A Smart Registration Method for Structured Light Point Clouds of Sound Barriers Based on Multi-Scale ICP

Abstract

This invention discloses an intelligent registration method for structured light point clouds of sound barriers based on multi-scale ICP. The method includes acquiring structured light point cloud data of sound barriers, denoising the point cloud data, filtering the point cloud data, and multi-scale ICP registration of the sound barrier point cloud. This invention relates to the field of point cloud data processing technology, specifically to an intelligent registration method for structured light point clouds of sound barriers based on multi-scale ICP. This scheme innovatively adopts a denoising method based on adaptive neighborhood scale factor calculation and neighborhood density calculation, which significantly improves the denoising efficiency and accurately distinguishes between main points and noise points. It also innovatively adopts a progressive filtering strategy to achieve full and accurate removal of residual clutter in non-main areas. Furthermore, it innovatively adopts a progressive registration strategy that includes bounding box alignment, large-scale offset detection in the Y direction, percentage point clipping in the X direction, anchoring region extraction, multi-scale ICP registration, and complete transformation application, which effectively improves the registration matching accuracy and enhances the stability of the registration process.

Description

Technical Field

[0001] This invention relates to the field of point cloud data processing technology, specifically to a method for intelligent registration of structured light point clouds of sound barriers based on multi-scale ICP. Background Technology

[0002] In the long-term structural health monitoring of rail transit sound barriers, a structured light 3D vision measurement system mounted on a mobile operating platform is used to collect 3D point cloud data of the sound barrier surface at different times. Due to differences in the installation of the detection equipment, it is necessary to filter, reconstruct, and register the point cloud data at different times in order to construct a 3D data model that truly reflects the spatial geometric features of the sound barrier. Therefore, a multi-scale ICP-based intelligent registration method for sound barrier structured light point clouds is proposed. This method can achieve accurate matching of sound barrier point clouds at different times, laying the foundation for studying the spatiotemporal structural changes of sound barriers.

[0003] However, traditional noise reduction algorithms for main body point clouds of sound barriers suffer from technical problems such as poor adaptability to changes in spatial scale, reliance on manually set noise removal thresholds, and high computational complexity. These problems easily lead to inaccurate removal of low-density isolated noise points and local discrete clutter points, thereby reducing the accuracy and stability of subsequent point cloud filtering and registration. In traditional sound barrier point cloud filtering methods, there is a technical problem that the spatial range of the main body of the sound barrier cannot be accurately defined for residual clutter in non-main body areas, invalid ground points, and low-density isolated areas in the point cloud. This results in incomplete filtering of the main body point cloud, leading to mismatches and decreased registration accuracy. Traditional sound barrier point cloud registration algorithms suffer from technical problems such as large data processing volume, high computational complexity, and lack of effective initial attitude calibration, large-scale offset correction, and stable registration area screening, resulting in excessively long registration calculation time and low registration accuracy. Summary of the Invention

[0004] To address the above issues and overcome the shortcomings of existing technologies, this invention provides an intelligent registration method for structured light point clouds of sound barriers based on multi-scale ICP. Traditional sound barrier main point cloud denoising algorithms suffer from poor adaptability to spatial scale changes, reliance on manually set noise removal thresholds, and high computational complexity, easily leading to inaccurate removal of low-density isolated noise points and local discrete clutter points, thus reducing the accuracy and stability of subsequent point cloud filtering and registration. This solution innovatively adopts an integrated denoising method that combines local point cloud partitioning, adaptive neighborhood scale factor calculation, and adaptive neighborhood density calculation. This significantly reduces the computational and spatial complexity of the algorithm, significantly improves denoising efficiency, accurately distinguishes between main points and noise points, reduces false noise point identification and false deletion of valid main points, enhances the targeting and robustness of noise identification, and greatly improves the purity of the sound barrier main point cloud. It achieves low-complexity, high-precision denoising of point clouds under different detection distances and sparsity, providing a clean point cloud foundation with no noise interference, structural integrity, and stable features for subsequent multi-scale ICP registration, ultimately improving the accuracy, stability, and robustness of point cloud registration. In sound barrier point cloud filtering methods, there is a technical problem that residual clutter in non-main areas, invalid ground points, and low-density isolated regions remain in the point cloud, making it impossible to accurately define the spatial range of the main sound barrier. This leads to incomplete filtering of the main point cloud, resulting in mismatched registration and decreased registration accuracy. This solution innovatively adopts a progressive filtering strategy, including depth clipping, intensity clipping, sound barrier edge detection, depth histogram clipping, density clipping, and salt-and-pepper noise filtering, to perform full-dimensional filtering on both the denoised reference and target main point cloud sets. The progressive screening and purification process effectively solves the technical problems of traditional point cloud filtering, such as weak targeting, incomplete purification, and inability to accurately distinguish between effective main points and residual noise. It accurately defines the final spatial range of the main body of the sound barrier, effectively eliminates various residual noise interferences, significantly improves the purity of the main point cloud, strengthens the structural integrity and characteristic stability of the main point cloud of the sound barrier, achieves full-dimensional accurate removal of residual clutter in non-main areas, avoids the adverse effects of invalid clutter on registration, and ultimately improves the accuracy, stability and robustness of point cloud registration.To address the technical problems of traditional sound barrier point cloud registration algorithms, such as large data processing volume, high computational complexity, and lack of effective initial attitude calibration, large-scale offset correction, and stable registration region selection, leading to excessively long registration calculation time and low registration accuracy, this solution innovatively adopts a progressive registration strategy consisting of bounding box alignment, large-scale offset detection in the Y direction, percentile point clipping in the X direction, anchoring region extraction, multi-scale ICP registration, and complete transformation application. This effectively reduces the computational complexity and power consumption of the registration algorithm, significantly improves the registration convergence speed and matching accuracy, enhances the stability and robustness of the registration process, and accurately avoids registration inaccuracies caused by initial position deviations and outliers. It can flexibly adapt to the registration requirements of sound barrier point clouds with different detection distances and sparsities, ultimately achieving fast and accurate global registration between the sound barrier reference point cloud and the target point cloud.

[0005] The technical solution adopted by this invention is as follows: The intelligent registration method for structured light point clouds of sound barriers based on multi-scale ICP provided by this invention includes the following steps:

[0006] Step S1: Acquisition of structured light point cloud data for the sound barrier;

[0007] Step S2: Denoising point cloud data;

[0008] Step S3: Point cloud data filtering;

[0009] Step S4: Multi-scale ICP registration of sound barrier point cloud.

[0010] Further, in step S1, the acquisition of the sound barrier structured light point cloud data specifically involves collecting point cloud data to obtain the original sound barrier structured light point cloud data, performing point cloud data preprocessing on the original sound barrier structured light point cloud data to obtain standardized sound barrier structured light point cloud data, and then dividing the standardized sound barrier structured light point cloud data into two sets of point cloud sets to be registered, which are respectively denoted as the reference point cloud set and the target point cloud set.

[0011] Point cloud data preprocessing specifically includes coordinate system calibration and format normalization.

[0012] Furthermore, in step S2, the denoising of the main point cloud data specifically includes the following steps:

[0013] Step S21: Coarse extraction of the main point cloud of the sound barrier. Specifically, based on the total length of the sound barrier in the detection section and the nominal detection distance between the structured light acquisition device and the sound barrier, the value ranges of the X, Y, and Z axes of the global detection coordinate system of the sound barrier are set respectively. According to the value ranges of the three axes, the point clouds in the reference point cloud set and the target point cloud set in the standardized sound barrier structured light point cloud data are marked as the temporary internal point cloud of the main body and the temporary external point cloud of the main body, respectively, to obtain the coarsely extracted reference main body point cloud set and target main body point cloud set.

[0014] Step S22: Construct a denoising algorithm for the main point cloud data, which specifically includes the following steps:

[0015] Step S221: Point cloud local block division, specifically, the input main point cloud set is divided at equal intervals along the Y-axis direction of the global detection coordinate system of the sound barrier with a preset step size to obtain N mutually independent point cloud local blocks;

[0016] Step S222: Calculate the adaptive neighborhood scaling factor. Specifically, for each local block of the point cloud, calculate the difference between the maximum and minimum values ​​of its coordinates in the X, Y, and Z axes of the global detection coordinate system of the sound barrier. Take the maximum value among the three as the maximum size of the local block in three-dimensional space, and then calculate the adaptive neighborhood scaling factor of the local block based on the maximum size in three-dimensional space. Finally, the adaptive neighborhood scale factor for each local block of the point cloud is calculated.

