Laser radar point cloud mapping method for power distribution line scanning in dynamic environment

Abstract

The application relates to the field of laser SLAM mapping of power scenes, and discloses a laser radar point cloud mapping method for power distribution line scanning in a dynamic environment. The method first extracts a segmented curvature voxel occupancy rate descriptor; then uses a voxelized generalized iterative closest point algorithm based on curvature voxel neighborhood search to obtain a current pose of a current frame by optimizing a pose residual error of associated voxels; then uses a curvature voxel clustering method based on laser intensity assistance to perform object segmentation on the point cloud of the current frame, label absolute static objects and potential dynamic objects, retain low dynamic objects and remove high dynamic objects; again uses the voxelized generalized iterative closest point algorithm based on curvature voxel neighborhood search to perform secondary optimization on the current frame pose; and finally performs pose transformation and splicing frame by frame to obtain a global static point cloud map after removing the high dynamic objects. The application can be used for handheld point cloud scanning modeling of scenes containing power distribution lines in a dynamic environment.

Description

Technical Field

[0001] This invention relates to the field of laser SLAM mapping in power scenarios, specifically to a laser radar point cloud mapping method for power line scanning in dynamic environments. Background Technology

[0002] Real-time localization and mapping (SLAM) is a technology that enables robots to perceive their surroundings in unknown environments using visual sensors, estimate the sensor poses during movement, and simultaneously perform self-localization based on a map and build incremental maps based on the localization results.

[0003] With the widespread application of SLAM technology, using handheld mobile 3D LiDAR scanning technology for scanning and modeling power scenarios such as urban power distribution lines offers advantages such as operational flexibility and low cost. Point cloud modeling of urban power distribution line environments is crucial for power line inspection and facility status assessment. However, urban power distribution lines contain numerous plants and buildings in addition to power poles and wires, and are often located alongside roads with vehicular traffic. Therefore, to generate accurate power scene point cloud models, addressing environmental scanning and modeling in dynamic environments with vehicles is key to improving mapping accuracy.

[0004] For the problem of point cloud instance segmentation, most algorithms increasingly use convolutional neural networks (CNNs) for point cloud segmentation and semantic annotation. Distance-based image-based point cloud instance segmentation methods directly apply modified image instance segmentation methods, such as Mask R-CNN (see "Segmenting unknown 3d objects from real depth images using mask r-cnn trained on synthetic data", Danielczuk et al.), which uses a multi-stage segmentation approach. Meanwhile, considering the 3D representation of point clouds, the 3D convolutional neural network PointNet (see "Pointnet: Deep learning on point sets for 3d classification and segmentation", Qi et al.) has emerged. However, deep learning methods require costly training with accurate labels. In the power industry, where point cloud annotations are lacking, pre-trained models have poor generalization ability for recognizing power equipment. Furthermore, the construction of point cloud models for power industry scenarios emphasizes real-time scanning and mapping by mobile scanning devices, which traditional point cloud segmentation networks struggle to meet diverse needs and real-time requirements.

[0005] To address the problem of removing dynamic objects, existing methods can effectively remove most dynamic objects by specifically representing dynamic information. Point and pixel cascade removal methods based on distance images, such as Removert (see "Remove, then revert: Static point cloud map construction using multiresolution range images", Kim et al.), employ multi-resolution pixel difference detection. Ray tracing methods based on occupancy grids, such as ERASOR (see "ERASOR: Egocentric ratio of pseudo occupancy-based dynamic object removal for static 3D point cloud map building", Lim et al.), design a pseudo-center spatial occupancy descriptor. However, existing methods ignore the object properties of dynamic information and are not robust to problems such as scan inconsistencies, leading to the erroneous removal of a large number of static points. This poses a challenge for the preservation of delicate components in power equipment, especially wires and insulator strings.

[0006] Currently, there are no domestic or international patents that can simultaneously integrate point cloud instance segmentation and dynamic object removal for point cloud scanning and modeling needs in power scenarios. Summary of the Invention

[0007] Purpose of the invention: To address the above-mentioned problems, a lidar point cloud mapping method for power distribution line scanning in dynamic environments is provided.

