Curvature adaptive sampling based background point cloud processing method and system

Abstract

The present application relates to a background point cloud processing method and system based on curvature adaptive sampling, which combines panoramic point cloud and sub-module point cloud, constructs KD tree space index, sets tolerance threshold, detects coincident points in superimposed point cloud through radius search, removes all coincident points, and obtains initial background point cloud composed of non-coincident points; calculates the normal vector and curvature value of the background point cloud, divides the point cloud point set into high curvature point set and low curvature point set according to the set curvature threshold; performs XYZ axis block on the background point cloud, sets differentiated voxel sampling resolution according to the proportion of high curvature points in the block; performs voxel sampling on the block using differentiated voxel sampling resolution to obtain voxel sampling point cloud meeting the point number requirement; and converts the voxel sampling point cloud into an octomap object. The present application can effectively improve the quality of the substation background point cloud, and has fast processing speed and high efficiency through efficient removal of coincident points, accurate filtering of outliers and adaptive sampling in different regions.

Description

Technical Field

[0001] This invention relates to a method and system for processing background point clouds based on curvature adaptive sampling. Background Technology

[0002] Substations are high-risk areas for live-line work, characterized by dense equipment and confined working spaces. During hoisting operations, the complex posture and large range of motion of the crane boom make it highly susceptible to accidents if it approaches or touches personnel, vehicles, or other live conductors. Currently, the industry commonly uses high-precision positioning terminals to locate personnel and vehicles within substations in real time. Based on the positioning coordinates, the complete posture of the crane boom is reconstructed in three-dimensional space, allowing for the calculation of safe distances between the boom and live conductors or other objects, thus achieving safe control of hoisting operations. To achieve accurate distance calculation and collision warning between the boom and live conductors, high-quality background point cloud modeling of the substation environment is required, providing a basis for machine recognition and positioning of the boom and other objects. However, existing background point cloud processing methods have the following drawbacks: 1) The sampling strategy is fixed, and the same sampling rate is used for high-curvature electrical equipment and low-curvature ground, resulting in the loss of details of electrical equipment or data redundancy; 2) The filtering efficiency of overlapping points and invalid points is low, and an efficient spatial indexing structure is not used, which can easily lead to misjudgment of distance calculation; 3) Outlier removal is not thorough, and the point cloud is not structured and converted into a collision detection format, which cannot be linked with the positioning data of cranes, personnel, and vehicles for calculation, thereby reducing the accuracy and reliability of machine recognition and positioning. Summary of the Invention

[0003] The purpose of this invention is to provide a background point cloud processing method and system based on curvature adaptive sampling, which can adapt to the scene characteristics involved in the safety management of substation hoisting operations and provide favorable conditions for related 3D collision detection.

[0004] The technical solution of this invention is: a background point cloud processing method based on curvature adaptive sampling, comprising the following steps: S1. Obtain the background point cloud; S2. Calculate the normal vector and curvature value of the background point cloud (each point), and divide the point cloud point set into a high curvature point set and a low curvature point set according to the set curvature threshold; S3. Divide the background point cloud into XYZ axis blocks and set the differential voxel sampling resolution according to the proportion of high curvature points in the block; S4. Implement block-based voxel sampling using differentiated voxel sampling resolution to obtain a voxel sampling point cloud that meets the point count requirement (not exceeding the point count threshold); S5. Convert the voxel sampling point cloud into an octomap object.

[0005] Therefore, the octomap object can be used for data processing related to safety management of substation hoisting operations, such as calculating the distance between the boom and live parts.

[0006] Preferably, step S1 (acquiring the background point cloud) includes the following steps: S11. Load the overall point cloud and the sub-module point clouds, remove invalid points and merge them into an overlay point cloud; S12. Construct a KD tree spatial index, set a tolerance threshold, detect (mark) overlapping points in the superimposed point cloud through radius search, remove all overlapping points, and obtain an initial background point cloud composed of non-overlapping points. S13. The initial background point cloud is filtered by a statistical outlier removal filter to remove outliers and form a background point cloud.

[0007] The tolerance threshold is set according to the actual situation. For example, in scenarios involving safety management of substation hoisting operations, the tolerance threshold can usually be set to 0.1m.

[0008] The point count threshold can be set according to the actual scenario and / or actual needs. Once the number of points in the voxel sampling point cloud does not exceed (is less than or equal to) the point count threshold, it is considered a voxel sampling point cloud that meets the point count requirements.

[0009] The panoramic point cloud used for merging should be the panoramic point cloud after removing invalid points. If the loaded panoramic point cloud is the original point cloud, invalid points can be filtered out to form the panoramic point cloud after invalid points are removed. The modular point cloud used for merging should be the complete modular point cloud. If the loaded modular point cloud is not the complete modular point cloud, it can be processed into a complete modular point cloud based on the panoramic point cloud before merging.

[0010] The overall point cloud and the modular point cloud should generally have the same source and coordinate system.

