line trajectory.

[0051] First, based on the power line candidate point extraction and clustering analysis of multi-geometric feature fusion, the specific steps are as follows:

[0052] (1) Setting a neighborhood search radius of 0.5m for each point in the preprocessed point cloud, which takes into account the typical diameter of the power line and the sampling density of the laser radar. The principal component analysis (PCA) method is used to calculate the covariance matrix of the neighborhood point cloud and obtain the eigenvalues and the corresponding eigenvectors , defining the linearity , curvature and normal vector variance as geometric feature descriptions, where is the normal vector of the i-th point, is the average normal vector, is the number of field point clouds.

[0053] (2) Based on the geometric characteristics of the power line, a multi-feature judgment criterion is established. The power line has the characteristics of high linearity, low curvature and consistent normal vector. When the linearity , curvature and normal vector variance of a point are satisfied, the point is marked as a power line candidate point. This multi-feature fusion criterion can effectively identify point clouds with power line characteristics while suppressing linear interference such as building edges and tree branches.

[0054] (3) The edge filtering algorithm is executed on the candidate points. The neighborhood points within a radius of each candidate point in the original point cloud are searched, and the number of non-compliant neighbor points is counted. When the number of non-compliant neighbor points exceeds the set threshold, it is considered that the candidate point is located in the interference area such as the edge of the building, and it is removed. The setting of the edge filtering radius of 0.1m is based on the typical diameter of the power line, which can effectively detect the local environmental characteristics around the candidate point.

[0055] (4) The remaining candidate points are spatially clustered by a region growing based manual clustering algorithm, with a clustering tolerance of 0.5 m, and a KdTree is used for neighborhood search. Adjacent candidate points with a Euclidean distance less than the clustering tolerance are merged into the same cluster. Then the geometric length of each cluster is calculated, and short clusters with a length less than 1.0 m are filtered out, and the power line candidate clusters with sufficient length are retained.

[0056] Secondly, based on the power line candidate point clustering, the power line spatial segment reconstruction and connectivity analysis are performed, and the power line spatial segments are merged to obtain the initial power line trajectory, and the specific steps are as follows:

[0057] (1) The geometric attribute of each power line candidate cluster obtained in the above step is calculated, the covariance matrix of the cluster point cloud is calculated by using the principal component analysis method, and the first principal component is extracted as the main direction vector of the cluster . The cluster center is calculated, all points in the cluster are projected onto the main direction, and the start point and the end point of the cluster are determined by the extreme value of the projection value, and the cluster length is calculated.

[0058] (2) The vertical line segment filtering is performed, and the included angle between the main direction of each segment and the z-axis is calculated , wherein is the main direction of the segment, and when the included angle is less than 45°, the segment is considered to be too vertical and does not meet the horizontal distribution characteristics of the power line, and is removed from the candidate segment. The setting of the vertical line segment threshold of 45° is based on the characteristics that the power line is usually horizontally erected.

[0059] (3) The connectivity judgment mechanism between segments is established, and for any two segments and . The start point is , and the end point is ; The start point is , and the end point is . Four distance combinations between the end points are calculated: start point-start point distance , start point-end point distance , end point-start point distance , and end point-end point distance . When the minimum end point distance , it is considered that the two segments have a connection possibility in space, is the maximum distance threshold.

[0060] (4) The vertical distance component of the connection vector is further calculated, and the connection vector , wherein is the center point of the first segment, is the center point of the second segment, average direction , is the main direction of the first segment, is the main direction of the second segment, projection of the connecting vector in the average direction , vertical component , vertical distance When the vertical distance , the two segments are considered to have good collinearity, meeting the geometric constraints of power lines.

[0061] (5) A connection scoring mechanism is constructed, considering the spatial distance and direction consistency between segments.

[0062] The distance score is: , where is the main direction distance between the center points of the two segments, is the main direction distance threshold.

[0063] The direction score is: , where is the main direction of the first segment, is the main direction of the second segment, is the connecting vector connecting the center points of the two segments.

[0064] The comprehensive score is: When the score exceeds the threshold, the connection relationship is confirmed, and are the weights of the distance and direction scores.

[0065] (6) The depth-first search (DFS) algorithm is used to analyze the connectivity of the segments, and the adjacency matrix of the segment connection graph is constructed. All connected groups are found by DFS traversal, and each connected group represents a complete power line. All segment point clouds in the connected group are merged, and the total length, main direction and geometric center of the merged power line are recalculated. The spline curve fitting is performed on the merged power line, and the main direction is uniformly sampled at a resolution of 0.5m to generate a sequence of spline control points , providing a smooth geometric representation of the power line.