[0008] Technical Solution: This invention proposes a lidar point cloud mapping method for power line scanning in dynamic environments, which includes the following steps:

[0009] Step 1: Use a handheld multi-line lidar to scan the power distribution lines on the street and obtain the current frame of the original lidar point cloud as the current frame. Construct a segmentation curvature voxel occupancy rate descriptor in it.

[0010] Step 2: Using the voxelization generalized iterative nearest point algorithm based on curvature voxel nearest neighbor search, the pose of the current frame is obtained by optimizing the pose residual of the associated voxel registration.

[0011] Step 3: Use the laser intensity-assisted curvature voxel clustering method to segment the point cloud of the current frame into objects, and then use the geometric feature-based object classification method to identify the segmented objects and label the absolutely static objects.

[0012] Step 4: Using the segmented curvature voxel occupancy rate descriptor of adjacent frames, perform curvature voxel alignment on adjacent frames, track potential dynamic objects, and then detect changes in curvature voxel occupancy rate to retain low-dynamic objects and remove high-dynamic objects.

[0013] Step 5: The pose residual of static curvature voxel registration in the reloaded curvature voxel occupancy descriptor is optimized again using the voxelization generalized iterative nearest point algorithm based on curvature voxel nearest neighbor search to obtain the current pose of the current frame after optimization.

[0014] Step 6: Based on the optimized pose, perform pose transformation frame by frame and stitch the images together to obtain a global static point cloud map after removing highly dynamic objects. Then, use the spatial distribution geometric features of power lines and towers to extract power lines and towers from the point cloud.

[0015] Furthermore, the segmented curvature voxel occupancy rate descriptor constructed in step 1 encodes the semantics and position of objects in the environment into segmented curvature voxels, as detailed below:

[0016] The preprocessed point cloud is encoded into a set of curvature voxels at a fixed resolution in three projection directions. Curvature voxels containing at least one point are stored in a hash table structure. The specific spatial expression describing the occupancy rate of the segmented curvature voxels is as follows:

[0017]

[0018] In equation (1), SCV ijk,t For the curvature voxel with index (i,j,k) in the segmented curvature voxel occupancy descriptor at the current timestamp t, N ρ N σ and It is the maximum index value in the three projection directions.

[0019] Furthermore, the voxelized generalized iterative nearest-neighbor algorithm based on curvature voxel nearest neighbor search described in step 2 is adapted to the spatial structure of the segmented curvature voxel occupancy descriptor. A Kd-tree search of the curvature voxel center point cloud is used to obtain associated curvature voxel pairs. Specifically, it is assumed that the local point clouds in the curvature voxel association pair {a,b} follow a Gaussian distribution: and Where a m This represents the m-th curvature voxel. Represents the curvature voxel a m The average position, Represents the curvature voxel a m The covariance matrix, b n This represents the nth curvature voxel. Represents curvature voxel b nThe average position, Represents curvature voxel b n The covariance matrix, Represents the curvature voxel a m Gaussian distribution representation, Represents curvature voxel b n If the Gaussian distribution is used to express the pose transformation error, then the corresponding pose transformation error is:

[0020]

[0021] In equation (2), T t+1,t The pose between timestamps t and t+1. Let T represent the pose transformation error. t+1,t It can be estimated using maximum likelihood as follows:

[0022]

[0023] In equation (3), N m For a m The number of points in the matrix, where the superscript T indicates the transpose of the matrix.

[0024] Furthermore, the laser intensity-assisted curvature voxel clustering method described in step 3 specifically includes the following sub-steps:

[0025] Sub-step 3-1 involves generating a laser intensity map by calculating the mean and variance of the intensity within each curvature voxel and projecting it onto the segmented curvature voxel occupancy descriptor. This process is represented as:

[0026]

[0027] In equation (4), This represents the generated laser intensity map, av ijk ,var ijk These represent the mean and variance of intensity within the curvature voxel, respectively;

[0028] Sub-step 3-2, based on the generated laser intensity map A constraint and lifting strategy based on laser intensity is used for object point cloud segmentation, where the constraint function E r (·) The constrained nearest neighbor voxel clustering process is represented as:

[0029]

[0030] In equation (5), r is the nearest neighbor search radius, h av and h var The search thresholds are the intensity matrix and variance, respectively, where (u,v,w) represents the curvature voxel index, and SCV is the variable. uvw,t Represents SCVijk,t neighborhood curvature voxels, av (i,j,k) av represents the mean intensity of the (i,j,k)th curvature voxel. (u,v,w) var represents the mean intensity of the (u,v,w)th curvature voxel. (u,v,w) Let represent the intensity variance of the (u,v,w)th curvature voxel;

