Dense power supply area visual reconstruction method based on laser radar guidance

Abstract

The invention relates to the technical field of power supply areas, in particular to a dense power supply area visual reconstruction method based on laser radar guidance, and the method comprises the steps: combining laser radar equipment and binocular image collection equipment, and carrying out the synchronous data collection of a target area, so as to obtain laser radar point cloud data and binocular image data; performing external parameter calibration on the laser radar and the binocular camera, solving a rotation matrix and a translation vector, and establishing a unified world coordinate system; projecting the laser radar point cloud to a binocular image coordinate system to obtain an image sparse reference depth; carrying out depth correction and restoration on a binocular reconstruction result by using the laser radar point cloud; constructing a joint optimization model comprising a binocular constraint term and a laser radar constraint term, and optimizing the depth map; according to the method, a three-dimensional point cloud or model with high precision and density is generated, the three-dimensional point cloud or model is used for measuring and identifying the power supply area equipment, and reliable technical support is provided for accurate measurement and identification of the power supply area equipment by guiding a repair and joint optimization scheme.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power supply area, and relates to a dense power supply area visual reconstruction method based on laser radar guidance. BACKGROUND

[0002] At present, binocular stereo vision is widely used in three-dimensional reconstruction, which has the advantages of being able to obtain dense depth maps and retain rich color texture information. However, in complex environments such as low texture, strong light interference and occlusion, the binocular vision reconstruction result often has errors, drifts and holes, resulting in insufficient measurement accuracy, which is difficult to meet the needs of high-precision application scenarios such as power supply areas. In contrast, laser radar can provide reliable spatial position point cloud data due to its high-precision ranging capability, but its point cloud data is sparse and difficult to express detailed structures, especially in small components (such as insulators and connectors) in power supply areas, making it difficult to achieve complete reconstruction, thereby limiting its application in dense modeling.

[0003] For example, the announcement number CN112132972B discloses a kind of laser and image data fusion's three-dimensional reconstruction method and system. The method is by obtaining the three-dimensional point cloud data of laser radar and the RGB image data of camera, laser radar point cloud is mapped to the RGB image coordinate system of camera, combined with point cloud registration and texture projection realizes three-dimensional reconstruction. It uses ICP algorithm to carry out point cloud rough registration, and realizes external parameter calibration by PnP algorithm. Although the method shows certain effect in global data alignment, its fusion is limited to point cloud registration level, and does not specially process the holes or local drift of binocular depth map. In low-texture scene, system deviation is easy to produce, reconstruction result is high in sparseness (lack of guided repair mechanism), measurement accuracy is insufficient 5 centimeters, which limits the practicability in complex environment.

[0004] As the announcement number CN117351140A proposes a kind of fusion panoramic camera and laser radar's three-dimensional reconstruction method, device and equipment. The method is by initial camera pose estimation to panoramic image, combined with laser radar pose optimization, generates panoramic depth map, and carries out three-dimensional reconstruction based on optimization result and laser radar point cloud data. It includes LiDAR pose guided panoramic depth filling, and optimization using energy minimization framework. However, this method is mainly designed for panoramic view, ignoring binocular disparity consistency constraint; after fusion, texture information is inaccurate in complex occlusion or low light environment, point cloud density is low (lack of weighted interpolation mechanism);At the same time, drift is easy to produce, resulting in measurement error more than 10 centimeters, which is difficult to meet the needs of high precision.

[0005] In summary, although the prior art has made some progress in the fusion of lidar and visual data, there are still problems such as insufficient fusion depth, poor robustness, and weak application targeting. In particular, in the low-texture and complex occlusion power station area scene, it is difficult to achieve three-dimensional reconstruction with high precision and density. SUMMARY

[0006] Therefore, the purpose of the present application is to provide a dense power supply area visual reconstruction method based on lidar guidance, which effectively improves the reconstruction accuracy and integrity by introducing LiDAR point cloud as geometric prior, combining the density of binocular vision, guiding the repair and joint optimization scheme, and providing reliable technical support for accurate measurement and identification of power supply area equipment.