[0017] Step S223: Calculate the adaptive neighborhood density. Specifically, construct a KD tree structure for each local block of the point cloud. Then, for any point within a single local block of the point cloud, use the adaptive neighborhood scale factor of the local block of the point cloud as the radius of the spherical neighborhood. Query the number of all effective neighboring points within the spherical neighborhood using the KD tree. Then, calculate the adaptive neighborhood density value of the point based on the number of effective neighboring points and the volume of the spherical neighborhood. Finally, complete the adaptive neighborhood density calculation for each point within all local blocks of the point cloud to obtain the adaptive neighborhood density value corresponding to each point in the point cloud set.

[0018] Step S224: Noise point removal, specifically, for each local block of the point cloud, calculate the adaptive removal threshold for that local block based on its corresponding adaptive neighborhood scale factor. Then, the adaptive neighborhood density within the local block is calculated. Points that are not identified as noise points are removed, while points that are retained are... After removing noise points from all local blocks of the sound barrier's main point cloud, the remaining point clouds from all local blocks are fused and stitched together to obtain the denoised main point cloud set; the formula used is as follows:

[0019] ;

[0020] In the formula, This indicates the total number of points within a local block of the point cloud. This represents the adaptive rejection threshold adjustment coefficient;

[0021] Step S23: Data denoising, specifically, inputting the coarsely extracted reference subject point cloud set and target subject point cloud set into the point cloud data denoising algorithm to obtain the denoised reference subject point cloud set and the denoised target subject point cloud set.

[0022] Furthermore, in step S3, the point cloud data filtering specifically includes the following steps:

[0023] Step S31: Depth clipping, specifically, based on the nominal detection distance between the structured light acquisition device and the sound barrier, combined with the actual ranging accuracy of the structured light acquisition device, a depth tolerance is set, and then the Z-axis depth clipping interval is set according to the nominal detection distance and the depth tolerance. Finally, based on the depth clipping interval, all point clouds in the main point cloud set that exceed the interval are removed.

[0024] Step S32: Intensity value clipping, specifically, calculating the structured light reflection intensity values ​​of all the main point cloud sets after depth clipping, and taking the intensity values. Quantiles as the lower limit threshold of intensity Intensity values ​​in the main point cloud set after depth clipping Point cloud, retain intensity values The point cloud, among which, Indicates the intensity quantile coefficient;

[0025] Step S33: Based on the edge detection of the sound barrier, specifically, the point cloud set of the main body point cloud after the intensity value is clipped is divided into B independent local blocks along the Y-axis with a fixed step size. Then, edge detection is performed on each local block. Finally, all the local blocks after edge detection are merged to obtain the main body point cloud set after edge detection.

[0026] The edge detection specifically involves traversing the point cloud within a single local block along the X-axis from the maximum to the minimum coordinate value, and calculating the structured light reflection intensity value of all point clouds in each row. And take the maximum sum of the intensity values ​​of all rows within the local block as the value. When h consecutive rows of all point clouds When the x-axis coordinates of the last row in a series of h rows are taken as the starting boundary of the sound barrier on the x-axis, then the boundary is determined by the x-axis coordinates of the last row in the series h rows. Then, for all rows within the same local block, traverse along the X-axis from the minimum to the maximum coordinate, calculating the intensity value of all point clouds in that row. And take the maximum sum of the intensity values ​​of all rows within the local block as the value. When h consecutive rows of all point clouds When the x-axis coordinate of the last row in a series of h rows is taken as the x-axis termination boundary of the sound barrier, then the x-axis coordinate of the last row is taken as the termination boundary of the sound barrier. Finally, based on this local block and Perform point cloud cropping, retaining the desired result. Point clouds; among them, This represents the initial boundary strength threshold coefficient. This represents the termination boundary strength threshold coefficient, and h represents the number of consecutive decision rows. Represents the X-axis coordinate value;

[0027] Step S34: Depth histogram cropping, specifically, cropping the main point cloud set after edge detection by layering it at equal intervals along the Z-axis. Perform layering and construct Z-axis coordinate values. Find the histogram distribution and locate the depth value corresponding to the peak value of the histogram. , and then with Define an asymmetric filtering window for the reference. Finally, the point cloud within the window is retained, while residual noise in the depth direction outside the interval is removed; among them, and These represent the lower and upper tolerances of the histogram clipping depth, respectively.

[0028] Step S35: Density clipping, specifically, projecting the point cloud from the main point cloud set after depth histogram clipping onto the Y-axis, dividing the Y-axis into equal segments according to segment length, and calculating the point cloud density of each Y-axis segment. Then, take the maximum point cloud density among all segments as the value. Then, the density clipping threshold is calculated based on the maximum point cloud density and the density scaling factor. Finally, retain the ones that satisfy the condition. And continuous Point clouds of Y-axis segments where all segments exceed the density clipping threshold, among which, This represents the threshold for determining consecutive valid segments, and its value is a positive integer.

[0029] Step S36: Salt and pepper noise filtering, specifically, performing median filtering and morphological opening operation combination operations sequentially on the density-trimmed main point cloud set to filter salt and pepper noise points, finally obtaining a clean main point cloud set;

[0030] Furthermore, in step S4, the multi-scale ICP registration of the sound barrier point cloud specifically includes the following steps:

[0031] Step S41: Bounding box alignment, specifically, constructing 3D axis-aligned bounding boxes for the clean reference subject point cloud set and the clean target subject point cloud set respectively, extracting the lower left corner vertex of the two bounding boxes as the registration reference point, translating the bounding box reference point of the target subject point cloud to the bounding box reference point of the reference subject point cloud, completing the initial position alignment of the two subject point cloud sets, and obtaining the target subject point cloud set after bounding box alignment;

[0032] Step S42: Large-scale offset detection in the Y direction, specifically, extracting points along the X-axis from the clean reference point cloud set and the target point cloud set after bounding box alignment. , , The three cross-sections of the location form a three-dimensional slice region. Then, the root mean square error of the target subject point cloud relative to the reference subject point cloud within each slice region is calculated. If the average root mean square error value of the three slice regions If the offset exceeds the preset threshold, the target point cloud is determined to have a large-scale offset along the Y-axis. Subsequently, the target point cloud is translated along the Y-axis according to a preset translation step size, and the average root mean square error value after each translation is calculated. The translation amount corresponding to the minimum average root mean square error value is taken as the optimal offset, and the target subject point cloud is translated and corrected to obtain the target subject point cloud set after large-scale offset.

[0033] Step S43: X-axis percentile cropping. Specifically, for the target subject point cloud set after large-scale migration and the clean reference subject point cloud set, the percentile distribution of their coordinate values ​​in the X-axis direction is calculated, and the points are cropped and removed if the X-axis coordinates are less than a certain value. quantile greater than Extreme outlier point clouds of quantiles, preserving the X-axis direction. The point cloud set within the quantile range yields the target subject point cloud set and the reference subject point cloud set after percentile cropping. This indicates the lower percentile of the percentile retention interval. This indicates the upper percentile of the percentile retention interval;

[0034] Step S44: Anchoring region extraction. Specifically, for the point clouds of the target subject point cloud set and the reference subject point cloud set after percentage cropping, the minimum value of the point cloud in its own X-axis direction is used as the benchmark, and the point clouds are extracted from the bottom according to the height percentage. to Point clouds within the height range are used as registration anchoring regions, resulting in point cloud sets for the target subject's anchoring region and reference subject's anchoring region. This indicates the lower limit of the anchoring area in the X-axis direction. This indicates the upper limit of the anchored area in the X-axis direction;

[0035] Step S45: Multi-scale ICP registration, specifically, using the reference subject anchoring region point cloud set as the baseline and the target subject anchoring region point cloud set as the registration object, and performing registration according to the downsampling ratio. , and Simultaneously downsampling is performed on the point cloud set of the target anchoring region and the point cloud set of the reference anchoring region to obtain three sets of downsampled point cloud pairs, and an offset radius limit is set for the three sets of downsampled point cloud pairs. , and Then, weighted ICP registration calculations are performed on each set of downsampled point cloud pairs from coarse to fine scales, and finally the optimal three-dimensional rotation matrix and the optimal three-dimensional registration translation vector of the anchoring region point cloud set are obtained.