[0066] S3, based on the initial power line trajectory, a closed evolutionary Bayesian inference is used to construct a probability map, and the region of interest of the probability map is extracted. This step constructs a dynamic and adaptive three-dimensional probability map to estimate the existence probability of the power line in space in real time. The core idea is to integrate four probability components through the theoretically optimal information fusion framework to form the final probability state : core Bayesian probability (current optimal fusion of multiple sources of observation), time evolution probability Historical confidence intelligent decay, continuity propagation probability Physical continuity neighborhood propagation, multi-modal physical probability Reflective intensity, height, etc. non-geometric corroboration. The specific steps are as follows:

[0067] (1) Establish a three-dimensional voxelized probability map data structure that enhances observation fusion. Discretize the continuous three-dimensional space into a voxel grid, and set the voxel size to 0.1m to balance the calculation efficiency and spatial resolution. Unlike traditional probability maps, each voxel not only maintains the basic power line existence probability and confidence information, but also introduces the core elements of the closed-form evolutionary Bayesian theory. Specifically, it stores the power line existence probability, enhanced observation variables based on BCDM theory, dynamically updated confidence, last observation timestamp, and a set of power line IDs supporting independent tracking of multiple power lines. To achieve optimal information fusion, three types of observation operator matrices are constructed: spatial observation operator responsible for processing geometric position and distance information, time observation operator processing historical probability information and time continuity, geometric observation operator processing geometric constraints such as linearity, direction consistency and curvature; and define observation noise covariance, spatial observation noise: , time observation noise: , geometric observation noise: .

[0068] (2) Construct an enhanced observation variable construction framework based on causal weight closed-form conditional reasoning. This framework innovatively introduces causal adaptive weights. Traditional methods simply combine different observation sources with weights, lacking theoretical guidance, while this method achieves optimal linear combination by constructing an information weight matrix. First, calculate three types of original observation values. Spatial observation only comes from the current frame, where is the shortest distance from the voxel to the detected power line , and is the voxel position; time observation is historical inheritance, where is the final probability value of the voxel at the last time, is the voxel position; geometric observation is local point cloud feature where is the linearity of the voxel position , and is the direction consistency score of the voxel position . Calculate the information weight: where is the voxel position, t is the current time, and the weight coefficients satisfy the normalization constraint and are dynamically adjusted according to the reliability of different sensors. The core innovation lies in constructing an enhanced observation: wherein is the voxel position, is the weight of the three different observations, which is a function of the voxel position and the time t. The mathematical significance of this construction is that, by means of an optimal linear transformation, the multi-source observation information is projected into a low-dimensional space, so that the complex multi-dimensional conditional expectation calculation is transformed into a simple one-dimensional problem, satisfying the important conditional expectation equivalence property: wherein is the spatial observation, the temporal observation, and the geometric observation above, represents the true existence state of the power line at the position . This property mathematically guarantees that the single enhanced observation constructed by us has been losslessly compressed and contains all the useful information from . Therefore, in the subsequent update, we use this simplified single to obtain the optimal estimate, which is completely equivalent to the optimal estimate obtained by simultaneously processing the three complex observation sources. This makes the complex multi-source fusion problem be transformed into a standard single-source update problem, thereby simplifying the implementation of the filter without losing information. Wherein, , , wherein is the voxel position, t is the time, and the confusion function is introduced to adaptively and dynamically adjust the weights. Wherein is the normalization factor to ensure that the sum of the weights is 1, is the confusion sensitivity parameter (fixed value), is the hyperbolic tangent function. Wherein the spatial confusion function evaluates the reliability of the spatial observation by calculating the degree of deviation between the actual spatial observation value of the current frame and the expected observation value inferred based on the historical geometric state; the temporal confusion function detects the abnormal jump of the historical probability by comparing the difference between the actual change rate of the probability value at the current time and the expected change rate predicted based on the historical trend; and the geometric confusion function evaluates the reliability of the geometric observation by comparing the consistency degree between the actual observed direction angle of the power line in the current frame and the expected direction angle predicted based on the kinematic model.

[0069] (3) Closed-form conditional probability computation with intelligent correction and spatio-temporal coupled evolution mechanism. Traditional Bayesian update methods usually assume simple Gaussian distribution or adopt numerical integration method, which has high computational complexity and lacks theoretical guarantee. This method establishes a closed-form expression of the conditional posterior probability, which accurately models the nonlinear probability distribution through the combination of sigmoid function and quadratic form: , is the standard posterior probability, where the parameter is learned by minimizing the conditional variance criterion, is the bias term, is the linear weight term, which captures the linear relationship between the augmented observation and the probability, is the quadratic interaction term, which models the nonlinear effect of the augmented observation and ensures theoretical optimality. In terms of spatio-temporal evolution, we abandon the traditional simple exponential decay model and establish a more accurate spatio-temporal coupled evolution dynamics. We introduce an adaptive evolution kernel function, and the decay rate is no longer a constant but is dynamically adjusted according to the confidence and historical consistency of the augmented observation variable. When the observation quality is high and the historical consistency is strong, the decay rate is small, and the probability remains stable. Conversely, it quickly decays to converge to the equilibrium state. We also introduce a lightweight correction factor: where the correction factor is , where is the correction strength parameter (fixed value), is the confusion function defined in step (2), is the core Bayesian probability after causal correction. When confusion factors are detected, the system will make a moderate downward correction to the probability estimate according to the degree of confusion and the strength of the observation.

[0070] Traditional probability maps use a simple exponential decay model to update the probability of unobserved regions. This method establishes an adaptive spatio-temporal coupled evolution dynamics, discards the fixed decay rate, and introduces an adaptive evolution kernel function. The probability evolution formula is: where is the observation interval time, is the final probability of the last observation in the voxel, is the voxel position, will be defined in step (6). The adaptive decay rate where is the historical average augmented observation, is the current augmented observation variable calculated in step (2).

