Data acquisition pre-filtering processing system based on three-dimensional laser scanning
By constructing an adaptive three-dimensional laser scanning data acquisition pre-filtering system, and dynamically adjusting the neighborhood search and filtering parameters, the instability problem of traditional filtering algorithms in landslide areas is solved, and high-precision landslide monitoring is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-12
- Publication Date
- 2026-04-07
AI Technical Summary
In existing technologies, traditional filtering algorithms have fixed parameters or rely on human experience, which cannot adapt to the problems of uneven point cloud density distribution and drastic slope changes in landslide areas, resulting in unstable processing results.
By employing modules such as multimodal data acquisition, multimodal vegetation penetration preprocessing, terrain adaptive dynamic parameter optimization, and micro-variation feature enhancement multi-scale analysis, a closed-loop control system is constructed. This system dynamically adjusts the neighborhood search strategy and filtering parameters, and combines a multi-constraint ICP algorithm and a multi-weight registration error function to achieve adaptive processing of point cloud data.
It significantly improves the robustness and automation of the system, enabling it to capture sub-millimeter-level terrain changes, providing high-precision support for early landslide warning, and reducing the complexity and error of subsequent processing.
Smart Images
Figure CN121811259A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of terrain monitoring, and in particular to a data acquisition pre-filtering processing system based on three-dimensional laser scanning. BACKGROUND
[0002] Landslide disasters, as a global frequent geological disaster, pose a serious threat to human life and property and the ecological environment. Traditional terrain monitoring methods such as manual measurement and photogrammetry have problems such as low efficiency, limited precision, and difficulty in real-time monitoring of full coverage. Three-dimensional laser scanning technology has become a key technology for monitoring post-landslide terrain changes due to its advantages of high precision, high efficiency, and non-contact acquisition of massive point cloud data. The pre-filtering processing system effectively removes isolated noise points, suppresses systematic errors, and smooths local abnormal fluctuations by designing a targeted filtering algorithm for pre-processing of raw point cloud data, thereby improving the quality and reliability of three-dimensional laser scanning data. The filtered point cloud data can more accurately reflect the morphological characteristics, displacement trend, and deformation characteristics of the landslide body, providing a scientific basis for landslide stability assessment, disaster warning, and emergency disposal. Meanwhile, as a key preprocessing step in the data acquisition process, pre-filtering can reduce the complexity and error propagation of subsequent data processing, improve the efficiency and accuracy of the overall monitoring process, and enhance the dynamic monitoring capability of landslide disasters, providing real-time and accurate terrain change information for disaster prevention and mitigation decision-making, ultimately serving the protection of people's life and property safety, the maintenance of ecological safety, and the innovation and development of geological disaster monitoring technology.
[0003] In the prior art, the parameters of traditional filtering algorithms are usually global fixed values or rely on manual experience adjustment, which cannot adapt to the non-uniform characteristics of uneven point cloud density distribution and steep slope changes in landslide areas, resulting in unstable processing results. Therefore, a data acquisition pre-filtering processing system based on three-dimensional laser scanning is proposed. SUMMARY
[0004] The purpose of the present application is to solve the problems existing in the prior art and to provide a data acquisition pre-filtering processing system based on three-dimensional laser scanning.
[0005] To achieve the above-mentioned purpose, the present application adopts the following technical solutions: A data acquisition pre-filtering processing system based on three-dimensional laser scanning, comprising: A multi-modal data acquisition module: synchronously acquires geometric features, multi-spectral intensity data, polarization parameters, and positioning and attitude information; A multi-modal vegetation penetration preprocessing module: separates vegetation and terrain based on the data provided by the multi-modal data acquisition module, and outputs pure terrain point cloud and vegetation mask; The terrain adaptive dynamic parameter optimization module: obtains the pure terrain point cloud output by the multi-modal vegetation penetration preprocessing module as the source point cloud and the target point cloud, finds a candidate corresponding point in the target point cloud for each point in the source point cloud based on a variable-scale neighborhood search algorithm, performs multiple filtering on the candidate corresponding point pair, constructs a multi-weight registration error function to solve an optimal rigid transformation matrix, applies the optimal rigid transformation matrix to coordinate transformation of the source point cloud, and calculates the registration error between the transformed point cloud and the target point cloud, specifically including: using the variable-scale neighborhood search algorithm to adaptively adjust the range and shape of neighborhood search according to local slope and point cloud density, using a multi-constraint ICP improved algorithm for the multiple filtering, including geometric distance constraint, curvature consistency constraint and intensity similarity constraint, the multi-weight registration error function fuses geometric distance, curvature consistency and intensity similarity, if the registration error is less than a preset threshold or the number of iterations reaches an upper limit, the final registration result is output, otherwise, the current registration error and the curvature consistency index are fed back to dynamically adjust the search radius and the weight coefficients of each constraint in the next iteration; The micro-variation feature enhancement multi-scale analysis module: receives the dynamically optimized parameters and performs multi-scale pyramid decomposition and feature saliency detection on the pure terrain point cloud; The quality evaluation and output module: performs quality evaluation on the processed point cloud, and outputs the pure terrain point cloud, the micro-variation feature enhancement result and the quality evaluation report.