[0036] Step S46: Complete transformation application, specifically based on the optimal 3D rotation matrix and optimal 3D registration translation vector of the anchored region point cloud set, the transformation parameters are applied to the target subject point cloud set after percentile clipping through rigid body space transformation, global spatial attitude correction is performed on the complete target point cloud, and finally the registered target subject point cloud set is obtained, realizing the matching of the two point cloud sets.

[0037] The beneficial effects achieved by the present invention using the above solution are as follows:

[0038] (1) In view of the technical problems of traditional noise barrier main point cloud denoising algorithms, such as poor adaptability to changes in spatial scale, reliance on manual setting of noise removal threshold, and high computational complexity, it is easy to cause inaccurate removal of low-density isolated noise points and local discrete clutter points, thereby reducing the accuracy and stability of subsequent point cloud filtering and registration. This solution innovatively adopts an integrated denoising method of point cloud local block division, adaptive neighborhood scale factor calculation, and adaptive neighborhood density calculation, which greatly reduces the computational and spatial complexity of the algorithm, significantly improves the denoising efficiency, accurately distinguishes between main points and noise points, reduces the misjudgment of noise points and the false deletion of effective main points, enhances the pertinence and robustness of noise identification, and greatly improves the purity of the main point cloud of the noise barrier. It realizes low-complexity and high-precision denoising of point clouds under different detection distances and different sparsity, and provides a pure point cloud foundation with no noise interference, complete structure, and stable features for subsequent multi-scale ICP registration, ultimately improving the accuracy, stability and robustness of point cloud registration.

[0039] (2) To address the technical problem in traditional sound barrier point cloud filtering methods where residual clutter, invalid ground points, and low-density isolated areas remain in the point cloud, making it impossible to accurately define the spatial range of the sound barrier's main body, thus leading to incomplete filtering of the main body point cloud and consequently registration mismatch and decreased registration accuracy, this solution innovatively adopts a progressive filtering strategy consisting of depth clipping, intensity clipping, sound barrier edge detection, depth histogram clipping, density clipping, and salt-and-pepper noise filtering. This strategy is used to filter the denoised reference main body point cloud set and the target main body point cloud set. The comprehensive and progressive screening and purification effectively solves the technical problems of traditional point cloud filtering, such as weak targeting, incomplete purification, and inability to accurately distinguish between effective main points and residual noise. It accurately defines the final spatial range of the main body of the sound barrier, effectively eliminates various residual noise interferences, significantly improves the purity of the main point cloud, strengthens the structural integrity and characteristic stability of the main point cloud of the sound barrier, achieves comprehensive and accurate removal of residual clutter in non-main areas, avoids the adverse effects of invalid clutter on registration, and ultimately improves the accuracy, stability and robustness of point cloud registration.

[0040] (3) In view of the technical problems of traditional sound barrier point cloud registration algorithms, such as large data processing volume, high computational complexity, lack of effective initial attitude calibration, large-scale offset correction and stable registration region screening, resulting in long registration calculation time and low registration accuracy, this solution innovatively adopts a progressive registration strategy of bounding box alignment, large-scale offset detection in the Y direction, percentage point clipping in the X direction, anchoring region extraction, multi-scale ICP registration and complete transformation application. This effectively reduces the computational complexity and computing power consumption of the registration algorithm, significantly improves the registration convergence speed and matching accuracy, enhances the stability and robustness of the registration process, and accurately avoids the registration misalignment problem caused by initial position deviation and outliers. It can flexibly adapt to the registration requirements of sound barrier point clouds with different detection distances and different sparsities, and finally realizes fast and accurate global registration of sound barrier reference point cloud and target point cloud. Attached Figure Description

[0041] Figure 1 A flowchart illustrating the intelligent registration method for structured light point clouds of a sound barrier based on multi-scale ICP provided by this invention.

[0042] Figure 2 This is a flowchart illustrating step S2;

[0043] Figure 3 This is a flowchart illustrating step S3;

[0044] Figure 4 This is a flowchart illustrating step S4;

[0045] Figure 5 This is a flowchart illustrating step S22;

[0046] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0047] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0048] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the system or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0049] Example 1, see Figure 1 The technical solution adopted by this invention is as follows: The intelligent registration method for structured light point clouds of sound barriers based on multi-scale ICP provided by this invention includes the following steps:

[0050] Step S1: Acquisition of structured light point cloud data for the sound barrier, specifically by acquiring and preprocessing point cloud data to obtain a reference point cloud set and a target point cloud set;

[0051] Step S2: Point cloud data denoising, used to remove low-density isolated noise points and local discrete clutter points in the main body area of ​​the sound barrier in the standardized sound barrier structured light point cloud data, eliminate noise interference in the point cloud body, improve the purity of the point cloud data, and at the same time take into account the denoising efficiency to adapt to the hardware computing power limitations of the mobile detection platform, providing high-quality basic point cloud data for subsequent point cloud filtering and registration; specifically, firstly, the main body point cloud of the sound barrier is coarsely extracted from the reference point cloud set and the target point cloud set, then the main body point cloud data denoising algorithm is constructed by dividing the point cloud into local blocks, calculating the adaptive neighborhood scale factor, calculating the adaptive neighborhood density and noise point removal, and finally, the coarsely extracted reference main body point cloud set and target main body point cloud set are input into the point cloud data denoising algorithm to perform denoising processing, resulting in the denoised reference main body point cloud set and the denoised target main body point cloud set;

[0052] Step S3: Point cloud data filtering, used to accurately remove non-target noise from the denoised reference subject point cloud set and the denoised target subject point cloud set in all dimensions, to finally define the spatial range of the sound barrier subject and to purify residual noise, to completely separate the subject point cloud from the surrounding clutter, invalid ground points, and low-density isolated areas, and to eliminate residual salt-and-pepper noise interference, so as to obtain a high-purity subject point cloud set containing only the effective structural features of the sound barrier. Specifically, a progressive filtering strategy is adopted, which includes depth clipping, intensity value clipping, sound barrier edge detection, depth histogram clipping, density clipping and salt-and-pepper noise filtering. The denoised reference subject point cloud set and the target subject point cloud set are screened and filtered layer by layer in turn, and finally the filtered pure reference subject point cloud set and the pure target subject point cloud set are obtained.

[0053] Step S4: Multi-scale ICP registration of the sound barrier point cloud. This step is used to perform high-precision global registration between the clean reference subject point cloud set and the clean target subject point cloud set, achieving fast and accurate matching of the sound barrier structured light point cloud. Specifically, a progressive registration strategy is adopted, which includes bounding box alignment, large-scale offset detection in the Y direction, percentage point cropping in the X direction, anchoring region extraction, multi-scale ICP registration, and complete transformation application, to achieve registration between the clean target subject point cloud set and the clean reference subject point cloud set, ultimately obtaining the registered sound barrier target subject point cloud.

[0054] Example 2, see Figure 1 This embodiment is based on the above embodiment. In step S1, the acquisition of sound barrier structured light point cloud data is used to collect and preprocess standardized sound barrier structured light point cloud data. Specifically, the structured light three-dimensional vision measurement system is mounted on a mobile detection platform to scan and sample the sound barrier of the detection section at different times, collect point cloud data, and obtain the original sound barrier structured light point cloud data containing three-dimensional spatial coordinate information and structured light reflection intensity value information. The original sound barrier structured light point cloud data is then preprocessed to obtain standardized sound barrier structured light point cloud data. The standardized sound barrier structured light point cloud data is then divided into two sets of point cloud sets to be registered, which are denoted as the reference point cloud set and the target point cloud set, respectively.