[0031] Sub-step 3-3, promotion function G τ,r (·) Merge nearest neighbor segmented objects, which is represented as:

[0032]

[0033] In equation (6), τ is the number of iterations. For nearest neighbor segmentation objects, The object with index k;

[0034] Sub-steps 3-4 involve calculating a 7-dimensional feature vector for each object using a geometric feature-based object classification method. This vector includes linearity, planarity, divergence, principal orientation, maximum height, minimum height, and distribution scale. First, the three eigenvalues ​​of the covariance matrix of the segmented point cloud objects are calculated and sorted from largest to smallest as σ1, σ2, and σ3. The linearity, planarity, and divergence are calculated as follows:

[0035]

[0036] Equation (7)f l ,f p and f s These represent linearity, surface properties, and divergence, respectively.

[0037] Ultimately, objects are categorized into absolutely stationary objects, including the ground, buildings, and trees, and potentially moving objects, including vehicles and pedestrians.

[0038] Furthermore, step 4, which involves aligning the curvature voxels of adjacent frames using the segmented curvature voxel occupancy rate descriptors to track potential dynamic objects, utilizes the registration of curvature voxel vertices to map and locate objects within adjacent segmented curvature voxel occupancy rate descriptors. This process specifically includes the following sub-steps:

[0039] Sub-step 4-1: Extract the three vertices of each curvature voxel, including the nearest vertex p. near,t , central vertex p central,t and the farthest vertex p far,t Then perform coordinate transformation:

[0040]

[0041] In equation (8), SCV t+1,tThis is the expression after registration of curvature voxels;

[0042] Sub-step 4-2, after registering the segmented curvature voxel occupancy rate descriptor, obtains the response object of the tracked object through the segmentation annotation of adjacent curvature voxels:

[0043]

[0044] In equation (9), id is the object index, and PD represents the potential dynamic object. For the object pointed to by id in frame t+1, for The object after curvature voxel registration. In response to a set of objects, v(·) represents the process of extracting the set of curvature voxels occupied by the objects. The process of extracting curvature voxel tags;

[0045] Sub-step 4-3, the method for removing highly dynamic objects based on curvature voxel occupancy rate change detection, involves tracking the object after curvature voxel alignment to obtain the corresponding response object and calculating the object overlap ratio r of the tracked object based on curvature voxel occupancy rate change detection. id,t for:

[0046]

[0047] Based on the obtained object overlap ratio, a threshold h is set. r To determine the motion properties of the tracked object, the basic assumptions are as follows:

[0048] a)r id,t <h r The tracked object is a low-dynamic object and should be preserved.

[0049] b)r id,t h r The tracked object is a high-dynamic object and should be removed.

[0050] Furthermore, the pose residual of static curvature voxel registration in the heavy-load curvature voxel occupancy rate descriptor mentioned in step 5 is calculated as follows:

[0051]

[0052] In formula (11) To account for pose errors due to dynamic weights, The pose residual is denoted by T, and T is an intermediate variable for optimizing the pose. To optimize pose.

[0053] Furthermore, the method for obtaining static instance maps based on point cloud pose and stitching described in step 6 involves optimizing the pose transformation of absolutely stationary objects and low-dynamic objects retained in each frame of point cloud, registering them to the global world coordinate system, and further segmenting the point clouds of power lines and towers based on the spatial distribution characteristics of power lines and towers using levitation analysis.

[0054] Compared with existing technologies, the technical solution provided by this invention has the following beneficial effects:

[0055] 1. The proposed segmentation curvature voxel descriptor is applicable to multiple types of 3D laser sensors, enabling the laser scanning mapping algorithm to be adapted to various mobile LiDAR devices. The mapping algorithm overcomes the interference of dynamic objects in the scene and removes dynamic objects during the mapping process, avoiding the ghosting caused by dynamic objects in point clouds that is traditionally removed manually in post-processing.