[0071] (4) Establish a probabilistic propagation and constraint mechanism for the independent evolution of multiple lines. Considering that there are often multiple parallel or intersecting power lines in actual scenarios, a probabilistic evolution framework that supports independent tracking of multiple lines needs to be established. For each power line, maintain an independent state space, including the true existence state, enhanced observation variables, time-varying probabilities, etc. To avoid interference between the probability estimates of different power lines, introduce inter-line coupling constraints: the sum of the existence probabilities of all power lines at the same location should not exceed the physical upper limit, and the coupling evolution is realized through Softmax normalization. In addition, use the physical continuity constraint of the power line for probability propagation. When the probability of a certain location is high, it will propagate probability information to the adjacent positions along the direction of the power line, and the propagation strength is determined by the distance decay function and the direction consistency. For the i-th power line, when the core Bayesian probability of the voxel at position is high, the probability information is propagated to the adjacent positions along the direction of the power line: where is the distance between two points, is the core Bayesian probability of the voxel at position p in power line i, is the propagation scale (fixed value), is the connection vector direction, is the main direction of the power line. The final probability is the maximum propagation value from all possible propagation source positions and power lines. The propagation strength is determined by the distance decay and the direction consistency.

[0072] (5) Construct a Bayesian fusion framework for multi-modal enhanced observations. Power line detection involves multiple physical phenomena and geometric features, and single-mode observation is often insufficient to accurately determine the existence of power lines. This method independently calculates the observation components of multiple modalities and then optimally fuses them through a Bayesian framework. The geometric observation component is obtained by calculating the linearity and direction consistency of the local point cloud, the reflection observation component is based on the reflection intensity information of the laser radar, the height observation component considers the prior knowledge that power lines are usually suspended within a certain height range, and the environmental observation component considers the spatial accessibility and topological consistency. Introduce non-geometric physical priors to fuse the geometric, reflective, height, and environmental four modal observations: . The geometric modal is multiplexed ; the reflection modal is based on laser intensity (typical range ), where is the average reflection intensity of the point cloud in the voxel, is the typical reflection intensity range of the power line (fixed value), is the indicator function. The height modal is based on the typical erection height (Gaussian distribution) where is the height coordinate of the voxel,​​ is the typical erection height (fixed value), is the height variance (fixed value). The spatial accessibility and topological consistency are considered, wherein is the spatial accessibility score (judged by ray casting), is the topological consistency score (judged whether located on the known electric tower connecting path). The multi-modal Bayesian fusion ensures that the detection results conform to the engineering layout rules of the power system, and the weight coefficients are determined by experiments.

[0073] (6) Implement an environment-adaptive prior updating mechanism. Traditional methods usually use fixed prior probability distribution, ignoring the differences in power line distribution density in different regions and time periods. This method establishes an environment-adaptive prior updating mechanism, dynamically adjusts the prior distribution according to historical observation data and regional characteristics. Define the equilibrium probability (i.e. in step (3)) as: . Wherein is the base prior (fixed value), the spatial factor , is the Euclidean distance from the voxel at position to the nearest electric tower, is the influence radius; the time factor considers the observation quality, is the observation quality score, which is learned from historical data by Gaussian mixture model. Specifically, the detection success rate, point cloud density, reflection intensity variance, etc. in different periods (morning, noon, evening, night) are calculated, and the time-related quality curve is fitted. Considering the time period difference of observation quality, the detection success rate, point cloud density, reflection intensity variance, etc. in different periods (morning, noon, evening, night) are learned from historical data by Gaussian mixture model, and the prior probability is improved in the time period with good observation quality. The adaptive updating mechanism integrates these regional and temporal prior information into the closed Bayesian inference framework, replacing the original fixed prior probability, so that the entire system can adapt to different environmental conditions and application scenarios.

[0074] (7) Perform multi-level integration and confidence quantification of the final probability state. The different probability components obtained from the previous steps are weighted and integrated to form the final voxel probability state. The integration formula is: where each component comes from closed conditional inference, spatio-temporal coupling evolution, continuity constraint propagation, and multi-modal fusion. The weight coefficients are optimized according to the reliability of each component and environmental conditions: wherein is the temperature parameter (fixed value), is the reliability score of the i-th component, the reliability score of the jth component, are the same set of scores, which all represent one of the four components of the reliability score, The calculation of each is: is the voxel confidence, is the observation interval time, is the average core Bayesian probability of the neighborhood point cloud, is the observation quality factor, which is the same as defined in step (6). In addition to the probability value, the system also needs to quantify the confidence of each probability estimate to provide uncertainty information for subsequent decision-making. The confidence calculation considers multiple factors such as the strength of the enhanced observation variable, temporal consistency, spatial continuity, etc., and obtains a comprehensive confidence index by weighted combination (observation of power line), where is the voxel position, is the confidence of the last voxel, is the confusion function of step (2); when the voxel is not observed, the confidence decays exponentially: is the confidence decay rate (fixed value), is the observation interval time.