[0055] Point cloud data preprocessing is used to eliminate coordinate system deviations and format heterogeneity issues in point cloud data under different acquisition times and device postures, ensuring the consistency of subsequent point cloud denoising, filtering, and registration algorithms in data parsing and calculation. Specifically, it includes coordinate system uniform calibration and format normalization processing.

[0056] The coordinate system calibration is used to unify the spatial reference of point cloud data at different acquisition times, and eliminate the inconsistency of coordinate systems caused by the position offset of the detection platform and the deviation of the equipment installation angle. Specifically, it transforms the local coordinate system of the acquisition equipment of the original structure light point cloud data of each batch of sound barriers to the preset global detection coordinate system of the sound barrier. The coordinate transformation is completed by the rigid body space transformation formula, so that the three-dimensional coordinate values ​​of all point cloud data are based on the same spatial reference, and the unification of the point cloud spatial reference is achieved.

[0057] The global detection coordinate system for the sound barrier is specifically defined with the origin at the intersection of the bottom of the sound barrier and the ground at the starting point of the detection section, the Y-axis along the extension direction of the sound barrier, the X-axis perpendicular to the ground and upward, and the Z-axis in the depth direction, with the unit uniformly in millimeters; wherein, the Z-axis coordinate value represents the vertical straight-line distance from a point on the sound barrier column structure component to the origin in the depth direction of the Z-axis.

[0058] The format normalization process is used to solve the problem of heterogeneous point cloud data formats output by different acquisition devices and software, and to ensure the uniformity of data storage structure and feature fields. Specifically, it converts the original structured light point cloud data of the sound barrier into the industrially common PCD point cloud format, and standardizes the field definitions of the point cloud data so that each point cloud data contains a three-dimensional coordinate field and an intensity value field, and the field encoding method and data type are consistent.

[0059] Example 3, see Figure 1 , Figure 2 and Figure 5 This embodiment is based on the above embodiment. In step S2, the main body point cloud data denoising is used to remove low-density isolated noise points and local discrete clutter points in the main body area of ​​the sound barrier in the standardized sound barrier structured light point cloud data, eliminate noise interference in the point cloud body, improve the purity of the point cloud data, and at the same time take into account the denoising processing efficiency to adapt to the hardware computing power limitations of the mobile detection platform, providing high-quality basic point cloud data for subsequent point cloud filtering and registration. Specifically, the main body point cloud is first coarsely extracted from the reference point cloud set and the target point cloud set. Then, the main body point cloud data denoising algorithm is constructed by dividing the point cloud into local blocks, calculating the adaptive neighborhood scale factor, calculating the adaptive neighborhood density, and removing noise points. Finally, the coarsely extracted reference main body point cloud set and target main body point cloud set are input into the point cloud data denoising algorithm to perform denoising processing, resulting in the denoised reference main body point cloud set and the denoised target main body point cloud set. The steps include:

[0060] Step S21: Coarse extraction of the point cloud of the main body of the sound barrier, used to quickly delineate the approximate spatial range of the main body of the sound barrier, reducing the interference of non-target noise points on the calculation of neighborhood features during subsequent denoising processes; specifically, based on the total length of the sound barrier of the detected road segment. The nominal detection distance between the structured light acquisition equipment and the sound barrier The X, Y, and Z axes of the global detection coordinate system for the sound barrier are set with value ranges respectively. Based on the value ranges of the three axes, the point clouds in the reference point cloud set and the target point cloud set in the standardized sound barrier structured light point cloud data are marked as the point cloud inside the temporary subject and the point cloud outside the temporary subject, respectively, to obtain the coarsely extracted reference subject point cloud set and target subject point cloud set. Among them, the point cloud inside the temporary subject is the point cloud of the area where the sound barrier subject is located, which is the effective object of subsequent denoising processing. The point cloud outside the temporary subject corresponds to the point cloud of the external noise area that is not the sound barrier subject, and does not participate in the subsequent denoising calculation. The temporary coarse judgment is only marked.

[0061] The set value range is specifically the value range in the X-axis direction. Y-axis direction value range Z-axis direction value range Point clouds that simultaneously satisfy the above three-axis value ranges are marked as point clouds inside the provisional subject, and point clouds that do not satisfy the value range of any of the above axes are marked as point clouds outside the provisional subject. and These represent the minimum and maximum values ​​for coarse extraction along the X-axis, respectively. Indicates the value in the X-axis direction. Indicates the value in the Y-axis direction. This indicates the allowable deviation in the Z-axis direction. Indicates the value in the Z-axis direction;

[0062] Step S22: Construct a main point cloud data denoising algorithm to perform fine denoising processing on the coarsely extracted main point cloud of the sound barrier, specifically eliminating low-density isolated noise points and local discrete clutter points within the main area; specifically including the following steps:

[0063] Step S221: Point cloud local block division, used to adapt to the structural features of the sound barrier extending linearly along the Y-axis, avoids feature calculation distortion caused by the difference in sparsity of the near and far points of the sound barrier when calculating the overall scale and density of the reference main point cloud set and the target main point cloud set after coarse extraction, improves the accuracy of subsequent neighborhood scale and density calculation, and realizes the block parallel processing of point cloud to improve the overall efficiency of noise reduction. Specifically, the input main point cloud set is divided into N independent point cloud local blocks along the Y-axis direction of the global detection coordinate system of the sound barrier at a preset step size. Then, each point cloud local block is uniquely identified and its contained point cloud data is recorded. Finally, the block-divided point cloud local block set is obtained.

[0064] The preset step size is adapted to the point cloud collection density of the actual detection section of the sound barrier.

[0065] Step S222: Calculate the adaptive neighborhood scale factor to dynamically match the optimal neighborhood calculation scale for each point cloud local block. This avoids the problem of not being able to simultaneously adapt to the point cloud features of sparse and dense local blocks when using a fixed neighborhood scale, thus improving the targeting and accuracy of subsequent neighborhood density calculations. Specifically, for each point cloud local block in the set of point cloud local blocks, calculate the difference between the maximum and minimum values ​​of its coordinates in the X, Y, and Z axes of the global detection coordinate system of the sound barrier. Take the maximum value of the three as the maximum size of the point cloud local block in three-dimensional space, and then calculate the adaptive neighborhood scale factor of the point cloud local block based on the maximum size in three-dimensional space. Finally, calculate the adaptive neighborhood scale factor for each point cloud local block. The formula used is as follows:

[0066] ;

[0067] ;

[0068] In the formula, This represents the maximum three-dimensional size of a local block in the point cloud. and These represent the maximum and minimum coordinates of the point cloud within a local block along the X-axis, respectively. and These represent the maximum and minimum values ​​of the point cloud coordinates on the Y-axis within the local block, respectively. and These represent the maximum and minimum coordinates of the point cloud within the local block along the Z-axis, respectively. This represents the adaptive neighborhood scaling factor of a local patch in a point cloud. This is represented as a scaling factor, with a value range of [value range missing]. ;

[0069] Step S223: Calculate the adaptive neighborhood density to quantify the spatial distribution characteristics of each point within a local block of the point cloud, enabling feature differentiation between the main point cloud of the sound barrier and low-density noise points. Simultaneously, based on the fast neighborhood query mechanism of the KD tree, the overall complexity of the neighborhood density calculation is controlled within a certain range. Balancing computational accuracy and processing efficiency, and adapting to the hardware computing power limitations of mobile detection platforms, this involves constructing a KD-tree structure for each local point cloud block in the set of local point cloud blocks, and then processing any point within a single local point cloud block. Using the adaptive neighborhood scale factor of the local patch of the point cloud as the radius of the spherical neighborhood, the number of all valid neighboring points within the spherical neighborhood is queried using a KD-tree. Then, based on the number of effective neighboring points and the volume of the spherical neighborhood, the adaptive neighborhood density value of the point is calculated. Finally, the adaptive neighborhood density calculation of each point in each local block of the point cloud is completed, and the adaptive neighborhood density value corresponding to each point in the point cloud set is obtained.