[0056] 2. To address the issues of insufficient cost annotation and high training cost faced by deep learning methods for point cloud instance segmentation, the strategy proposed in this invention combines laser intensity-assisted voxel clustering and geometric feature-based object classification. This non-deep learning approach does not rely on GPUs and achieves good segmentation results for street-side power distribution line scenes.

[0057] 3. To address the problem of incorrect removal of static point clouds due to improper dynamic information representation in traditional dynamic object removal methods, the high dynamic object removal method proposed in this invention based on curvature voxel occupancy rate change detection has consistently demonstrated high robustness to motion blur and scanning, removing most dynamic objects while retaining as many static objects as possible. Attached Figure Description

[0058] Figure 1 This is a basic flowchart of the present invention;

[0059] Figure 2 This is a flowchart of the laser intensity-assisted voxel clustering method proposed in this invention;

[0060] Figure 3 The flowchart shows the high dynamic object removal method based on curvature voxel occupancy rate change detection proposed in this invention.

[0061] Figure 4 Laser scanning modeling rendering of a power scene. Figure 4 (a) is the original global map, (b) is the global map with semantic segmentation, (c) is the result of removing vehicles from the point cloud, and (d) is the result of extracting power lines and towers. Detailed Implementation Plan

[0062] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0063] Step 1: Use a handheld multi-line lidar to scan the street-side power distribution lines, including power lines and towers, as well as ground features, vegetation, buildings, vehicles, pedestrians, and other objects. Obtain the current frame of the original lidar point cloud as the current frame, and construct a segmentation curvature voxel occupancy descriptor within it.

[0064] Step 2: Using the voxelization generalized iterative nearest point algorithm based on curvature voxel nearest neighbor search, the pose of the current frame is obtained by optimizing the pose residual of the associated voxel registration.

[0065] Step 3: Use a laser intensity-assisted curvature voxel clustering method to segment the point cloud of the current frame into objects. Then, use a geometric feature-based object classification method to identify the segmented objects and label absolutely static objects, such as ground, vegetation, and buildings, and potentially dynamic objects, such as vehicles and pedestrians.

[0066] Step 4: Using the segmentation curvature voxel occupancy rate descriptors of adjacent frames, perform curvature voxel alignment on adjacent frames to track potential dynamic objects. Then, based on the curvature voxel occupancy rate change detection, retain low-dynamic objects, such as parked vehicles and stopped pedestrians, and remove high-dynamic objects, such as moving vehicles and walking pedestrians.

[0067] Step 5: The pose residual of the static curvature voxel registration in the reloaded curvature voxel occupancy descriptor is optimized again using the voxelization generalized iterative nearest point algorithm based on curvature voxel nearest neighbor search to obtain the current pose of the current frame after optimization.

[0068] Step 6: Based on the optimized pose, perform pose transformation frame by frame and stitch the images together to obtain a global static point cloud map after removing highly dynamic objects. Then, use the spatial distribution geometric features of power lines and towers to extract power lines and towers from the point cloud.

[0069] Specifically, the process of constructing the segmented curvature voxel occupancy rate descriptor in step 1 involves encoding the preprocessed point cloud into a set of curvature voxels at a fixed resolution in three projection directions. Curvature voxels containing at least one point are stored in a hash table structure. The specific spatial expression for the segmented curvature voxel occupancy rate descriptor is as follows:

[0070]

[0071] In equation (1), SCV ijk,t For the curvature voxel with index (i,j,k) in the segmented curvature voxel occupancy descriptor at the current timestamp t, N ρN σ and It is the maximum index value in the three projection directions.

[0072] Specifically, the voxelized generalized iterative nearest-neighbor algorithm based on curvature voxel nearest neighbor search described in step 2 is adapted to the spatial structure of the segmented curvature voxel occupancy descriptor. It uses a Kd-tree search of the curvature voxel center point cloud to obtain associated curvature voxel pairs. It is assumed that the local point clouds in the curvature voxel association pair {a,b} follow a Gaussian distribution: and Where a m This represents the m-th curvature voxel. Represents the curvature voxel a m The average position, Represents the curvature voxel a m The covariance matrix, b n This represents the nth curvature voxel. Represents curvature voxel b n The average position, Represents curvature voxel b n The covariance matrix, Represents the curvature voxel a m Gaussian distribution representation, Represents curvature voxel b n If the Gaussian distribution is used to express the pose transformation error, then the corresponding pose transformation error is:

[0073]

[0074] In equation (2), T t+1,t The pose between timestamps t and t+1. Let T represent the pose transformation error. t+1,t It can be estimated using maximum likelihood as follows:

[0075]

[0076] In equation (3), N m The superscript T represents the number of points in the associated pair, and the superscript T denotes the transpose of the matrix.