[0075] (8) Intelligent Region of Interest (ROI) extraction based on multi-level probability map. Intelligent Region of Interest extraction using the updated probability map is one of the important applications of probability map. First, a high-probability voxel set is obtained by probability threshold screening, and the threshold is set to 0.7 to balance detection accuracy and recall rate. Then, a connectivity clustering algorithm based on graph theory is used to merge voxels that are spatially adjacent and have similar probabilities into connected regions. The adjacency condition for clustering not only considers spatial distance, but also considers probability difference to ensure semantic consistency of the clustering result. For each connected component, an Axis Aligned Bounding Box (AABB) is calculated, which is calculated by traversing all points in a connected component to find the minimum and maximum values of the X, Y, and Z coordinate axes respectively. These six extreme coordinates define two opposite vertices of a cuboid that can just wrap all points, with each side parallel to the coordinate axis. To handle the possible overlap of bounding boxes, an overlap detection and merging mechanism is introduced, and finally the intelligent region of interest is generated as the input of the subsequent fine detection and tracking algorithm.

[0076] ​​​​S4, initializing the MaxUKF tracker based on the initial power line trajectory and the probability map, using the initialized MaxUKF tracker to predict the trajectory of the point cloud data of the subsequent frame, updating the probability map, and using the region of interest to crop the point cloud data of the subsequent frame, performing the local detection of S2 on the cropped point cloud data, completing the predicted trajectory according to the local detection result, and outputting the complete power line. The specific steps are as follows:

[0077] (1) Establish an enhanced 15-dimensional power line state vector and a MaxUKF parameter model. Traditional Kalman filters usually only consider the kinematic state of the target, such as position, velocity, etc., which is not enough for complex power line tracking tasks. This method designs a 15-dimensional enhanced state vector to unify the geometric state and the probability state of the power line. The geometric state contains 11 dimensions: the three-dimensional coordinates of the starting point and the ending point for describing the spatial position of the power line, the normalized principal direction unit vector representing the orientation of the power line, the geometric length and the average curvature respectively describing the scale and shape features of the power line. The probability state contains 4 dimensions: the average probability reflecting the average support of the probability map on the trajectory path, the maximum probability indicating the highest probability value on the path, the confidence quantifying the reliability of state estimation, and the observation quality evaluating the pros and cons of the current observation condition. The complete state vector is: The state transition equation is: , MaxUKF parameter design, where is the state transition matrix (block diagonal structure), is the complete state vector at the previous time, is the probability guide matrix, is the probability map correction vector, is the process noise. is the process noise covariance, which is self-adaptively adjusted according to the confidence: , where is the gain coefficient, is the basic process noise covariance value.

[0078] (2) Instantiate the probabilistic map guided MaxUKF tracker. Whenever a new power line is detected, the system creates a tracker instance. Instantiation includes state initialization, covariance setup, and trajectory attribute configuration. When initializing the state vector, the geometric state is directly extracted from the S2 detection result, and the probabilistic state is obtained from the S3 query probabilistic map. The trajectory state management adopts a finite state machine model, defining four basic states: the TENTATIVE state represents a newly created trajectory that needs to be continuously observed to verify its authenticity. The CONFIRMED state represents a stable tracking trajectory that has been verified. The PREDICTED state represents a trajectory that is not detected in the current frame but is maintained based on historical information. The COMPLETED state represents a trajectory that is intelligently completed based on probabilistic map information. Each trajectory instance also maintains rich attribute information, including confidence score, number of successful matches, number of consecutive losses, probabilistic map association index, etc. The initialized state, covariance, and trajectory attributes are output and passed to step (3) for prediction.

[0079] (3) Perform the probabilistic map guided MaxUKF prediction step. For the dynamic model defined in step (1) and the probabilistic map evolution information constructed in step S3, predict the state distribution at the next time step. Unlike the standard UKF, this method uses Gaussian sum to represent the posterior distribution, which can model multi-modal uncertainty. MaxUKF is an extension of the traditional unscented Kalman filter, specifically designed to handle non-Gaussian state estimation problems. The core of the prediction step is to propagate the probability distribution through sigma point sampling technology. Unlike the standard UKF, MaxUKF needs to handle probability density functions in Gaussian sum form: where and are the system state vectors, is all the observation history we already have, from the 1st frame to the k-1th frame observation vector, is the weight and the sum of all Gaussian component weights is 1, is the predicted mean, is the covariance, is the number of Gaussian components. In step (5), the Gaussian components will be split out. For each Gaussian component, a set of representative sampling points, called sigma points , is generated using a deterministic sampling strategy. These sigma points can accurately capture the mean and covariance information of the state distribution. The center point is set at the component mean, and the spread points are distributed along the principal directions of the covariance matrix. The number of sampling points depends on the state dimension. Each sigma point is equipped with a corresponding weight for subsequent mean and covariance calculation. State prediction is to propagate these sigma points through the nonlinear state transition equation, which is divided into two parallel channels: based on the kinematics model of the power line (e.g. constant velocity model) considering the physical motion characteristics of power lines; the probability state prediction is updated according to the evolution law of the probability map, specifically by querying the probability value on the predicted trajectory path. The system discretizes the predicted trajectory into a series of points, queries the probability value of each point in the probability map, and calculates the average probability and the maximum probability as the predicted probability state. After prediction, the predicted mean and covariance of each Gaussian component are calculated by the weighted statistics of the sigma points after propagation, and the predicted Gaussian sum distribution is output for subsequent steps. It is passed to step (5) for splitting enhancement, and also provides the predicted state for step (4) for observation fusion.