[0070] The number of effective neighboring points Specifically, it involves statistically analyzing the spherical neighborhood excluding the target point. The number of effective neighboring point clouds outside itself;

[0071] The formula used is as follows:

[0072] ;

[0073] In the formula, Represents the volume of a spherical neighborhood;

[0074] Step S224: Noise point removal, used to accurately identify and remove low-density isolated noise points and local discrete clutter points in the reference point cloud set and the target point cloud set based on adaptive neighborhood density features. Specifically, for each point cloud local block in the point cloud local block set, the adaptive removal threshold of the point cloud local block is calculated based on its corresponding adaptive neighborhood scale factor. Then, the adaptive neighborhood density within the local block is calculated. Points that are identified as noise points are removed, while points that are retained are... After removing noise points from all local blocks of the sound barrier's main point cloud, the remaining point clouds from all local blocks are fused and stitched together to obtain the denoised main point cloud set; the formula used is as follows:

[0075] ;

[0076] In the formula, This indicates the total number of points within a local block of the point cloud. This represents the adaptive rejection threshold adjustment coefficient, with a value range of [value range missing]. ;

[0077] Step S23: Data denoising, specifically, inputting the coarsely extracted reference subject point cloud set and target subject point cloud set into the point cloud data denoising algorithm to obtain the denoised reference subject point cloud set and the denoised target subject point cloud set.

[0078] By performing the above operations, this solution addresses the technical problems of traditional sound barrier main point cloud denoising algorithms, such as poor adaptability to changes in spatial scale, reliance on manually set noise removal thresholds, and high computational complexity. These problems easily lead to inaccurate removal of low-density isolated noise points and local discrete clutter points, thereby reducing the accuracy and stability of subsequent point cloud filtering and registration. This solution innovatively adopts an integrated denoising method that combines point cloud local block partitioning, adaptive neighborhood scale factor calculation, and adaptive neighborhood density calculation. This significantly reduces the computational and spatial complexity of the algorithm, significantly improves denoising efficiency, accurately distinguishes between main points and noise points, reduces false noise point judgments and false deletion of valid main points, enhances the targeting and robustness of noise identification, and greatly improves the purity of the sound barrier main point cloud. It achieves low-complexity, high-precision denoising of point clouds under different detection distances and sparsity, providing a clean point cloud foundation with no noise interference, complete structure, and stable features for subsequent multi-scale ICP registration, ultimately improving the accuracy, stability, and robustness of point cloud registration.

[0079] Example 4, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In step S3, the point cloud data filtering is used to perform full-dimensional non-target noise removal, final definition of the sound barrier's main spatial range, and residual noise purification on the denoised reference subject point cloud set and the denoised target subject point cloud set. This completely separates the main point cloud from peripheral clutter, invalid ground points, and low-density isolated areas, eliminates residual salt-and-pepper noise interference, and obtains a high-purity main point cloud set containing only the effective structural features of the sound barrier. Specifically, a progressive filtering strategy is adopted, including depth clipping, intensity value clipping, sound barrier edge detection, depth histogram clipping, density clipping, and salt-and-pepper noise filtering. The denoised reference subject point cloud set and the target subject point cloud set are sequentially screened and filtered layer by layer to finally obtain the filtered pure reference subject point cloud set and the pure target subject point cloud set. The steps include:

[0080] Step S31: Depth trimming, used to quickly remove distant noise points and too-close invalid points far from the main body of the sound barrier along the Z-axis depth direction, initially narrowing the processing range for subsequent filtering; specifically, based on the nominal detection distance D between the structured light acquisition device and the sound barrier, and combined with the actual ranging accuracy of the structured light acquisition device, a depth tolerance is set. and Then, the Z-axis depth clipping range is set according to the nominal detection distance and depth tolerance. Finally, based on this depth clipping interval, all point clouds in the main point cloud set that exceed this interval are removed;

[0081] The main subject point cloud set includes a denoised reference main subject point cloud set and a target main subject point cloud set;

[0082] Step S32: Intensity value clipping, used to remove non-sound barrier material noise with excessively low structured light reflection intensity values, enhancing the feature recognition of the main point cloud; specifically, calculating the structured light reflection intensity values ​​of all the main point cloud sets after depth clipping, and taking the intensity value... Quantiles as the lower limit threshold of intensity Intensity values ​​in the main point cloud set after depth clipping Point cloud, retain intensity values The point cloud, among which, This represents the intensity quantile coefficient, with a value range of [value missing]. ;

[0083] Step S33: Based on sound barrier edge detection, this step accurately defines the effective edge of the sound barrier in the X-axis height direction, removing side noise and invalid extension points in the X-axis direction. Specifically, it involves cropping the intensity values ​​and then moving the point cloud in the main point cloud set along the Y-axis direction at a fixed step size. The system is divided into B independent local blocks at equal intervals. Then, edge detection is performed on each local block separately. Finally, all the local blocks after edge detection are merged to obtain the main point cloud set after edge detection.

[0084] The edge detection specifically involves traversing the point cloud within a single local block along the X-axis from the maximum to the minimum coordinate value, and calculating the structured light reflection intensity value of all point clouds in each row. And take the maximum sum of the intensity values ​​of all rows within the local block as the value. When h consecutive rows of all point clouds When the x-axis coordinates of the last row in a series of h rows are taken as the starting boundary of the sound barrier on the x-axis, then the boundary is determined by the x-axis coordinates of the last row in the series h rows. Then, for all rows within the same local block, traverse along the X-axis from the minimum to the maximum coordinate, calculating the point cloud values ​​for each row. And take the maximum sum of the intensity values ​​of all rows within the local block as the value. When h consecutive rows of all point clouds When the x-axis coordinate of the last row in a series of h rows is taken as the x-axis termination boundary of the sound barrier, then the x-axis coordinate of the last row is taken as the termination boundary of the sound barrier. Finally, based on this local block and Perform point cloud cropping, retaining the desired result. For point clouds, remove all point clouds within a local block that exceed the specified range; among them... This represents the initial boundary strength threshold coefficient, with a value range of [value missing]. , This represents the termination boundary strength threshold coefficient, with a value range of [value range missing]. h represents the number of consecutive rows to be evaluated, and its value range is... , Represents the X-axis coordinate value;

[0085] Each row is layered in the height direction of the local block point cloud based on a fixed step size along the X-axis, corresponding to a fixed height interval of the sound barrier; the formula used is as follows:

[0086] ;

[0087] In the formula, This represents the sum of the structured light reflection intensity values ​​of all point clouds in the j-th row. This indicates the total number of points contained in the point cloud. This represents the structured light reflection intensity value of the i-th point cloud in the j-th row;

[0088] Step S34: Depth histogram cropping, used to accurately locate the core area of ​​the sound barrier's main body in the Z-axis depth direction and remove residual non-target noise points in the depth direction; specifically, the point cloud set of the main body after edge detection is cropped in the Z-axis direction with equidistant layering steps. Perform layering and construct Z-axis coordinate values. Find the histogram distribution and locate the depth value corresponding to the peak value of the histogram. , and then with Define an asymmetric filtering window for the reference. Finally, the point cloud within the window is retained, while residual noise in the depth direction outside the interval is removed; among them, and These represent the lower and upper tolerances of the histogram clipping depth, respectively, with value ranges of [missing values]. and ;

[0089] Step S35: Density clipping, used to remove low-density isolated noise areas along the Y-axis extension of the sound barrier, ensuring the continuity of the main point cloud in the extension direction, avoiding misjudgment of isolated noise points, and achieving accurate definition of the main point cloud in the Y-axis direction; specifically, projecting the point cloud set of the main point cloud after depth histogram clipping onto the Y-axis direction, according to segment length Divide the Y-axis into equal segments and calculate the point cloud density for each segment. Then, take the maximum point cloud density among all segments as the value. Then, the density clipping threshold is calculated based on the maximum point cloud density and the density scaling factor. Finally, retain the ones that satisfy the condition. And continuous Point clouds of Y-axis segments where all segments exceed the density clipping threshold; among them... The range of values ​​is , This represents the density proportionality coefficient, and its value range is... , This represents the threshold for determining continuous valid segments, and its value is a positive integer, ranging from 1 to 2. ;

[0090] ;

[0091] In the formula, This represents the number of point clouds in the y-th Y-axis segment;

[0092] Step S36: Salt and pepper noise filtering, used to eliminate high-intensity salt and pepper noise introduced by factors such as sunlight and sensor sampling errors, purify residual noise in the main point cloud of the sound barrier, and improve the final purity of the point cloud; specifically, median filtering and morphological opening operation are performed sequentially on the density-trimmed main point cloud set to filter salt and pepper noise points, and finally a pure main point cloud set is obtained.