[0007] To achieve the above purpose, the present application provides the following technical scheme:

[0008] A dense power supply area visual reconstruction method based on lidar guidance, comprising the following steps:

[0009] S1, data acquisition: synchronously acquiring target area through a joint acquisition system to obtain lidar point cloud data and binocular image data; wherein the joint acquisition system comprises a lidar device and a binocular image acquisition device;

[0010] S2, external parameter calibration and coordinate unification: calibrating the external parameters of the lidar device and the binocular image acquisition device, solving the rotation matrix and the translation vector, and establishing a unified world coordinate system;

[0011] S3, initial reconstruction and projection:

[0012] Stereoscopic matching is performed on the binocular image data to generate an initial dense depth map;

[0013] The lidar point cloud data is projected to the image coordinate system of the binocular image acquisition device based on the unified world coordinate system to obtain image sparse reference depth;

[0014] S4, depth repair and error correction: using the image sparse reference depth to correct the depth of the initial dense depth map globally and interpolate the repair of the local area to generate a repaired depth map;

[0015] S5, sparse-dense fusion optimization: constructing a joint optimization model containing binocular constraint items and lidar constraint items, optimizing the repaired depth map, and obtaining a final depth map;

[0016] S6, result output: the final depth map is converted into a three-dimensional space point cloud, a three-dimensional point cloud or model with high precision and density is generated, and is used for power station area equipment measurement and identification.

[0017] The laser radar device and the binocular image acquisition device are fixed on the same support, the angles of the two are adjusted to make the acquisition directions the same, more overlapping areas are covered between the respective fields of view, the relative positions between the laser radar device and the binocular image acquisition device are ensured to remain unchanged during acquisition, and data is acquired through synchronous acquisition.

[0018] The application further sets that the step S2 comprises:

[0019] A joint calibration scene is built and a joint calibration board is placed, and a plurality of groups of calibration board point clouds and image data under different postures are acquired,

[0020] A perspective point algorithm is used to solve a rotation matrix R and a translation vector t, wherein R represents a 3*3 rotation matrix, and t represents a 3*1 translation vector;

[0021] An extrinsic parameter matrix is established ; the laser radar point cloud is converted into the coordinate system of the binocular image acquisition device;

[0022] The intrinsic parameter matrix of the binocular image acquisition device is:

[0023] , wherein , is a focal length, , is an optical center coordinate; the binocular baseline is B, and the focal length is taken as ,

[0024] The laser radar point cloud point .

[0025] The binocular depth , wherein represents a parallax value, is a pixel coordinate, the baseline B represents the distance between two lenses of a binocular camera,

[0026] The confidence degree is calculated based on a matching cost, and ranges from 0 to 1.

[0027] The application further sets that, in the step S3, the left and right images are subjected to de-warping and stereoscopic correction, and a half-global block matching (SGBM) algorithm is used to obtain a parallax map and a confidence map ,

[0028] wherein

[0029] calculating the initial dense depth map according to the disparity map :

[0030] ;

[0031] coordinate transformation of the lidar point:

[0032] , wherein represents the three-dimensional coordinates of the point in the camera coordinate system;

[0033] the point in the binocular image acquisition device coordinate system is projected to a pixel:

[0034] ;

[0035] obtaining the image sparse reference depth as ; retaining and , and marking low confidence, wherein is an occlusion threshold parameter;

[0036] performing point cloud coarse registration by using an iterative closest point (ICP) algorithm.