[0006] The above technical solution further includes: Further, the multi-modal data acquisition module specifically integrates a polarization laser radar, a multi-spectral camera and a GPS / IMU positioning and orientation system; The polarization laser radar is used to obtain the three-dimensional coordinates of the point cloud and the corresponding polarization parameters; The multi-spectral camera is used to synchronously acquire multi-spectral intensity information in the near-infrared and short-wave infrared bands; The GPS / IMU positioning and orientation system is used to provide a space-time reference for each scanning point.
[0007] Further, the multi-modal vegetation penetration preprocessing module is based on a VTS-Net model, the input of which is the point cloud geometric features, multi-spectral intensity and polarization parameters provided by the multi-modal data acquisition module, and the output is a vegetation mask, a confidence weight and an initial terrain estimate.
[0008] Further, the VTS-Net model distinguishes between vegetation with random polarization characteristics and rock-soil bodies with stable polarization characteristics by analyzing the polarization parameters, and fuses the high reflectivity characteristics of multi-spectral intensity in the near-infrared band to jointly construct a material classification model, and finally identifies and filters out vegetation points through a time-series penetration algorithm, and outputs a vegetation mask, a confidence weight and an initial terrain estimate.
[0009] Furthermore, the variable-scale neighborhood search algorithm specifically includes: In steep slope areas, slope-guided neighborhood ellipticization is performed to stretch the search range along the slope direction. Based on the calculated local slope principal direction, the search neighborhood is adjusted from a standard sphere to an ellipse, and the search range is stretched along the slope direction. A density-aware KD-tree is constructed in sparse point cloud regions, automatically expanding the search radius in sparse regions and shrinking the search radius in dense regions.
[0010] Furthermore, the geometric distance constraint means eliminating point pairs whose Euclidean distance is greater than a dynamic threshold, the curvature consistency constraint means calculating the curvature difference between point pairs and eliminating point pairs whose curvature difference exceeds a preset tolerance, ensuring that matching point pairs are located on similar terrain structures, and the intensity similarity constraint means calculating the intensity value difference between point pairs based on the intensity information obtained by the multispectral camera and eliminating point pairs with excessive differences.
[0011] Furthermore, the micro-variation feature enhancement multi-scale analysis module includes a crack sensing filter unit and a curvature-preserving downsampling unit: The crack sensing filtering unit identifies linear crack features based on directional variance analysis, and maintains filtering sharpness in the normal direction of the crack and performs moderate smoothing in the tangential direction. The curvature-preserving downsampling unit retains more sampling points in regions of abrupt curvature change and introduces information entropy constraints to ensure that the loss of geometric information during the downsampling process is minimized.
[0012] Furthermore, the quality assessment and output module includes a local point cloud consistency detection unit, a confidence-weighted fusion unit, and a differential point cloud analysis unit. The local point cloud consistency detection unit is used to evaluate the signal quality in the loose rock and soil area; The confidence-weighted fusion unit is used to optimize the final terrain estimate by combining the confidence weights output by the front-end module; The differential point cloud analysis unit is used to compare multi-period scan data to extract minute terrain changes at the millimeter to sub-millimeter level.
[0013] The present invention has the following beneficial effects: 1. In this invention, a closed-loop control system of evaluation-search-adjustment-feedback is constructed through a terrain-adaptive dynamic parameter optimization module. This system can dynamically adjust the neighborhood search strategy and filtering parameters based on a real-time terrain complexity map, reducing registration errors in steep slope areas and completely solving the algorithm failure problem caused by sudden terrain changes, significantly improving the system's robustness and automation level.
[0014] 2、In the application, based on the aforementioned pure point cloud and optimization parameters provided, the micro-variation feature enhancement multi-scale analysis module adopts crack perception filtering and curvature preserving downsampling and other targeted algorithms, which realizes the removal of noise while enhancing and preserving the directionality and significance of micro-variation features, which enables the system to capture sub-millimeter level terrain changes and provide unprecedented precision support for early landslide warning. BRIEF DESCRIPTION OF DRAWINGS
[0015] Fig. 1 A system block diagram of a data acquisition pre-filtering processing system based on three-dimensional laser scanning is proposed in the application; Fig. 2 A flowchart of the terrain adaptive dynamic parameter optimization module in the application. DETAILED DESCRIPTION
[0016] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.