[0093] The clean subject point cloud set includes a clean reference subject point cloud set and a clean target subject point cloud set;

[0094] The median filtering is used to remove isolated high-intensity salt-and-pepper noise points; a 3×3×3 three-dimensional neighborhood window is used; the morphological opening operation is used to eliminate small connected regions formed by noise points, using a spherical structuring element with a radius of 3mm.

[0095] By performing the above operations, this solution addresses the technical problem in traditional sound barrier point cloud filtering methods: the inability to accurately define the spatial extent of the sound barrier's main body, resulting in incomplete filtering of the main body point cloud and consequently, mismatched registration and decreased registration accuracy, due to residual clutter in non-main areas, invalid ground points, and low-density isolated regions remaining in the point cloud. This is achieved by employing a progressive filtering strategy that combines depth clipping, intensity clipping, sound barrier edge detection, depth histogram clipping, density clipping, and salt-and-pepper noise filtering. This strategy is applied to both the denoised reference main body point cloud set and the target main body point cloud. This approach employs a comprehensive, layer-by-layer screening and purification process, effectively addressing the technical challenges of traditional point cloud filtering, such as insufficient targeting, incomplete purification, and the inability to accurately distinguish between effective points and residual noise. It precisely defines the final spatial range of the sound barrier's main body, effectively eliminates various residual noise interferences, significantly improves the purity of the main point cloud, strengthens the structural integrity and characteristic stability of the main point cloud, achieves comprehensive and precise removal of residual clutter in non-main areas, avoids the adverse effects of invalid clutter on registration, and ultimately improves the accuracy, stability, and robustness of point cloud registration.

[0096] Example 5, see Figure 1 and Figure 4This embodiment is based on the above embodiment. In step S4, the multi-scale ICP registration of the sound barrier point cloud is used to perform high-precision global registration of the clean reference subject point cloud set and the clean target subject point cloud set, realizing fast and accurate matching of the sound barrier structured light point cloud. Specifically, it adopts a progressive registration strategy of bounding box alignment, large-scale offset detection in the Y direction, percentage point cropping in the X direction, anchoring region extraction, multi-scale ICP registration, and complete transformation application to achieve the registration of the clean target subject point cloud set and the clean reference subject point cloud set, and finally obtain the registered sound barrier target subject point cloud; including the following steps:

[0097] Step S41: Bounding box alignment, used to provide a unified initial registration pose for the reference and target point clouds, eliminate the overall spatial offset of the point cloud caused by the device installation angle and the position deviation of the detection platform, and reduce the convergence difficulty of subsequent registration; specifically, three-dimensional axis-aligned bounding boxes are constructed for the clean reference subject point cloud set and the clean target subject point cloud set respectively, the lower left corner vertex of the two bounding boxes is extracted as the registration reference point, the bounding box reference point of the target subject point cloud is translated to the bounding box reference point of the reference subject point cloud, and the initial position alignment of the two subject point cloud sets is completed, resulting in the bounding box aligned target subject point cloud set;

[0098] Step S42: Large-scale offset detection in the Y direction, used to detect and correct the large-scale linear offset of the point clouds of the two sound barriers along the Y-axis, to avoid large offsets causing subsequent ICP registration to fall into a local optimum; specifically, extracting points along the X-axis from the clean reference point cloud set and the bounding box aligned target point cloud set. , , Three cross-sections of the location, each with a bandwidth of W, form a three-dimensional slice region. Then, the root mean square error of the target subject point cloud relative to the reference subject point cloud within each slice region is calculated. If the average root mean square error value of the three slice regions If the offset exceeds the preset threshold, the target point cloud is determined to have a large-scale offset along the Y-axis. Subsequently, the target point cloud is translated along the Y-axis according to a preset translation step size, and the average root mean square error value after each translation is calculated. The translation amount corresponding to the minimum average root mean square error is taken as the optimal offset, and the target point cloud is translated and corrected to obtain the target point cloud set after large-scale offset. , , Expressed as the cross-sectional extraction ratio, The value is 30%. The value is 50%. The value is 70%, and the range of W is [missing value]. millimeters;

[0099] ;

[0100] In the formula, This indicates the number of matching point pairs in the slice region. This represents the 3D coordinates of the i-th point within a 3D slice region of the clean reference subject point cloud. This indicates the area within the 3D slice of the target subject's point cloud after bounding box alignment. The 3D coordinates of the i-th matched point;

[0101] Step S43: X-direction percentile clipping is used to remove extreme outliers in the X-axis direction of the reference and target point clouds, avoiding interference from outliers on registration accuracy, ensuring the consistency of the effective area of ​​the two point clouds in the X-axis height direction, and improving the stability of the registration process. Specifically, for the target subject point cloud set after large-scale migration and the clean reference subject point cloud set, the percentile distribution of their coordinate values ​​in the X-axis direction is calculated, and points with X-axis coordinates less than a certain value are clipped and removed. quantile greater than Extreme outlier point clouds of quantiles, preserving the X-axis direction. The point cloud set within the quantile range yields the target subject point cloud set and the reference subject point cloud set after percentile cropping. This represents the lower percentile of the percentile retention interval, with a value of 0.5%. This represents the upper percentile of the percentile retention interval, with a value of 99.5%.

[0102] Step S44: Anchoring region extraction, used to extract stable registration regions of the sound barrier point cloud in the X-axis direction, avoiding the influence of dense noise areas at the bottom and easily deformable areas at the top on registration accuracy, and improving the reliability of the registration results. Specifically, for the point clouds in the target subject point cloud set and the reference subject point cloud set after percentage point cropping, the minimum value of their own X-axis coordinates is used as the benchmark, and the points are extracted from the bottom according to the height percentage. to The point cloud within the specified height range is used as the registration anchoring region, while other remaining height regions are not included in subsequent registration calculations. This yields the point cloud set of the target subject's anchoring region and the point cloud set of the reference subject's anchoring region. This represents the lower limit of the anchored area in the X-axis direction, with a value of 10%. This indicates the upper limit of the anchored area in the X-axis direction, with a value of 80%.

[0103] Step S45: Multi-scale ICP registration is used to achieve high-precision and fine-grained registration of point clouds within the anchoring region. Multi-scale downsampling balances registration efficiency and accuracy, avoiding the problem of single-scale ICP registration easily getting trapped in local optima and the difficulty in balancing efficiency and accuracy. Specifically, the reference subject anchoring region point cloud set is used as the baseline, and the target subject anchoring region point cloud set is used as the registration object. Registration is performed according to the downsampling ratio... , and Simultaneously downsampling is performed on the point cloud set of the target anchoring region and the point cloud set of the reference anchoring region to obtain three sets of downsampled point cloud pairs, and an offset radius limit is set for the three sets of downsampled point cloud pairs. , , Subsequently, weighted ICP registration calculations were performed on each set of downsampled point cloud pairs from coarse to fine scales, ultimately yielding the optimal 3D rotation matrix and the optimal 3D registration translation vector for the anchored region point cloud set; among which, The value is 20%. The value is 50%. The value is 100%. The value is 100 millimeters. The value is 50 mm. The value is 10 millimeters;

[0104] The initial iteration values ​​for the weighted ICP registration calculation are as follows: for coarse-scale registration, the unit 3D rotation matrix and the zero-registration translation vector are used as the initial iteration values; for meso-scale and fine-scale registration, the 3D rotation matrix and the 3D registration translation vector obtained from the previous level are used as the initial iteration values ​​for the current level of weighted ICP registration calculation.

[0105] The weighted ICP registration calculation is based on the iterative nearest point algorithm and introduces a depth-direction weighting mechanism. Specifically, given a reference point set and a target point set, based on a preset initial 3D rotation matrix and 3D registration translation vector, the nearest point is found for each point in the target point set within the offset radius limit of the reference point set, forming a matching point pair. The sum of the weighted Euclidean distances of all matching point pairs is used as the objective function. The 3D rotation matrix and 3D registration translation vector that minimize the objective function are solved by singular value decomposition. The target point set is then spatially transformed using these parameters. The above iterative process is repeated until the root mean square error difference between two adjacent iterations is less than a preset convergence threshold, thus completing the weighted ICP registration calculation.