[0011] Based on the overall scenic area cloud, any suitable existing technology can be used to form a target modular point cloud. The initial modular point cloud can then be refined into a complete modular point cloud.

[0012] Preferably, overlapping point detection can be achieved using pcl::KdTreeFLANN.

[0013] Preferably, when the proportion of high curvature points within a block is greater than 30%, the voxel resolution is set to 0.01 times the block size; otherwise, it is set to 0.05 times the block size.

[0014] Preferably, in step S5, if the number of points after the initial voxel sampling exceeds the point number threshold, the number of points is reduced by iterative voxel downsampling and / or random sampling until a voxel sampling point cloud that meets the point number requirement (does not exceed the point number threshold) is obtained.

[0015] Preferably, an iteration number threshold (maximum iteration number) is set. When the number of points after the initial voxel sampling exceeds (is greater than) the number of points threshold, the voxel size is first adjusted iteratively to reduce the number of points. If the number of points still exceeds the number of points after reaching the iteration number threshold, random sampling is then used to control the number of points within the number of points threshold range (not greater than the number of points threshold).

[0016] Preferably, when iteratively adjusting the voxel size, the voxel is increased by cbrt(ratio×1.2f). The threshold for the number of iterations to adjust the voxel size can be set according to requirements, for example, to 10 times.

[0017] The substation background point cloud processing system based on curvature adaptive sampling uses any of the background point cloud processing methods disclosed in this invention to perform substation-related background point cloud processing.

[0018] This background point cloud processing system can include a background point cloud acquisition module, a segmentation module, a voxel sampling module, and a conversion output module. The background point cloud acquisition module acquires the background point cloud; the segmentation module segments the background point cloud along the XYZ axes; the voxel sampling module calculates the normal vector and curvature value of each point in the background point cloud, divides the point cloud into high-curvature and low-curvature point sets according to a set curvature threshold, sets a differentiated voxel sampling resolution based on the proportion of high-curvature points within each segment, and performs voxel sampling on each segment using this differentiated resolution to obtain a voxel-sampled point cloud that meets the point count requirement (not exceeding the point count threshold); the conversion output module converts the voxel-sampled point cloud into an octomap object and outputs it.

[0019] If the number of points after initial voxel sampling exceeds the point count threshold, the number of points is reduced by iterative voxel downsampling and / or random sampling until a voxel sampling point cloud that meets the point count requirement (does not exceed the point count threshold) is obtained.

[0020] Preferably, the background point cloud acquisition module includes a point cloud merging module, a point cloud deduplication module, and a point cloud filtering module. The point cloud merging module is used to merge the overall point cloud and the sub-module point clouds into a superimposed point cloud. The point cloud deduplication module is used to construct a KD tree spatial index, set a tolerance threshold, detect (mark) overlapping points in the superimposed point cloud through radius search, remove all overlapping points, and generate an initial background point cloud composed of non-overlapping points. The point cloud filtering module is used to filter the initial background point cloud with a statistical outlier removal filter to remove outliers and form the background point cloud.

[0021] Furthermore, the background point cloud acquisition module can also include a sub-module point cloud module, which is used to generate a sub-module point cloud of the target based on the overall point cloud.

[0022] The beneficial effects of the present invention are: (1) It is suitable for densely equipped substations with high curvature areas, retains details in high curvature areas and simplifies data in low curvature areas, which can reduce the amount of data processing and improve efficiency while meeting actual needs; (2) It effectively removes invalid points and outliers and is fast; (3) It outputs octomap in a structured manner, which can be directly connected to the collision detection system to improve real-time performance.

[0023] This invention improves the quality of substation background point clouds by efficiently removing overlapping points, accurately filtering outliers, and performing regional adaptive sampling, thus better adapting to the needs of boom posture reproduction, live distance calculation, and collision detection.

[0024] This invention is adapted to substations with densely packed live equipment. While preserving high curvature details, it simplifies data, providing a high-precision environmental model for boom posture reproduction and distance calculation between the boom and live parts, thereby improving the efficiency and reliability of safety management in lifting operations. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the background point cloud processing flow involved in the present invention; Figure 2 This is a schematic diagram of the background point cloud processing system involved in the present invention. Detailed Implementation

[0026] See Figure 1 and Figure 2 The processing method of this invention can be implemented using any of the processing systems disclosed in this invention, and mainly includes the following processing steps: loading point cloud and removing invalid points; fast deduplication using KD-tree; outlier removal using statistical filtering; curvature calculation and classification; XYZ block division; adaptive voxel sampling based on curvature; if the number of points exceeds the limit, iterative or random sampling as a fallback; conversion to octomap for collision detection. The general process can be: invalid point removal → overlapping point filtering → outlier removal → curvature classification → adaptive voxel sampling → octomap conversion. Example