[0017] Please refer to Figs. 1-2 The application is a data acquisition pre-filtering processing system based on three-dimensional laser scanning, which comprises: A multi-modal data acquisition module: synchronously acquires geometric features, multi-spectral intensity data, polarization parameters and positioning and attitude information; A multi-modal vegetation penetration preprocessing module: separates vegetation and terrain based on the data provided by the multi-modal data acquisition module, and outputs pure terrain point cloud and vegetation mask; The terrain adaptive dynamic parameter optimization module: obtains the pure terrain point cloud output by the multi-modal vegetation penetration preprocessing module as the source point cloud and the target point cloud, finds a candidate corresponding point in the target point cloud for each point in the source point cloud based on a variable scale neighborhood search algorithm, performs multiple filtering on the candidate corresponding point pairs, constructs a multi-weight registration error function to solve an optimal rigid transformation matrix, applies the optimal rigid transformation matrix to coordinate transformation of the source point cloud, and calculates the registration error between the transformed point cloud and the target point cloud, specifically including: adopting the variable scale neighborhood search algorithm, adaptively adjusting the range and shape of neighborhood search according to local slope and point cloud density, the multiple filtering adopts a multi-constraint ICP improved algorithm, including geometric distance constraint, curvature consistency constraint and intensity similarity constraint, to improve the point cloud registration accuracy in areas with dramatic slope changes, the multi-weight registration error function fuses geometric distance, curvature consistency and intensity similarity, if the registration error is less than a preset threshold or the number of iterations reaches an upper limit, the final registration result is output, otherwise, the current registration error and the curvature consistency index are fed back (by calculating the local point cloud consistency index and / or comparing the differential point clouds generated by multiple scans to evaluate the point cloud processing quality and generate a feedback signal), and the search radius and the weight coefficients of each constraint in the next iteration are dynamically adjusted; The micro-variation feature enhancement multi-scale analysis module: receives the dynamically optimized parameters and performs multi-scale pyramid decomposition and feature saliency detection on the pure terrain point cloud to smooth the noise while enhancing and preserving the micro-variation terrain features; The quality evaluation and output module: evaluates the processed point cloud to ensure the terrain estimation accuracy and reliability, and outputs the pure terrain point cloud, micro-variation feature enhancement result and quality evaluation report that meet the landslide monitoring requirements.
[0018] In one embodiment, the multi-modal data acquisition module specifically integrates a polarized laser radar, a multi-spectral camera and a GPS / IMU positioning and orientation system; The polarized laser radar is used to obtain the three-dimensional coordinates of the point cloud and the corresponding polarization parameters; The multi-spectral camera is used to synchronously acquire multi-spectral intensity information in the near-infrared and short-wave infrared bands; The GPS / IMU positioning and orientation system is used to provide a space-time reference for each scanning point.
[0019] In one embodiment, the multi-modal vegetation penetration preprocessing module is based on a VTS-Net model, the input of which is the point cloud geometric features, multi-spectral intensity and polarization parameters provided by the multi-modal data acquisition module, and the output is a vegetation mask, a confidence weight and an initial terrain estimate, thereby realizing accurate separation of vegetation points and terrain points and providing the subsequent modules with pure terrain point clouds with improved signal-to-noise ratio.
[0020] In one embodiment, the VTS-Net model distinguishes vegetation with random polarization characteristics from rock-soil with stable polarization characteristics by analyzing polarization parameters, and fuses the high reflectivity characteristics of multispectral intensity in the near-infrared band to jointly construct a material classification model, and finally identifies and filters out vegetation points through a time series penetration algorithm, and outputs a vegetation mask, a confidence weight and an initial terrain estimate.