[0077] Specifically, the laser intensity-assisted curvature voxel clustering method described in step 3 first generates a laser intensity map by calculating the mean and variance of the intensity within each curvature voxel and projecting it onto the segmented curvature voxel occupancy descriptor. This process is represented as:

[0078]

[0079] In equation (4), This represents the generated laser intensity map, av ijk ,varijk These represent the mean and variance of intensity within the curvature voxels, respectively. Then, based on the generated intensity map... A laser intensity-based constraint and lifting strategy is used for object point cloud segmentation. The constraint function E... r (·) The constrained nearest neighbor voxel clustering process is represented as:

[0080]

[0081] In equation (5), r is the nearest neighbor search radius, h av and h vae The search thresholds are the intensity matrix and variance, respectively, where (u,v,w) represents the curvature voxel index, and SCV is the variable. uvw,t Represents SCV ijk,t neighborhood curvature voxels, av (i,j,k) av represents the mean intensity of the (i,j,k)th curvature voxel. (u,v,w) var represents the mean intensity of the (u,v,w)th curvature voxel. (u,v,w) Let G represent the intensity variance of the (u,v,w)th curvature voxel. Then, the lifting function G... τ,r (·) Merge nearest neighbor segmented objects, which is represented as:

[0082]

[0083] In equation (6), SCV uvw,t Represents SCV ijk,t The neighborhood curvature voxels, where τ is the iteration number. The nearest neighbor is the dividing object.

[0084] Then, a geometric feature-based object classification method calculates a 7-dimensional feature vector for each object, including linearity, planarity, divergence, principal orientation, maximum height, minimum height, and distribution scale. First, the three eigenvalues ​​of the covariance matrix of the segmented point cloud objects are calculated and sorted from largest to smallest as σ1, σ2, and σ3. Linearity, planarity, and divergence are calculated as follows:

[0085]

[0086] Equation (7)f l ,f p and f s These represent linearity, planarity, and divergence, respectively. Ultimately, objects are classified into absolutely stationary objects, including the ground, buildings, and trees, and potentially moving objects, including vehicles and pedestrians.

[0087] For the detailed process of step 3, please refer to [link / details]. Figure 2 .

[0088] Specifically, the strategy for tracking potential dynamic objects based on curvature voxel alignment described in step 4 extracts three vertices of each curvature voxel, including the nearest vertex p. near,t , central vertex p central,t and the farthest vertex p far,t Then perform coordinate transformation:

[0089]

[0090] In equation (8), SCV t+1,t This is the expression after curvature voxel registration. After implementing segmented curvature voxel occupancy rate descriptor registration, the response object of the tracked object can be obtained through the segmentation annotation of adjacent curvature voxels.

[0091]

[0092] In equation (9), id is the object index, and PD represents the potential dynamic object. For the object pointed to by id in frame t+1, for The object after curvature voxel registration. In response to a set of objects, v(·) represents the process of extracting the set of curvature voxels occupied by the objects. This is the process of extracting curvature voxel tags.

[0093] Then, a method for removing highly dynamic objects based on curvature voxel occupancy rate change detection is used. After object tracking following curvature voxel alignment, the corresponding response object is obtained, and the object overlap ratio r of the tracked object is calculated based on the curvature voxel occupancy rate change detection. id,t for:

[0094]

[0095] In equation (10), v(·) represents the process of extracting the set of curvature voxels occupied by the extracted object. Based on the obtained object overlap ratio and setting a threshold h... r To determine the motion properties of the tracked object, the basic assumptions are as follows:

[0096] a)r id,t <h r The tracked object is a low-dynamic object and should be preserved.

[0097] b)r id,t h r The tracked object is a highly dynamic object and should be removed.

[0098] For the detailed process of step 4, please refer to [link / details]. Figure 3 .