[0106] The depth-direction weighting mechanism specifically assigns different weights to points at different locations based on the local variation features of the point cloud along the Z-axis depth direction.

[0107] The weight is calculated as follows: for each section in the point cloud, the standard deviation of its Z coordinate is calculated, and the standard deviation is used as the rate of change of the depth direction of the Z axis of the section, and as the weight coefficient of the target point. The larger the standard deviation of the Z coordinate, the more drastic the structural changes and the more significant the geometric features in the depth direction of the region, and the higher the weight is assigned. Flat regions with small standard deviations of the Z coordinate are assigned low weights to suppress the influence of noise.

[0108] The weighted Euclidean distance is calculated as follows: for each pair of matching points, first calculate its ordinary Euclidean distance, and then multiply the distance by the weight of the target point to obtain the weighted Euclidean distance of the pair.

[0109] The anchoring region point cloud set specifically includes the target subject anchoring region point cloud set and the reference anchoring region registration point cloud set;

[0110] The downsampling method is specifically voxel downsampling;

[0111] Step S46: Complete transformation application, used to apply the obtained optimal spatial transformation parameters to the complete target subject point cloud set, to achieve global high-precision registration between the reference point cloud and the target point cloud, ensuring that the registration result covers all valid point clouds; specifically, based on the optimal 3D rotation matrix and optimal 3D registration translation vector of the anchored region point cloud set, the transformation parameters are applied to the target subject point cloud set after percentile clipping through rigid body space transformation, and global spatial attitude correction is performed on the complete target point cloud, finally obtaining the registered target subject point cloud set that is globally registered with the reference subject point cloud, realizing the matching of the two point cloud sets;

[0112] The global spatial attitude correction specifically involves using the reference subject point cloud set after percentage point cropping as a fixed reference. This set does not undergo any spatial transformation. By performing a rigid body spatial transformation on the target subject point cloud set after percentage point cropping, the spatial attitude of the target subject point cloud is adjusted to a state that fully matches the reference subject point cloud, ultimately obtaining the registered target subject point cloud set that is globally registered with the reference subject point cloud.

[0113] By performing the above operations, this solution addresses the technical problems of traditional sound barrier point cloud registration algorithms, such as large data processing volume, high computational complexity, and lack of effective initial attitude calibration, large-scale offset correction, and stable registration region screening, which lead to excessively long registration calculation time and low registration accuracy. This solution innovatively adopts a progressive registration strategy consisting of bounding box alignment, large-scale offset detection in the Y direction, percentile pruning in the X direction, anchoring region extraction, multi-scale ICP registration, and complete transformation application. This effectively reduces the computational complexity and power consumption of the registration algorithm, significantly improves the registration convergence speed and matching accuracy, enhances the stability and robustness of the registration process, and accurately avoids registration inaccuracies caused by initial position deviations and outliers. It can flexibly adapt to the registration requirements of sound barrier point clouds with different detection distances and sparsities, ultimately achieving fast and accurate global registration of the sound barrier reference point cloud and the target point cloud.

[0114] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0115] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0116] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for intelligent registration of structured light point clouds of sound barriers based on multi-scale ICP, characterized in that: The method includes the following steps: Step S1: Acquisition of structured light point cloud data for the sound barrier, specifically by acquiring and preprocessing point cloud data to obtain a reference point cloud set and a target point cloud set; Step S2: Point cloud data denoising, used to eliminate noise interference from the target subject in the point cloud; specifically, firstly, coarse extraction of the main subject point cloud of the sound barrier is performed on the reference point cloud set and the target point cloud set, then, through point cloud local block division, calculation of adaptive neighborhood scale factor, calculation of adaptive neighborhood density and noise point removal, a main subject point cloud data denoising algorithm is constructed, and finally, the coarsely extracted reference subject point cloud set and target subject point cloud set are input into the point cloud data denoising algorithm to perform denoising processing, resulting in a denoised reference subject point cloud set and a denoised target subject point cloud set; Step S3: Point cloud data filtering, used to separate the target subject point cloud from surrounding clutter, invalid ground points, and low-density isolated areas. Specifically, a progressive filtering strategy is adopted, which includes depth clipping, intensity clipping, sound barrier edge detection, depth histogram clipping, density clipping, and salt-and-pepper noise filtering. The denoised reference subject point cloud set and the target subject point cloud set are sequentially screened and filtered layer by layer to finally obtain the filtered clean reference subject point cloud set and the clean target subject point cloud set. Step S4: Multi-scale ICP registration of the sound barrier point cloud. Specifically, a progressive registration strategy is adopted, which includes bounding box alignment, large-scale offset detection in the Y direction, percentage point cropping in the X direction, anchoring region extraction, multi-scale ICP registration, and complete transformation application, to achieve the registration of the clean target subject point cloud set and the clean reference subject point cloud set, and finally obtain the registered sound barrier target subject point cloud.

2. The method for intelligent registration of structured light point clouds of sound barriers based on multi-scale ICP according to claim 1, characterized in that: In step S2, the denoising of the main point cloud data specifically includes the following steps: Step S21: Coarse extraction of the main point cloud of the sound barrier. Specifically, based on the total length of the sound barrier in the detection section and the nominal detection distance between the structured light acquisition device and the sound barrier, the value ranges of the X, Y, and Z axes of the global detection coordinate system of the sound barrier are set respectively. According to the value ranges of the three axes, the point clouds in the reference point cloud set and the target point cloud set in the standardized sound barrier structured light point cloud data are marked as the temporary internal point cloud of the main body and the temporary external point cloud of the main body, respectively, to obtain the coarsely extracted reference main body point cloud set and target main body point cloud set. Step S22: Construct a denoising algorithm for the main point cloud data; Step S23: Data denoising, specifically, inputting the coarsely extracted reference subject point cloud set and target subject point cloud set into the point cloud data denoising algorithm to obtain the denoised reference subject point cloud set and the denoised target subject point cloud set.

3. The method for intelligent registration of structured light point clouds of sound barriers based on multi-scale ICP according to claim 2, characterized in that: In step S22, the construction of the main point cloud data denoising algorithm specifically includes the following steps: Step S221: Point cloud local block division, specifically, the input main point cloud set is divided at equal intervals along the Y-axis direction of the global detection coordinate system of the sound barrier with a preset step size to obtain N mutually independent point cloud local blocks; Step S222: Calculate the adaptive neighborhood scaling factor. Specifically, for each local block of the point cloud, calculate the difference between the maximum and minimum values ​​of its coordinates in the X, Y, and Z axes of the global detection coordinate system of the sound barrier. Take the maximum value among the three as the maximum size of the local block in three-dimensional space, and then calculate the adaptive neighborhood scaling factor of the local block based on the maximum size in three-dimensional space. Finally, the adaptive neighborhood scale factor for each local block of the point cloud is calculated. Step S223: Calculate the adaptive neighborhood density. Specifically, construct a KD tree structure for each local block of the point cloud. Then, for any point within a single local block of the point cloud, use the adaptive neighborhood scale factor of the local block of the point cloud as the radius of the spherical neighborhood. Query the number of all effective neighboring points within the spherical neighborhood using the KD tree. Then, calculate the adaptive neighborhood density value of the point based on the number of effective neighboring points and the volume of the spherical neighborhood. Finally, complete the adaptive neighborhood density calculation for each point within all local blocks of the point cloud to obtain the adaptive neighborhood density value corresponding to each point in the point cloud set. Step S224: Noise point removal, specifically, for each local block of the point cloud, calculate the adaptive removal threshold for that local block based on its corresponding adaptive neighborhood scale factor. Then, the adaptive neighborhood density within the local block is calculated. Points that are not identified as noise points are removed, while points that are retained are... After removing noise points from all local blocks of the sound barrier's main point cloud, the remaining point clouds from all local blocks are fused and stitched together to obtain the denoised main point cloud set; the formula used is as follows: ； In the formula, This indicates the total number of points within a local block of the point cloud. This represents the adaptive rejection threshold adjustment coefficient.