[0021] In this embodiment: Multi-modal data input and feature fusion Synchronous data from the multi-modal data acquisition module is received, the data including point cloud geometric features, multispectral intensity and polarization parameters, the multispectral intensity and polarization parameters are taken as feature attributes of the point cloud, and are associated and fused with the point cloud geometric features point by point to form a multi-modal fusion feature point cloud , wherein the point cloud geometric features at least include three-dimensional space coordinates and normal vector information; the multispectral intensity at least includes near-infrared band reflectivity and short-wave infrared band reflectivity, which utilizes the high reflectivity difference of vegetation and rock-soil in the near-infrared band and the water absorption difference in the short-wave infrared band; the polarization parameter is a polarization degree of laser echo, which is used to distinguish random polarization caused by vegetation leaves and stable polarization caused by rock-soil surface; The multi-modal fusion feature point cloud is input into a pre-trained VTS-Net vegetation and terrain separation network; the VTS-Net extracts deep fusion features of each point through a multi-layer perceptron, and the deep fusion features are represented as , wherein and are network parameters, is a nonlinear activation function, the deep fusion features comprehensively include spatial coordinates, normal vectors, near-infrared and short-wave infrared band reflectivity, and polarization degree information, the VTS-Net model uses a point cloud data set containing accurate vegetation and terrain labels, the point cloud of the data set contains the multi-modal fusion features, cross-entropy loss is taken as a main loss function to supervise the accuracy of semantic segmentation, and a regression loss is used to supervise the output of the confidence weight, so that the model simultaneously learns the class attribution and prediction reliability of the points during the training process; Based on the deep fusion features, the VTS-Net performs semantic segmentation on each point through a classifier, and outputs the probability that it is a vegetation point or a terrain point , wherein , according to a preset probability threshold , a binary vegetation mask is generated The vegetation points are filtered out based on the confidence weight and the semantic segmentation result, and the remaining terrain points are weighted and optimized to generate an initial terrain estimation point cloud set The VTS-Net outputs a confidence weight of each point being a terrain point in parallel The confidence filtering threshold is used to filter out points with low confidence. When processing multi-period scanning data, a time series penetration algorithm is called to compare and analyze the semantic segmentation result and the confidence weight of the same geographical location in different periods of point cloud, identify and retain data that is stably determined as a terrain point in multiple periods, and further remove transient false points introduced by seasonal changes of vegetation (such as fallen leaves and new branches and leaves), thereby outputting a time series stable pure terrain point cloud. More specifically, the time series penetration algorithm establishes a time series point cloud stack. For each point in the stack, the frequency of being determined as a terrain point by the VTS-Net in different periods is calculated. When the frequency exceeds a preset stability threshold, the point is confirmed as a reliable terrain point and included in the final output. Otherwise, it is considered as an unstable point and is filtered out, thereby effectively dealing with the seasonal change interference of the vegetation covered area.
[0022] In one embodiment, the variable-scale neighborhood search algorithm specifically includes: In steep slope areas, slope-guided neighborhood ellipsizing is performed to stretch the search range in the slope direction. According to the calculated local slope main direction, the search neighborhood is adjusted from a standard sphere to an ellipse, and the search range is stretched in the slope direction. In point cloud sparse areas, a density-aware KD-tree is constructed, and the search radius is automatically expanded in point cloud sparse areas and reduced in dense areas to ensure data integrity and avoid neighborhood point selection failure.
[0023] In one embodiment, the geometric distance constraint removes point pairs with a Euclidean distance greater than a dynamic threshold, the curvature consistency constraint calculates the curvature difference between point pairs and removes point pairs with a curvature difference exceeding a preset tolerance, and the intensity similarity constraint calculates the intensity value difference between point pairs based on the intensity information obtained by a multi-spectral camera and removes point pairs with a large difference.
[0024] In this embodiment: Multi-feature source point cloud preparation: obtaining the pure terrain point cloud output by the multi-modal vegetation penetration preprocessing module as the source point cloud and the target point cloud, and extracting the geometric coordinates, intensity features, and normal vector and curvature value calculated based on local surface fitting of each point from the point cloud; Multi-constraint corresponding point pair search: based on the variable scale neighborhood search strategy, for each point in the source point cloud, find the candidate corresponding point in the target point cloud; the variable scale neighborhood search strategy dynamically adjusts the search radius according to the terrain complexity map provided by the terrain adaptive dynamic parameter optimization module in real time , in the point cloud sparse or steep slope area, the search range is automatically expanded, which is expressed as , wherein is the basic radius, is the adjustment coefficient, According to the local point cloud density and slope, in the steep slope area, the value increases, thereby expanding the search range; Multi-dimensional corresponding point pair screening: multiple filtering is performed on the candidate corresponding point pairs obtained in the multi-constraint corresponding point pair search, and the filtering constraints at least include: Geometric distance constraint: eliminating point pairs with a Euclidean distance greater than a dynamic threshold; Curvature consistency constraint: calculating the curvature difference between the point pairs, and eliminating the point pairs with a curvature difference exceeding a preset tolerance, to ensure that the matching point pairs are located on similar terrain structures; Intensity similarity constraint: based on the intensity information obtained by the multi-spectral camera, calculating the intensity value difference between the point pairs, and eliminating the point pairs with a too large difference; Multi-weight registration error function construction and solving: a multi-weight registration error function is constructed, which integrates the geometric distance, curvature consistency and intensity similarity, and a singular value decomposition or Levenberg-Marquardt algorithm is used to solve the optimal rigid transformation matrix to minimize the error function; Iterative transformation and quality feedback: applying the optimal rigid transformation matrix to perform coordinate transformation on the source point cloud, and calculating the registration error between the transformed point cloud and the target point cloud; feeding back the current registration error and the curvature consistency index to the quality feedback closed loop of the terrain adaptive dynamic parameter optimization module, for dynamically adjusting the search radius and the weight coefficients of each constraint in the next iteration; Convergence judgment and result output step: if the registration error is less than a preset threshold or the number of iterations reaches an upper limit, the iteration is terminated and the final registration result is output; otherwise, return to the multi-constraint corresponding point pair search for the next iteration.