[0080] (4) Establish an enhanced measurement model for multi-source observation fusion. One of the challenges of power line tracking is how to effectively fuse observation information from different sources. This method designs a unified enhanced measurement model to optimally combine traditional geometric observations and newly introduced probability observations. The geometric observation equation is: where is the geometric observation vector, is the geometric observation matrix, which only observes the geometric state, is the measurement noise, is the measurement noise covariance, is the complete state vector, is the geometric observation vector. The probability observation equation is: , the probability observation vector contains statistical quantities such as probability distribution along the trajectory path, maximum probability position, probability consistency score, and spatial coverage, is the probability measurement noise, is the measurement noise covariance, is the probability observation matrix. To achieve optimal fusion, we draw on the idea of closed-form evolution Bayesian theory to construct an information weight matrix: where , is the information fusion weight (determined according to the confidence of each source) where is the geometric observation confidence, is the probability observation confidence. Then the enhanced observation variable is obtained: This construction satisfies the condition of expectation equivalence: This property mathematically guarantees that the single enhanced observation we constructed has been losslessly compressed and contains information from both geometric observation and probability observation ​all the useful information. Therefore, in the following update, we use this simplified single The resulting optimal estimation is equivalent to that of processing both complex observation sources simultaneously. This construction theoretically guarantees the optimality of information fusion, minimizing the conditional variance of the fused observation. This makes the complex multi-source fusion problem transformed into a standard single-source update problem, thus simplifying the implementation of the filter without losing information. Augmented observation and its equivalent observation matrix are passed to the measurement update in step (6), while providing the observation intensity information for split decision in step (5).

[0081] (5) Implement maximum point prediction and adaptive Gaussian split. The core innovation of MaxUKF lies in the introduction of the maximum point concept to accurately model the multi-modal characteristics of non-Gaussian distribution. Maximum point prediction includes geometric maximum points and probability maximum points, which are obtained through kinematic model and probability map query respectively, and then fused by weighted fusion: where is the peak position of the state probability density function, is the geometric maximum point obtained by propagation through the kinematic model, is the probability maximum point obtained by querying the probability map, is the fusion weight, where is the geometric observation confidence, is the probability observation confidence. For each prior Gaussian component output by step (3), calculate the normalized deviation of the maximum point relative to the component mean , is the prediction covariance matrix of the jth Gaussian component in step (3), is the prediction mean of the jth Gaussian component in step (3). Determine whether each dimension needs to be split: if the absolute value of the normalized deviation is greater than or equal to 1.0 (fixed threshold), the dimension triggers splitting. For dimensions that need to be split, solve the intermediate parameters satisfy: where is the normalized deviation of the jth Gaussian component. After solving the intermediate parameters, construct the parameters of the two sub-Gaussian components, including their respective weights, means and variances. Map back to the original state space through inverse normalization transformation to generate new 15-dimensional Gaussian components. If multiple dimensions need to be split, combine to generate multiple sub-components. After adaptive splitting, the original Gaussian component is expanded into a refined prior distribution where is the system state vector, is all the observation history we already have, from the 1st frame to the k-1th frame observation vector, weight of the jth component, predicted mean, covariance, split number, which can accurately express the complex non-Gaussian and multi-modal characteristics and significantly improve the state estimation accuracy in the occlusion scene. The maximum point of the prior is output as the initialization reference for subsequent update. and the maximum point of the prior are passed to step (6) for measurement update.

[0082] (6) Implement the probability weight MaxUKF data association and measurement update. Realize the matching association of the current frame detection and historical tracking, and perform measurement update based on the refined prior distribution output in step (5). A multi-dimensional similarity measurement framework is established, considering factors such as spatial distance, direction consistency, probability support, and statistical consistency. The similarity calculation formula is: where each component is the spatial similarity calculated based on the end-point distance; the direction similarity calculated based on the dot product of the principal direction vector; the probability similarity calculated by querying the probability map to calculate the average probability support of the trajectory path; and the uncertainty similarity measured based on the Mahalanobis distance to measure whether the observation falls within the predicted uncertainty range.

[0083] ​​A cost matrix is constructed with elements of 1 minus the integrated similarity, the higher the similarity the lower the cost. The Hungarian algorithm is used to solve the optimal bipartite matching, with constraints that each detection is matched to at most one track and each track is associated with at most one detection. Three types of results are obtained: a set of matched pairs representing successfully associated detection-track pairs, a set of unmatched tracks possibly due to occlusion, and a set of unmatched detections possibly new power lines or false detections. A similarity threshold is set, and pairs below the threshold are forced to be unmatched even if the algorithm assigns them, to avoid false associations. Then the Gaussian sum measurement update is performed, which is the core step of the system. The key is to perform the UKF measurement update independently for each prior component output from step (5). For each component, the measurement Sigma points are generated and propagated, and the predicted measurement, measurement covariance, and cross-covariance are calculated. The state update and covariance update are performed through the Kalman gain. The weight update is based on the likelihood function: the closer the observation is to the component prediction, the larger the component weight, reflecting the relative likelihood of different hypothesis modalities. Normalization ensures that the weights sum to 1. After updating, low-weight components are removed through weight truncation. After all components are updated, the posterior maximum point is solved through the gradient ascent method, which is the peak position of the posterior Gaussian sum probability density function. The initial value is set to the posterior weighted mean, the step size is 0.5 and is adjusted adaptively according to the Armijo criterion, and the convergence criterion is that the step size is less than 0.01 or the maximum number of iterations is 20.