4. The method for intelligent registration of structured light point clouds of sound barriers based on multi-scale ICP according to claim 1, characterized in that: In step S3, the point cloud data filtering specifically includes the following steps: Step S31: Depth clipping, specifically, based on the nominal detection distance between the structured light acquisition device and the sound barrier, combined with the actual ranging accuracy of the structured light acquisition device, a depth tolerance is set, and then the Z-axis depth clipping interval is set according to the nominal detection distance and the depth tolerance. Finally, based on the depth clipping interval, all point clouds in the main point cloud set that exceed the interval are removed. Step S32: Intensity value clipping, specifically, calculating the structured light reflection intensity values ​​of all the main point cloud sets after depth clipping, and taking the intensity values. Quantiles as the lower limit threshold of intensity Intensity values ​​in the main point cloud set after depth clipping Point cloud, retain intensity values The point cloud, among which, Indicates the intensity quantile coefficient; Step S33: Based on sound barrier edge detection; Step S34: Depth histogram cropping, specifically, cropping the main point cloud set after edge detection by layering it at equal intervals along the Z-axis. Perform layering and construct Z-axis coordinate values. Find the histogram distribution and locate the depth value corresponding to the peak value of the histogram. , and then with Define an asymmetric filtering window for the reference. Finally, the point cloud within that window is retained; among them, and These represent the lower and upper tolerances of the histogram clipping depth, respectively. Step S35: Density clipping, specifically, projecting the point cloud from the main point cloud set after depth histogram clipping onto the Y-axis, dividing the Y-axis into equal segments according to segment length, and calculating the point cloud density of each Y-axis segment. Then, take the maximum point cloud density among all segments as the value. Then, the density clipping threshold is calculated based on the maximum point cloud density and the density scaling factor. Finally, retain the ones that satisfy the condition. And continuous Point clouds of Y-axis segments where all segments exceed the density clipping threshold, among which, This represents the threshold for determining consecutive valid segments, and its value is a positive integer. Step S36: Salt and pepper noise filtering, specifically, performing median filtering and morphological opening operations on the density-trimmed main point cloud set to filter salt and pepper noise points, finally obtaining a clean main point cloud set.

5. The method for intelligent registration of structured light point clouds of sound barriers based on multi-scale ICP according to claim 4, characterized in that: In step S33, the sound barrier edge detection specifically involves dividing the point cloud set of the main body point cloud set after the intensity value is clipped into B independent local blocks along the Y-axis direction at fixed step lengths, then performing edge detection on each local block, and finally merging all edge-detected local blocks to obtain the edge-detected main body point cloud set. The edge detection specifically involves traversing the point cloud within a single local block along the X-axis from the maximum to the minimum coordinate value, and calculating the structured light reflection intensity value of all point clouds in each row. And take the maximum sum of the intensity values ​​of all rows within the local block as the value. When h consecutive rows of all point clouds When the x-axis coordinates of the last row in a series of h rows are taken as the starting boundary of the sound barrier on the x-axis, then the boundary is determined by the x-axis coordinates of the last row in the series h rows. Then, for all rows within the same local block, traverse along the X-axis from the minimum to the maximum coordinate, calculating the point cloud values ​​for each row. And take the maximum sum of the intensity values ​​of all rows within the local block as the value. When h consecutive rows of all point clouds When the x-axis coordinate of the last row in a series of h rows is taken as the x-axis termination boundary of the sound barrier, then the x-axis coordinate of the last row is taken as the termination boundary of the sound barrier. Finally, based on this local block and Perform point cloud cropping, retaining the desired result. Point clouds; among them, This represents the initial boundary strength threshold coefficient. This represents the termination boundary strength threshold coefficient, and h represents the number of consecutive decision rows. This represents the X-axis coordinate value.

6. The method for intelligent registration of structured light point clouds of sound barriers based on multi-scale ICP according to claim 1, characterized in that: In step S4, the multi-scale ICP registration of the sound barrier point cloud specifically includes the following steps: Step S41: Bounding box alignment, specifically, constructing 3D axis-aligned bounding boxes for the clean reference subject point cloud set and the clean target subject point cloud set respectively, extracting the lower left corner vertex of the two bounding boxes as the registration reference point, translating the bounding box reference point of the target subject point cloud to the bounding box reference point of the reference subject point cloud, completing the initial position alignment of the two subject point cloud sets, and obtaining the target subject point cloud set after bounding box alignment; Step S42: Large-scale offset detection in the Y direction, specifically, extracting points along the X-axis from the clean reference point cloud set and the target point cloud set after bounding box alignment. , , The three cross-sections of the location form a three-dimensional slice region. Then, the root mean square error of the target subject point cloud relative to the reference subject point cloud within each slice region is calculated. If the average root mean square error value of the three slice regions If the offset exceeds the preset threshold, the target point cloud is determined to have a large-scale offset along the Y-axis. Subsequently, the target point cloud is translated along the Y-axis according to a preset translation step size, and the average root mean square error value after each translation is calculated. The translation amount corresponding to the minimum average root mean square error is taken as the optimal offset, and the target point cloud is translated and corrected to obtain the target point cloud set after large-scale offset. , , This is expressed as the cross-sectional extraction ratio; Step S43: X-axis percentile cropping. Specifically, for the target subject point cloud set after large-scale migration and the clean reference subject point cloud set, the percentile distribution of their coordinate values ​​in the X-axis direction is calculated, and the points are cropped and removed if the X-axis coordinates are less than a certain value. quantile greater than Extreme outlier point clouds of quantiles, preserving the X-axis direction. The point cloud set within the quantile range yields the target subject point cloud set and the reference subject point cloud set after percentile cropping. This indicates the lower percentile of the percentile retention interval. This indicates the upper percentile of the percentile retention interval; Step S44: Anchoring region extraction. Specifically, for the point clouds of the target subject point cloud set and the reference subject point cloud set after percentage cropping, the minimum value of the point cloud in its own X-axis direction is used as the benchmark, and the point clouds are extracted from the bottom according to the height percentage. to Point clouds within the height range are used as registration and anchoring regions, resulting in point cloud sets for the target subject's anchoring region and reference subject's anchoring region. This indicates the lower limit of the anchoring area in the X-axis direction. This indicates the upper limit of the anchored area in the X-axis direction; Step S45: Multi-scale ICP registration; Step S46: Complete transformation application, specifically based on the optimal 3D rotation matrix and optimal 3D registration translation vector of the anchored region point cloud set, the transformation parameters are applied to the target subject point cloud set after percentile clipping through rigid body space transformation, global spatial attitude correction is performed on the complete target point cloud, and finally the registered target subject point cloud set is obtained, realizing the matching of the two point cloud sets.

7. The method for intelligent registration of structured light point clouds of sound barriers based on multi-scale ICP according to claim 1, characterized in that: In step S45, the multi-scale ICP registration specifically involves using the reference subject anchoring region point cloud set as a reference and the target subject anchoring region point cloud set as the registration object, and registering them according to the downsampling ratio. , and Simultaneously downsampling is performed on the point cloud set of the target anchoring region and the point cloud set of the reference anchoring region to obtain three sets of downsampled point cloud pairs, and an offset radius limit is set for the three sets of downsampled point cloud pairs. , and Then, weighted ICP registration calculations are performed on each set of downsampled point cloud pairs from coarse to fine scales, finally obtaining the optimal 3D rotation matrix and the optimal 3D registration translation vector of the anchored region point cloud set.

8. The method for intelligent registration of structured light point clouds of sound barriers based on multi-scale ICP according to claim 1, characterized in that: In step S1, the acquisition of the sound barrier structured light point cloud data specifically involves collecting point cloud data to obtain the original structured light point cloud data of the sound barrier, preprocessing the original structured light point cloud data of the sound barrier to obtain standardized structured light point cloud data of the sound barrier, and then dividing the standardized structured light point cloud data of the sound barrier into two sets of point cloud sets to be registered, denoted as the reference point cloud set and the target point cloud set, respectively; the point cloud data preprocessing specifically includes coordinate system calibration and format normalization processing.

Images