[0037] The application is further provided as follows: the depth repair and error correction in the step S4 includes global affine correction and local depth repair,

[0038] the global affine correction is used for correcting the overall deviation or linear drift of the binocular depth scale, and a scale factor s and an offset b are solved by minimizing a robust cost function:

[0039] ;

[0040] and a global correction depth map is generated accordingly: ;

[0041] , wherein represents a robust function, represents a binocular depth value, represents a lidar depth value, represents a lidar projection point coordinate;

[0042] the local depth repair adopts kernel weighted local linear regression, and a model is established for the neighborhood of the lidar projection point :

[0043] , wherein represents a repaired depth value, represents a neighborhood pixel coordinate, denote local model parameters;

[0044] by minimizing an objective function:

[0045] solving ;

[0046] weights wherein denote pixel distances, denote standard deviation parameters, denote confidence values, propagate depth to fill holes, and finally obtain the repaired depth map.

[0047] The application further provides that in the step S5, a joint energy function is constructed:

[0048] ;

[0049] wherein denote depth fields, is a photometric consistency constraint term, is a smoothness constraint term, is a lidar constraint term, and denote weight parameters;

[0050] The energy function is optimized by a gradient descent method, and a bilateral filtering algorithm is used to process the point cloud to eliminate noise and retain edges, a guided filtering is used to guide the depth filling from a color image, holes are repaired, and finally the final depth map is obtained.

[0051] The application further provides that in the step S6, the final depth map is converted into a three-dimensional point cloud, and the point cloud points are:

[0052] ;

[0053] wherein denote three-dimensional point coordinates, denote inverses of intrinsic matrices, denote homogeneous pixel coordinate vectors, the target device is segmented and geometrically fitted, including a plane, a cylinder or a box model, and the size and spacing are calculated.

[0054] The application further provides that the iterative closest point algorithm further comprises: down-sampling the dense point cloud and voxelizing the lidar point cloud, using a point-to-plane ICP to minimize the distance, and removing outliers.

[0055] The application further provides that in the sparse point cloud projection, the confidence map is obtained by right-left consistency checking, and constraints are established for the projection points satisfying the confidence and occlusion conditions.

[0056] The application is further configured to solve the local model parameters by using an optimization method with a regularization term , which can be applied to power supply station scenarios with low texture or strong light interference. Compared with the shortcomings of the prior art, the application has the following beneficial effects:

[0057] The application effectively solves the problems of inaccurate depth, susceptibility to light, poor reconstruction effect in texture sparse areas, and insufficient precision in traditional power supply station complex scenarios that only rely on monocular or binocular stereo matching by combining the precise depth measurement capability of laser radar with the dense 3D perception capability of binocular cameras, and introducing fine external parameter calibration, depth repair, and sparse-dense fusion optimization algorithm. The application can generate a three-dimensional point cloud or model with high precision (laser radar precision) and density (binocular camera density), which is crucial for accurate measurement, state recognition, defect detection, and three-dimensional modeling of power supply station equipment (such as transformers, circuit breakers, insulators, etc.), significantly improving the efficiency and accuracy of power supply station inspection and maintenance. BRIEF DESCRIPTION OF DRAWINGS

[0058] Figure 1 The flowchart of the application is shown in the figure;

[0059] Figure 2 The external parameter calibration and coordinate unification flowchart of the application is shown in the figure;

[0060] Figure 3 The depth repair and error correction flowchart of the application is shown in the figure;

[0061] Figure 4 The sparse-dense fusion optimization flowchart of the application is shown in the figure;

[0062] Figure 5 The depth repair schematic diagram of the application is shown in the figure;

[0063] Figure 6 The joint optimization schematic diagram of the application is shown in the figure. DETAILED DESCRIPTION

[0064] Reference Figures 1 to 6 The application further describes a laser radar guided dense power supply station visual reconstruction method embodiment. To achieve high-precision and dense power supply station three-dimensional reconstruction, the application designs a complete process including data acquisition, external parameter calibration and coordinate unification, sparse point cloud projection, depth repair and error correction, sparse-dense fusion optimization, and result output.