[0084] (6) Perform the probability-guided multi-hypothesis intelligent completion algorithm. Determine whether to start the completion mechanism for the unmatched track. The triggering conditions include multiple checks: the track state must be PREDICTED, indicating that it has been confirmed in the previous frame but is not detected in the current frame; the number of consecutive missing frames must not exceed the set threshold, to avoid false completion of long-term missing tracks; the historical confidence must be sufficient; and the average probability of the predicted path on the probability map must exceed the threshold, to verify that the predicted path is consistent with the historical information. Tracks that meet all conditions enter the completion process, otherwise they directly enter the lifecycle management. For tracks that meet the completion conditions, three different completion hypotheses are generated for comprehensive decision-making. Tracks that meet the completion conditions enter the completion process, otherwise they directly enter the lifecycle management in step (8). The completion conditions include track state checking, continuous loss limitation, confidence threshold, and probability support degree evaluation. The system generates three different completion hypotheses for comprehensive decision-making. Hypothesis 1, is a probability-guided path, which searches for a high-probability path from the predicted starting point to the target endpoint based on the probability map using the A* algorithm. The cost function includes two terms: the negative probability, which encourages travel along high-probability areas, and the Euclidean distance, which ensures target orientation. Hypothesis 2, is a physically constrained path, which is based on the catenary mechanics model of power lines, considering the effects of gravity and typical wind forces, and generates the path by numerically integrating the differential equation. Hypothesis 3, For environmental avoidance, obstacle regions are identified in the environmental point cloud, and a B-spline curve is used to generate a smooth path that avoids obstacles, with the optimization goal being to minimize curvature variation and directional deviation. Three hypothetical paths are weighted and combined to generate the final completed trajectory. The final completion result is obtained by weighted combination: wherein are the three hypothetical paths, are the weights of each hypothesis, is the final completed trajectory, and a Catmull-Rom spline interpolation is used to generate a smooth power line trajectory based on the path quality evaluation, ensuring geometric continuity and physical rationality of the completion result. The completion confidence is calculated, which comprehensively reflects the reliability of the three hypotheses. If the completion confidence exceeds the acceptance threshold, the completion result is accepted and the trajectory state is marked as COMPLETED; otherwise, the completion is abandoned and the trajectory remains in the PREDICTED state, waiting for subsequent observations or deletion. The completed trajectory is accompanied by confidence information, which is used for subsequent probability map update weight adjustment.

[0085] (7) Perform trajectory lifecycle management and bidirectional probability map update. Trajectory lifecycle management uses a state machine-based design, defining clear state transition rules and condition judgment logic. The TENTATIVE to CONFIRMED transition requires a threshold number of consecutive successful matches and a confidence level above a threshold. The CONFIRMED to PREDICTED transition occurs when the current frame does not match but the confidence is sufficient. The PREDICTED to COMPLETED transition requires that the number of consecutive lost frames is within the tolerance range, the historical confidence is high, and step 7 is successful. The transition condition from any state to deletion is that the consecutive loss exceeds the maximum allowed number of frames or the confidence falls below the deletion threshold, indicating that the trajectory is not recoverable. The confidence dynamic update strategy includes: increasing the amount of successful matching is proportional to the similarity, with an upper limit of 1; exponentially decaying when matching fails, with a lower limit of 0. Confidence quantifies the historical reliability of the trajectory, affecting the uncertainty of subsequent prediction and the completion trigger condition. The bidirectional probability map update mechanism realizes the information loop between the tracking system and the probability map: for successfully matched trajectories, the probability of the corresponding region is enhanced, and the probability update formula is: for probability enhancement, wherein is the voxel position, is the updated probability of the corresponding voxel, is the probability before updating the corresponding voxel, is the learning rate, is the reliability evaluation parameter, determined by confidence and similarity; is the distance weight function, based on the distance from the point to the trajectory; for unmatched paths for probability decay, wherein is the updated probability of the corresponding voxel, is a probability before a voxel is updated, is a voxel position, is an equilibrium probability, is a time interval, is an adaptive decay rate, which is adaptively adjusted according to an observation deviation. Finally, the system outputs a complete power line list including all power lines detected, predicted and completed, and each trajectory is provided with detailed confidence quantification information, which comprehensively considers geometric consistency, probability support, time continuity and other factors, thereby providing reliable decision basis for subsequent applications.