[0099] Specifically, the secondary pose optimization method based on static curvature voxel reloading described in step 5 ignores the curvature voxels occupied by highly dynamic objects, considers only static curvature voxels, and recalculates the pose residuals of curvature voxel registration. This process is as follows:

[0100]

[0101] In formula (11) To account for pose errors due to dynamic weights, The pose residual is denoted by T, and T is an intermediate variable for optimizing the pose. To optimize pose.

[0102] Specifically, in step 6, the method for obtaining static instance maps based on point cloud pose and stitching involves optimizing the pose transformation of absolutely stationary objects and low-dynamic objects retained in each frame of the point cloud and registering them to the global world coordinate system. Considering the spatial distribution characteristics of power lines and towers—that is, the tower point cloud has elevation continuity in the z-axis direction and the power line point cloud has suspension—the suspension characteristics are used to coarsely segment the suspended power line and tower point clouds.

[0103] The original global map was as follows Figure 4 As shown in (a), the global map with semantic segmentation is as follows: Figure 4 As shown in (b), the effect of removing vehicles from the point cloud is shown in [the image]. Figure 4 (c) The effect of extracting power lines and towers is shown in [the image]. Figure 4 (d)

[0104] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A method for LiDAR point cloud mapping for power line scanning in dynamic environment, characterized in that, The method includes the following steps: Step 1: Use a handheld multi-line lidar to scan the power distribution lines on the street and obtain the current frame of the original lidar point cloud as the current frame. Construct a segmentation curvature voxel occupancy rate descriptor in it. Step 2: Using the voxelization generalized iterative nearest point algorithm based on curvature voxel nearest neighbor search, the pose of the current frame is obtained by optimizing the pose residual of the associated voxel registration. Step 3: Use the laser intensity-assisted curvature voxel clustering method to segment the point cloud of the current frame into objects, and then use the geometric feature-based object classification method to identify the segmented objects and label the absolutely static objects. Step 4: Using the segmented curvature voxel occupancy rate descriptor of adjacent frames, perform curvature voxel alignment on adjacent frames, track potential dynamic objects, and then detect changes in curvature voxel occupancy rate to retain low-dynamic objects and remove high-dynamic objects. Step 5: The pose residual of static curvature voxel registration in the reloaded curvature voxel occupancy descriptor is optimized again using the voxelization generalized iterative nearest point algorithm based on curvature voxel nearest neighbor search to obtain the current pose of the current frame after optimization. Step 6: Based on the optimized pose, perform pose transformation frame by frame and stitch the images together to obtain a global static point cloud map after removing highly dynamic objects. Then, use the spatial distribution geometric features of power lines and towers to extract power lines and towers from the point cloud. Specifically, the segmented curvature voxel occupancy rate descriptor constructed in step 1 encodes the semantics and position of objects in the environment into segmented curvature voxels, as detailed below: The preprocessed point cloud is encoded into a set of curvature voxels at a fixed resolution in three projection directions. Curvature voxels containing at least one point are stored in a hash table structure. The specific spatial expression describing the occupancy rate of the segmented curvature voxels is as follows: （1） In formula (1), is the curvature voxel indexed by (i, j, k) in the partitioned curvature voxel occupancy descriptor at the current time stamp t, , and is the maximum index value in the three projection directions.

2. The lidar point cloud mapping method for power line scanning in a dynamic environment according to claim 1, characterized in that: The voxelized generalized iterative nearest-neighbor algorithm based on curvature voxel nearest neighbor search described in step 2 is adapted to the spatial structure of the segmented curvature voxel occupancy descriptor. It uses a Kd-tree search of the curvature voxel center point cloud to obtain associated curvature voxel pairs. Specifically, it assumes that curvature voxel pairs are related. The local point cloud in the image follows a Gaussian distribution: and ,in This represents the m-th curvature voxel. Represents curvature voxels The average position, Represents curvature voxels The covariance matrix, This represents the nth curvature voxel. Represents curvature voxels The average position, Represents curvature voxels The covariance matrix, Represents curvature voxels Gaussian distribution representation, Represents curvature voxels If the Gaussian distribution is used to express the pose transformation error, then the corresponding pose transformation error is: （2） In equation (2), The pose between timestamps t and t+1. Let represent the pose transformation error. It is estimated by maximum likelihood as follows: （3） In equation (3), for The number of points in the matrix, where the superscript T indicates the transpose of the matrix.