[0065] Step one, data acquisition:

[0066] Device selection and parameter optimization: In order to ensure the reconstruction accuracy of small and medium-sized components (such as insulators) in the power distribution area, the laser radar device is preferably a high-line beam (such as 32 lines or 64 lines) mechanical laser radar, and the ranging accuracy is preferably ±2 cm. The binocular image acquisition device is preferably a pair of high-resolution (such as 2048x1536 pixels) global shutter industrial cameras, and the baseline (the distance between the two camera lenses) B is accurately set to 150 mm.

[0067] System integration: The laser radar device and the binocular image acquisition device are rigidly fixed on the same T-shaped alloy bracket. Adjust the installation angle of both to make their acquisition directions the same (for example, both facing straight ahead), and ensure that there is more than 70% overlap between their respective fields of view. Rigid fixation ensures that the relative position between the laser radar device and the binocular image acquisition device remains unchanged during acquisition.

[0068] Data synchronization: Use a hardware synchronization clock based on PTP (Precision Time Protocol) to synchronize and trigger acquisition of the laser radar device and the binocular image acquisition device, ensuring that the timestamp error of the laser radar point cloud data and the binocular image data is less than 1 ms.

[0069] In the data acquisition phase, the invention uses a joint acquisition system composed of a laser radar device and a binocular image acquisition device (hereinafter referred to as a binocular camera).

[0070] Step two, external parameter calibration and coordinate unification:

[0071] Build a joint calibration scene and place a joint calibration board with a visual checkerboard pattern and a laser radar high-reflective feature (such as a high-reflective material attached to the corners of the checkerboard) in the scene. Collect point cloud and image data of the calibration board from at least 20 different poses and distances.

[0072] Parameter solving: Extract the pixel coordinates of the corner points in the image and the three-dimensional coordinates of the corner points in the point cloud, and use a robust perspective point (PnP) algorithm to solve the rotation matrix R (a 3x3 matrix) and the translation vector t (a 3x1 vector) from the laser radar coordinate system to the binocular image acquisition device coordinate system (based on the left camera)

[0073] Coordinate system establishment:

[0074] Establishing the external parameter matrix , taking the binocular image acquisition device coordinate system as the basis, as the unified world coordinate system,

[0075] Obtain the intrinsic parameter matrix of the binocular image acquisition device in advance through Zhang Zhengyou calibration method :

[0076]

[0077] wherein and is the normalized focal length of the camera in the x, y axis, and is the optical center coordinate, Take 1600 pixels.

[0078] A laser radar point cloud point to be transformed is represented as

[0079] In the extrinsic parameter calibration process in step two, in order to improve the accuracy of calibration, after obtaining the initial camera-laser radar pose using the PnP algorithm, the iterative closest point (ICP) algorithm is further used for rough correction. The specific method is to downsample the dense point cloud generated by the binocular camera (obtained in steps four and five) and voxelize the laser radar point cloud. Then, the point-to-plane ICP matching algorithm is adopted to minimize the distance between the two sets of point clouds, and obvious outliers are removed, so as to obtain a more accurate extrinsic parameter.

[0080] Step three, sparse point cloud projection:

[0081] After completing the extrinsic parameter calibration and coordinate system unification, the present application enters the sparse point cloud projection stage. First, the binocular images to be collected are subjected to distortion removal and stereo correction processing to compensate for lens distortion and make the imaging planes of the left and right images coplanar, thereby improving the subsequent binocular ranging accuracy. Then, the semi-global block matching (SGBM) algorithm or graph cut method is used to calculate the disparity map of the image pair . Based on the disparity map, the depth map obtained by binocular stereo matching can be calculated .

[0082] Next, the laser radar point cloud under the laser radar coordinate system is converted to the binocular camera coordinate system through the extrinsic parameter matrix , to obtain . Wherein, is the three-dimensional coordinate of the point under the binocular camera coordinate system.

[0083] Then, the three-dimensional point under the camera coordinate system is projected onto the image plane by using the camera intrinsic parameter matrix K, to obtain the pixel coordinate . For the laser radar point successfully projected into the image, the depth value of the point under the camera coordinate system is the laser radar depth of the point.