[0027] 1) Load the PLY format global point cloud and remove invalid points using removeNaNFromPointCloud; 2) Merge the point clouds of different modules; 3) Use KdTreeFLANN to build an index and search for and filter overlapping points with a tolerance radius of 0.1m; 4) Use the StatisticalOutlierRemoval filter and set the neighborhood number and standard deviation thresholds to remove outliers; 5) Calculate the normal vector and curvature using NormalEstimation, and distinguish high and low curvature points based on a threshold. 6) Divide the blocks into fixed sizes and set the overlap rate. If the proportion of high curvature is >30%, use 0.01 times the block size resolution; otherwise, use 0.05 times. 7) When the number of points exceeds the limit, iteratively adjust the voxel size, up to 10 times; if it still exceeds the limit, randomly sample. 8) Convert the segmented point cloud into an octomap, save and output the model information.

[0028] Unless otherwise specified, the preferred and optional technical means disclosed in this invention can be arbitrarily combined to form several different specific embodiments when one preferred or optional technical means is a further limitation of another technical means.

Claims

1. A background point cloud processing method based on curvature adaptive sampling, characterized in that... Includes the following steps: S1. Obtain the background point cloud; S2. Calculate the normal vector and curvature value of the background point cloud, and divide the point cloud point set into a high curvature point set and a low curvature point set according to the set curvature threshold; S3. Divide the background point cloud into XYZ axis blocks and set the differential voxel sampling resolution according to the proportion of high curvature points in the block; S4. Implement block-based voxel sampling using differentiated voxel sampling resolution to obtain a voxel sampling point cloud that meets the point count requirement; if the number of points after voxel sampling exceeds the point count threshold, reduce the number of points through iterative voxel downsampling and / or random sampling until a voxel sampling point cloud that meets the point count requirement is obtained. S5. Convert the voxel sampling point cloud into an octomap object.

2. The background point cloud processing method according to claim 1, characterized in that... Step S1 includes the following steps: S11. Load the overall point cloud and the sub-module point clouds, remove invalid points and merge them into an overlay point cloud; S12. Construct a KD tree spatial index, set a tolerance threshold, detect overlapping points in the superimposed point cloud through radius search, remove all overlapping points, and obtain an initial background point cloud composed of non-overlapping points. S13. The initial background point cloud is filtered by a statistical outlier removal filter to remove outliers and form a background point cloud.

3. The background point cloud processing method according to claim 2, characterized in that... Fast detection of overlapping points is achieved using pcl::KdTreeFLANN.

4. The background point cloud processing method according to any one of claims 1-3, characterized in that... When the proportion of high curvature points within a block is greater than 30%, the voxel resolution is set to 0.01 times the block size; otherwise, it is set to 0.05 times the block size.

5. The background point cloud processing method according to claim 4, characterized in that... Set an iteration threshold (maximum number of iterations). If the number of points after initial voxel sampling exceeds (is greater than) the threshold, first use iterative adjustment of voxel size to reduce the number of points. If the number of points still exceeds the threshold after reaching the iteration threshold, then use random sampling to control the number of points within the threshold range (not greater than the threshold).

6. The background point cloud processing method according to claim 5, characterized in that... When iteratively adjusting the voxel size, increase the voxel size by cbrt(ratio×1.2f).

7. A substation background point cloud processing system based on curvature adaptive sampling, characterized in that... Background point cloud processing related to substations is performed using the background point cloud processing method described in any one of claims 1-6.

8. The background point cloud processing system according to claim 7, characterized in that... The system includes a background point cloud acquisition module, a segmentation module, a voxel sampling module, and a conversion output module. The background point cloud acquisition module acquires the background point cloud; the segmentation module segments the background point cloud along the XYZ axes; the voxel sampling module calculates the normal vector and curvature value of the background point cloud, divides the point cloud into high-curvature and low-curvature point sets according to a set curvature threshold, sets a differentiated voxel sampling resolution based on the proportion of high-curvature points within each segment, and performs voxel sampling on each segment using this differentiated resolution to obtain a voxel-sampled point cloud that meets the required number of points; the conversion output module converts the voxel-sampled point cloud into an octomap object and outputs it.

9. The background point cloud processing system according to claim 8, characterized in that... The background point cloud acquisition module includes a point cloud merging module, a point cloud deduplication module, and a point cloud filtering module. The point cloud merging module merges the overall point cloud and the sub-module point clouds into a superimposed point cloud. The point cloud deduplication module constructs a KD-tree spatial index, sets a tolerance threshold, detects overlapping points in the superimposed point cloud through radius search, removes all overlapping points, and generates an initial background point cloud composed of non-overlapping points. The point cloud filtering module uses a statistical outlier removal filter to filter the initial background point cloud, removes outliers, and forms the background point cloud.

10. The background point cloud processing system according to claim 9, characterized in that... The background point cloud acquisition module also includes a sub-module point cloud module, which is used to generate a sub-module point cloud of the target based on the overall point cloud.

Images