3. The lidar point cloud mapping method for power line scanning in a dynamic environment according to claim 1, characterized in that: Step 3, the laser intensity-assisted curvature voxel clustering method, specifically includes the following sub-steps: Sub-step 3-1 involves generating a laser intensity map by calculating the mean and variance of the intensity within each curvature voxel and projecting it onto the segmented curvature voxel occupancy descriptor. This process is represented as: （4） In equation (4), This represents the generated laser intensity map. These represent the mean and variance of intensity within the curvature voxel, respectively; Sub-step 3-2, based on the generated laser intensity map A constraint and lifting strategy based on laser intensity is used to segment object point clouds, where the constraint function... The constrained nearest neighbor voxel clustering process is represented as follows: （5） In equation (5), r is the nearest neighbor search radius. and Here, (u, v, w) represents the search thresholds for the intensity matrix and variance, respectively, and (u, v, w) represents the curvature voxel index. represent neighborhood curvature voxels, Indicates the first Mean intensity of curvature voxels Indicates the first Mean intensity of curvature voxels Indicates the first Intensity variance of curvature voxels; Sub-step 3-3, promotion function The nearest neighbor segmentation object is represented as follows: （6） In formula (6) For the number of iterations, For nearest neighbor segmentation objects, The object with index k; Sub-steps 3-4 involve calculating the 7-dimensional feature vector for each object using a geometric feature-based object classification method. This vector includes linearity, planarity, divergence, principal orientation, maximum height, minimum height, and distribution scale. First, the three eigenvalues ​​of the covariance matrix of the segmented point cloud objects are calculated and sorted from largest to smallest. , and Linearity, surface properties, and divergence are calculated as follows: （7） Equation (7) , and These represent linearity, surface properties, and divergence, respectively. Ultimately, objects are categorized into absolutely stationary objects, including the ground, buildings, and trees, and potentially moving objects, including vehicles and pedestrians.

4. The lidar point cloud mapping method for power line scanning in a dynamic environment according to claim 1, characterized in that: Step 4, which involves using the segmented curvature voxel occupancy rate descriptors of adjacent frames to align the curvature voxels and track potential dynamic objects, utilizes the registration of curvature voxel vertices to map and locate objects within adjacent segmented curvature voxel occupancy rate descriptors. Specifically, this includes the following sub-steps: Sub-step 4-1: Extract the three vertices of each curvature voxel, including the nearest vertex. Central vertex and farthest vertex Then perform coordinate transformation: （8） In equation (8), This is the expression after registration of curvature voxels; Sub-step 4-2, after registering the segmented curvature voxel occupancy rate descriptor, obtains the response object of the tracked object through the segmentation annotation of adjacent curvature voxels: （9） In equation (9), For object indexing, PD represents a potential dynamic object. In frame t+1 The object being referred to, for The object after curvature voxel registration. In response to a collection of objects, The process of extracting the set of curvature voxels occupied by the body. The process of extracting curvature voxel tags; Sub-step 4-3, the method for removing highly dynamic objects based on curvature voxel occupancy rate change detection, involves tracking the object after curvature voxel alignment to obtain the corresponding response object and calculating the object overlap ratio of the tracked object based on curvature voxel occupancy rate change detection. for: （10） Based on the obtained object overlap ratio and a threshold is set. To determine the motion properties of the tracked object, the basic assumptions are as follows: a) The tracked object is a low-dynamic object and should be preserved. b) The tracked object is a high-dynamic object and should be removed.

5. The lidar point cloud mapping method for power line scanning in a dynamic environment according to claim 1, characterized in that: The pose residual of static curvature voxel registration in the heavy-load curvature voxel occupancy rate descriptor mentioned in step 5 is calculated as follows: （11） In formula (11) To account for pose errors due to dynamic weights, The pose residual after heavy load. It is an intermediate variable for optimizing pose. To optimize pose.

6. The lidar point cloud mapping method for power line scanning in a dynamic environment according to claim 1, characterized in that: The method for obtaining static instance maps based on point cloud pose and stitching described in step 6 involves optimizing the pose transformation of absolutely stationary objects and low-dynamic objects retained in each frame of point cloud, registering them to the global world coordinate system, and further segmenting the point clouds of power lines and towers based on the spatial distribution characteristics of power lines and towers using levitation analysis.

Images