[0084] In order to improve the robustness of subsequent depth repair, the laser radar points are filtered, and only the points with depth and projection point points within the image valid range. In addition, occlusion detection is also performed. When the difference between binocular depth and lidar depth exceeds a preset occlusion threshold parameter (e.g. 0.2 meters),

[0085] If , the pixel point is marked as a low confidence area.

[0086] In the sparse point cloud projection stage, in order to achieve coarse-grained alignment of lidar point cloud and binocular image pixels, an iterative closest point (ICP) algorithm is used for point cloud coarse registration. The ICP algorithm first reduces the number of points in the point cloud through voxelization downsampling to speed up the registration, and then iteratively optimizes the rotation and translation through point-to-plane distance minimization, so that the lidar point cloud can be roughly aligned to the point cloud generated by the binocular camera (or the sparse point cloud obtained by binocular depth and internal parameter back projection).

[0087] In sparse point cloud projection, in order to ensure the effectiveness of the lidar point constraint, the calculation of the confidence of the matching candidate disparity value should be consistent in the matching from the left image to the right image and the matching from the right image to the left image. Only the projection points that meet this consistency condition are established with the lidar constraint, and combined with the visibility judgment, the invalid points caused by the poor back projection effect due to the long distance of the lidar point, insufficient resolution or being blocked by other objects are excluded.

[0088] Step four, depth repair and error correction:

[0089] Depth repair and error correction are divided into two sub-stages of global affine correction and local depth repair to overcome the inherent scale bias and linear drift of the binocular system, and fill the depth holes caused by texture loss, occlusion or lidar sparsity.

[0090] Global affine correction: this stage aims to correct the bias of binocular depth estimation in the overall scale factor and linear offset. The optimal scale factor s and offset b are solved by minimizing a robust cost function. The goal is to find s and b such that for all valid lidar projection points , the following objective function is minimized:

[0091] where is a robust function, for example, Huber loss function, which is insensitive to outliers. is the original binocular depth value, is the corresponding lidar depth value, is the pixel coordinate corresponding to the lidar point. After solving the optimal s and b, the corrected binocular depth map is:

[0092] .

[0093] Local depth inpainting: After global rectification, local depth inpainting uses a kernel-weighted local linear regression method to more finely integrate the sparse depth information of the lidar into the binocular depth map, especially in low-texture areas or areas where the lidar data is sparse. For the neighborhood of the lidar projection point , a local model is established: . Wherein is the inpainted depth value, u' is the pixel coordinate in the neighborhood, and are local model parameters. These parameters are solved by minimizing the following objective function:

[0094]

[0095] The weight combines spatial distance and confidence information: . Wherein pixel distance, is a standard deviation parameter that controls the neighborhood size. is the confidence value of the pixel point u', indicating the reliability of the original binocular depth estimation. This method can smoothly propagate the accurate depth information of the lidar to the adjacent area, effectively repairing the inaccurate or empty depth estimation caused by texture loss, occlusion or sparse lidar point cloud. In order to improve the solving efficiency and robustness, the local model parameters and can be obtained by solving the normal equation or adding a regularization term.

[0096] For the solution of the local model parameters and in the local depth inpainting (step four), if the robustness in low-texture or strong light interference power supply station area scenes is considered, the normal equation or L2 regularization term can be added to the least squares objective function to solve, in order to avoid the drastic fluctuation of the model parameters, and improve the accuracy of depth inpainting.

[0097] Step five, sparse-dense fusion optimization:

[0098] The goal of this stage is to further integrate and optimize the depth map after depth inpainting with the accurate three-dimensional information provided by the lidar to construct a joint energy function. The energy function is usually in the following form:

[0099] where D denotes the depth field to be optimized; is a photometric consistency or disparity consistency term, which encourages local regions of the binocular images to remain visually similar when the depth changes; is a smoothness term, which encourages the depth map to remain spatially smooth, preventing unnatural abrupt changes, but at the same time allowing depth jumps at image edges; is a lidar constraint term, which constrains the repaired depth map with the calibrated lidar depth values, especially for the regions where the lidar points are projected to; and are weight parameters, used to balance the importance of different constraint terms.