[0086] As shown in the multi-geometry feature power line extraction effect, Figure 2 As shown in the multi-geometry feature power line extraction effect, Figure 2 In (a), the original point cloud, Figure 2 In (b), the candidate point extraction result, Figure 2 In (c), the final clustering result, the algorithm effectively identifies the power line feature points and filters the building edges and other interference. As shown in the power line segment reconstruction effect, Figure 3 As shown in the power line segment reconstruction effect, Figure 4 As shown in the power line segment reconstruction effect, Figure 4 In (a), the closed evolutionary probability map construction effect is shown, different colors represent different levels of power line existence probability, Figure 4 In (b), the intelligent ROI bounding box accurately encloses the high-probability area. As shown in the MaxUKF non-Gaussian tracking and intelligent completion effect, Figure 5 The green trajectory represents the detection result, and the red trajectory represents the intelligent completion result. The system can successfully complete the power line trajectory during the occlusion period. The experimental results verify the effectiveness and practicability of the method, which significantly improves the processing efficiency and intelligence level while ensuring the detection accuracy, thereby providing reliable technical support for real-time monitoring of power lines.

[0087] Embodiment 2

[0088] The embodiment provides a power line detection system based on a closed evolutionary Bayesian probability map, comprising:

[0089] The data acquisition and preprocessing module is configured to acquire laser radar three-dimensional point cloud data, and perform frame processing on the point cloud data;

[0090] The initial power line trajectory construction module is configured to extract power line candidate points of the first frame of point cloud data, cluster the candidate points, perform power line spatial segment reconstruction and connectivity analysis based on the power line candidate point clustering, and merge the power line spatial segments to obtain the initial power line trajectory;

[0091] The probability map construction module is configured to: based on the initial power line trajectory, construct a probability map by using closed-form evolutionary Bayesian inference, and extract a region of interest of the probability map;

[0092] The power line detection module is configured to: initialize a MaxUKF tracker based on the initial power line trajectory and the probability map, use the initialized MaxUKF tracker to predict a trajectory of point cloud data of a subsequent frame, update the probability map, use the region of interest to crop the point cloud data of the subsequent frame, perform local detection on the cropped point cloud data, complete the predicted trajectory according to a local detection result, and output a complete power line.

[0093] It should be noted that the above modules correspond to the steps in Embodiment 1, and the above modules have the same examples and application scenarios as the steps to which the above modules correspond, but are not limited to the content disclosed in Embodiment 1. It should be noted that the above modules can be executed in a computer system as part of the system.

[0094] In more embodiments, there are also provided:

[0095] An electronic device includes a memory and a processor, and computer instructions stored in the memory and running on the processor, when the computer instructions are run by the processor, the method in Embodiment 1 is completed. For brevity, it will not be described here.

[0096] It should be understood that in the embodiments, the processor can be a central processing unit CPU, and the processor can also be other general-purpose processors, digital signal processors DSPs, application-specific integrated circuits ASICs, field programmable gate arrays FPGAs, or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0097] A computer-readable storage medium for storing computer instructions, when the computer instructions are executed by a processor, the method in Embodiment 1 is completed.

[0098] The method in Embodiment 1 can be directly executed by a hardware processor, or executed by a combination of hardware and software modules in the processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, or electrically erasable programmable memory, registers, etc. The storage medium is located in the memory, and the processor reads the information in the memory, and combines the hardware to complete the steps of the above method. To avoid repetition, it will not be described in detail here.

[0099] A computer program product includes a computer program, when the computer program is executed by a processor, the method in Embodiment 1 is implemented.

[0100] The present application also provides at least one computer program product tangibly stored on a non-transitory computer readable storage medium. The computer program product includes computer executable instructions, for example, instructions embodied in program modules, executed by devices at the target real or virtual processor to perform the processes / methods as described above. Generally, program modules include routines, programs, libraries, objects, classes, components, data structures, etc. that perform particular tasks or implement particular abstract data types. In various embodiments, the functionality of program modules can be combined or split between program modules as desired. Machine executable instructions for program modules can be executed within a local or distributed device. In a distributed device, program modules can be located in local and remote memory storage devices.

[0101] Computer program code for carrying out operations of the present application can be written in one or more programming languages. These computer program code can be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the program code, which executes via the processor of the computer or other programmable data processing apparatus, transforms the computer or other programmable data processing apparatus into a particular machine for executing the described functions / operations in the flowcharts and / or block diagrams. The program code can be executed entirely on a computer, partially on a computer, as a stand-alone software package, partially on a computer and partially on a remote computer or entirely on a remote computer or server.

[0102] In the context of the present application, the computer program code or related data can be carried by any suitable carrier to enable the device, apparatus or processor to perform the various processes and operations described above. Examples of carriers include signals, computer readable media, and the like. Examples of signals can include electrical, optical, radio, sound or other forms of propagated signals, such as carrier waves, infrared signals, and the like.

[0103] Those skilled in the art can realize that the units and algorithm steps of the examples described in conjunction with the present embodiments can be realized in electronic hardware or a combination of computer software and electronic hardware. Whether the functions are performed in hardware or software depends on the specific application and design constraints of the technical solutions. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0104] The above describes the specific embodiments of the present application in conjunction with the accompanying drawings, but is not a limitation on the scope of protection of the present application. Those skilled in the art should understand that various modifications or variations made by those skilled in the art on the basis of the technical solutions of the present application without inventive labor are still within the scope of protection of the present application.