[0100] The joint energy function can be solved iteratively by optimization algorithms such as gradient descent to obtain the optimal depth map D. After optimization, to further improve the quality of the point cloud, the invention uses a bilateral filtering algorithm to process the generated point cloud, which can effectively eliminate noise while preserving depth edge information as much as possible. In addition, Guided Filter is also used, which uses color images (texture information) to guide depth filling, which can more intelligently repair the hole regions in the depth map, ensuring that the final fused point cloud model has both high precision and density.

[0101] Step six, result output:

[0102] Convert the optimized depth map D(uBic into a three-dimensional point cloud P(u). The three-dimensional coordinate calculation formula of the point cloud point is:

[0103] where, is the three-dimensional coordinate of the point in the camera coordinate system is the value of the depth map at pixel is the inverse matrix of the intrinsic matrix K, is the homogeneous pixel coordinate vector.

[0104] The generated three-dimensional point cloud can be used for measurement and identification of power station area equipment. Specifically, the target equipment is segmented (for example, based on the normal vector, curvature and other features of the point cloud), and then the geometric parameters such as size, position and spacing of the equipment are accurately extracted through geometric fitting (such as plane, cylindrical or box model), providing high-precision data support for subsequent automated detection, maintenance and management.

[0105] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Those skilled in the art can make common changes and substitutions within the technical solution of the present application, which should be included in the protection scope of the present application.​

Claims

1. A method for visual reconstruction of dense power supply areas based on lidar guidance, characterized in that, Includes the following steps: S1. Data Acquisition: The target area is synchronously acquired through a joint acquisition system to obtain lidar point cloud data and binocular image data; wherein, the joint acquisition system includes lidar equipment and binocular image acquisition equipment; S2. External parameter calibration and coordinate unification: Perform external parameter calibration on the lidar equipment and the binocular image acquisition equipment, solve the rotation matrix and translation vector, and establish a unified world coordinate system; S3. Initial Reconstruction and Projection: Stereo matching is performed on the binocular image data to generate an initial dense depth map; The lidar point cloud data is projected onto the image coordinate system of the binocular image acquisition device based on the unified world coordinate system to obtain the image sparse reference depth. S4. Depth Restoration and Error Correction: Using the sparse reference depth of the image, perform global-scale depth correction and local region interpolation restoration on the initial dense depth map to generate a restored depth map. S5. Sparse-Dense Fusion Optimization: Construct a joint optimization model that includes binocular constraints and lidar constraints, optimize the repaired depth map, and obtain the final depth map; S6. Output Results: The final depth map is back-projected into a 3D spatial point cloud, generating a 3D point cloud or model with both high accuracy and density, which is used for the measurement and identification of equipment in the power supply area.

2. The method for visual reconstruction of dense power supply areas based on lidar guidance according to claim 1, characterized in that, In step S1: The lidar device and the binocular image acquisition device are fixed on the same bracket, and their angles are adjusted to make the acquisition directions the same, so that their respective fields of view cover more overlapping areas, ensuring that the relative positions between the lidar device and the binocular image acquisition device remain unchanged during acquisition, and data is acquired through synchronous acquisition.

3. The method for visual reconstruction of dense power supply areas based on lidar guidance according to claim 2, characterized in that, Step S2 includes: A joint calibration scenario was set up and a joint calibration board was placed. Point cloud and image data of the calibration board under different poses were collected. The perspective point algorithm is used to solve for the rotation matrix R and the translation vector t, where R represents a 3×3 rotation matrix and t represents a 3×1 translation vector. Establish the extrinsic parameter matrix Using the coordinate system of the binocular image acquisition device as a reference, the lidar point cloud is transformed into that coordinate system; Among them, the intrinsic parameter matrix of the binocular image acquisition device: ,in , Focal length , The coordinates of the optical center are given; the binocular baseline is B, and the focal length is taken as... , LiDAR point cloud ; Binocular Depth ,in Indicates the disparity value. The coordinates are in pixels, and baseline B represents the distance between the two lenses of the stereo camera. Confidence Calculated based on matching cost, ranging from 0 to 1.

4. The method for visual reconstruction of dense power supply areas based on lidar guidance according to claim 3, characterized in that, In step S3, distortion correction and stereo correction are performed on the left and right images, and the disparity map is obtained using the semi-global block matching (SGBM) algorithm. and confidence plot , in Calculate the initial density depth map based on the disparity map. : ； Perform coordinate transformation on the lidar points: ,in This represents the three-dimensional coordinates of a point in the camera coordinate system; The points in the coordinate system of the binocular image acquisition device are projected onto pixels: ； The image sparse reference depth is obtained as follows ;reserve and Mark low confidence, where This refers to the occlusion threshold parameter. Coarse registration of point clouds was performed using the Iterative Closest Point (ICP) algorithm.

5. The method for visual reconstruction of dense power supply areas based on lidar guidance according to claim 4, characterized in that, The depth restoration and error correction in step S4 include global affine correction and local depth restoration. The global affine correction is used to correct the overall bias or linear drift of the binocular depth scale, and is solved by minimizing the robust cost function to obtain the scale factor s and the offset b: ； And based on this, a global corrected depth map is generated: ; in Represents a robust function. Indicates the depth value of the binoculars. Indicates the depth value of the lidar. Indicates the coordinates of the lidar projection point; The local depth repair employs kernel-weighted local linear regression for the LiDAR projection points. neighborhood Model building: ,in This indicates the depth value after repair. Represents the coordinates of neighboring pixels. Represents local model parameters; By minimizing the objective function: Solve ; Weight ,in Indicates pixel distance. Represents the standard deviation parameter. The confidence level value represents the propagation depth used to repair the holes, ultimately resulting in the repaired depth map.

6. The method for visual reconstruction of dense power supply areas based on lidar guidance according to claim 4, characterized in that, In step S5, the joint energy function is constructed: ； in Represents the depth field. For photometric consistency constraints, For the smoothing constraint term, For lidar constraints, and Indicates the weighting parameter; The energy function is optimized using gradient descent, and a bilateral filtering algorithm is used to process the point cloud to eliminate noise and preserve edges. Guided filtering is then used to guide depth filling from the color image to repair hole areas, resulting in the final depth map.

7. The method for visual reconstruction of dense power supply areas based on lidar guidance according to claim 5, characterized in that, In step S6, the final depth map is generated. Convert to 3D point cloud, point cloud points: ； in Represents the coordinates of a three-dimensional point. Represents the inverse of the intrinsic parameter matrix. Represents a homogeneous pixel coordinate vector, used for segmentation and geometric fitting of the target device, including planar, cylindrical, or box models, calculating dimensions and spacing.

8. The method for visual reconstruction of dense power supply areas based on lidar guidance according to claim 3, characterized in that, The iterative nearest point algorithm further includes: downsampling dense point clouds and voxelizing LiDAR point clouds, minimizing the distance using point-to-plane ICP, and removing outliers.

9. The method for visual reconstruction of dense power supply areas based on lidar guidance according to claim 8, characterized in that, In the sparse point cloud projection, the confidence map The right-left consistency test was used to establish constraints on the projection points that meet the confidence and occlusion conditions.

10. A method for visual reconstruction of dense power supply areas based on lidar guidance according to claim 9, characterized in that, The local model parameters are solved using an optimization method with regularization terms. It can be applied to power supply areas with low texture or strong light interference.

Images