Claims

1. A power line detection method based on a closed-loop evolutionary Bayesian probabilistic map, characterized in that, Includes the following steps: S1. Acquire 3D point cloud data from the LiDAR and perform frame-by-frame processing on the point cloud data; S2. Extract candidate power line points from the first frame of point cloud data, cluster the candidate points, reconstruct and analyze the spatial segments of power lines based on the clustering of candidate power line points, merge the spatial segments of power lines, and obtain the initial power line trajectory. S3. Based on the initial electric field line trajectory, a probabilistic map is constructed using closed-loop evolutionary Bayesian inference, and the region of interest in the probabilistic map is extracted. A probability map is constructed using closed-loop evolutionary Bayesian inference, specifically as follows: Construct multimodal augmented observation variables, calculate information weight matrix based on causal adaptive weights, perform optimal linear combination of multimodal augmented observation variables based on information weight matrix, replace traditional exponential decay with spatiotemporal coupled evolution dynamics, model the nonlinear probability distribution of observation variables, and introduce inter-line coupling constraints to obtain the final probability map. S4. Initialize the MaxUKF tracker based on the initial power line trajectory and probability map. Use the initialized MaxUKF tracker to predict the trajectory of the point cloud data in subsequent frames, update the probability map, and use the region of interest to crop the point cloud data of subsequent frames. Perform local detection on the cropped point cloud data in S2. Complete the predicted trajectory based on the local detection results and output the complete power line. The step of using the initialized MaxUKF tracker to predict the trajectory of point cloud data in subsequent frames specifically involves: An enhanced state vector is designed to unify the geometric and probabilistic states of the power line. Trajectory state management employs a finite state machine model, defining four basic states: newly created trajectory (TENTATIVE), validated stable tracking trajectory (CONFIRMED), trajectory not detected in the current frame but maintained based on historical information prediction (PREDICTED), and trajectory intelligently completed based on probabilistic map information (COMPLETED). A Gaussian sum is used to represent the posterior distribution, and the probability distribution is propagated through sigma point sampling technology. An information weight matrix is ​​constructed to achieve optimal fusion of geometric and probabilistic observations, resulting in enhanced observation variables. Maximum point prediction and adaptive Gaussian splitting are introduced to accurately model the multimodal characteristics of non-Gaussian distributions. The Hungarian algorithm is used to solve for optimal bipartite graph matching, achieving matching association between current frame detection and historical tracking. For unmatched trajectories that meet preset conditions, three completion hypotheses—probabilistic guided path, physical constraint path, and environmental avoidance path—are generated and weighted to generate the final completed trajectory. Trajectory lifecycle management and bidirectional updates of the probabilistic map are executed to achieve an information closed loop between the tracking system and the probabilistic map.

2. The electric field line detection method based on a closed-loop evolutionary Bayesian probabilistic map as described in claim 1, characterized in that, It also includes establishing an environment-adaptive prior update mechanism to dynamically adjust the prior distribution based on historical observation data and regional characteristics.

3. The electric field line detection method based on a closed-loop evolutionary Bayesian probabilistic map as described in claim 1, characterized in that, The weight coefficients of the information weight matrix satisfy the normalization constraint and are dynamically adjusted according to the reliability of different observation sensors.

4. The electric field detection method based on a closed-loop evolutionary Bayesian probabilistic map as described in claim 1, characterized in that, The line coupling constraint is that the sum of the probabilities of all power lines existing at the same location should not exceed the physical upper limit, and the coupling evolution is achieved through Softmax normalization.

5. The electric field detection method based on a closed-loop evolutionary Bayesian probabilistic map as described in claim 1, characterized in that, The enhanced observation variables in the multimodal study include geometric observation components, reflection observation components, and altitude observation components.

6. A power line detection system based on a closed-loop evolutionary Bayesian probabilistic map, employing the power line detection method based on a closed-loop evolutionary Bayesian probabilistic map as described in any one of claims 1-5, characterized in that, include: The data acquisition and preprocessing module is configured to: acquire 3D point cloud data from the LiDAR and perform frame-segmentation processing on the point cloud data; The initial power line trajectory construction module is configured to: extract candidate power line points from the first frame of point cloud data, cluster the candidate points, reconstruct and analyze the power line spatial segments based on the power line candidate point clustering, merge the power line spatial segments, and obtain the initial power line trajectory. The probabilistic map construction module is configured to: construct a probabilistic map based on the initial electric field line trajectory using closed-form evolutionary Bayesian inference, and extract the region of interest from the probabilistic map; The power line detection module is configured to: initialize a MaxUKF tracker based on the initial power line trajectory and probability map; use the initialized MaxUKF tracker to predict the trajectory of point cloud data in subsequent frames; update the probability map; crop the point cloud data of subsequent frames using regions of interest; perform local detection on the cropped point cloud data; complete the predicted trajectory based on the local detection results; and output the complete power line.

7. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the method according to any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, perform the method described in any one of claims 1-5.

9. A computer program product, characterized in that, Includes a computer program, which, when executed by a processor, implements the method described in any one of claims 1-5.

Images