[0025] In the multi-weight registration error function construction and solving, the mathematical expression of the multi-weight registration error function is: , wherein is the rigid transformation matrix to be solved, is a point in the source point cloud, is a corresponding point in the target point cloud; represents a curvature value of the point, represents an intensity value of the point; are respectively a geometric distance weight, a curvature consistency weight and an intensity similarity weight, and the three weight coefficients are dynamically allocated by the terrain adaptive dynamic parameter optimization module according to local terrain complexity and point cloud quality.
[0026] The dynamic allocation logic of the weight coefficients is that, in an area with uniform point cloud density and rich features, the geometric distance weight is given a higher value; in an area with sparse point cloud or complex terrain structure, the contribution of the curvature consistency weight and the intensity similarity weight is increased, so as to make up for the lack of geometric information by using more dimensional information and ensure the robustness of registration.
[0027] In the iterative transformation and quality feedback, the quality feedback closed loop specifically includes: monitoring the satisfaction rate of the curvature consistency constraint in the screened point pair in real time, if the satisfaction rate continuously falls below a dynamic threshold, then automatically reducing the tolerance value of the curvature consistency constraint in the multi-dimensional corresponding point pair screening, and correspondingly increasing the weight coefficient of the multi-weight registration error function in construction and solving , so as to realize adaptive optimization of the registration process.
[0028] In one embodiment, the micro-variation feature enhanced multi-scale analysis module includes a crack perception filtering unit and a curvature preserving down-sampling unit: The crack perception filtering unit identifies linear crack features based on directional variance analysis, and maintains filtering sharpness in the normal direction of the crack and performs moderate smoothing in the tangent direction; The curvature preserving down-sampling unit retains more sampling points in the curvature mutation area, and introduces an information entropy constraint to ensure that the loss of geometric information in the down-sampling process is minimized.
[0029] The specific analysis process of the micro-variation feature enhanced multi-scale analysis module for feature enhancement is as follows: Multi-scale feature response calculation: on the voxel down-sampling layers of different scales, the normal change rate (curvature) and feature value divergence of each point are calculated; Linear feature recognition and protection: for points whose eigenvalues' divergence is greater than the fourth threshold and one of the eigenvalues is significantly greater than the other two, they are determined as potential crack edge points and only slightly smoothed in their normal direction while keeping the original coordinates in their tangent plane during the filtering process; Information entropy oriented down-sampling: in the down-sampling stage, points with the least contribution to the local surface information entropy are preferentially deleted to ensure that the point density retention rate in areas with high curvature is higher than that in flat areas.
[0030] In one embodiment, the quality assessment and output module includes a local point cloud consistency detection unit, a confidence weighted fusion unit, and a differential point cloud analysis unit: The local point cloud consistency detection unit is used to evaluate the signal quality of the loose rock-soil region; The confidence weighted fusion unit is used to optimize the final terrain estimation in combination with the confidence weight output by the front-end module; The differential point cloud analysis unit is used to compare multi-period scanning data to extract millimeter to sub-millimeter level micro-topographic changes.
[0031] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.
Claims
1. A data acquisition pre-filtering system based on three-dimensional laser scanning, characterized in that, include: Multimodal data acquisition module: synchronously acquires geometric features, multispectral intensity data, polarization parameters, and positioning and attitude information; Multimodal vegetation penetration preprocessing module: Based on the data provided by the multimodal data acquisition module, vegetation and terrain are separated, and a clean terrain point cloud and vegetation mask are output; The terrain adaptive dynamic parameter optimization module obtains the clean terrain point cloud output by the multimodal vegetation penetration preprocessing module as the source point cloud and the target point cloud. Based on the variable-scale neighborhood search algorithm, it searches for candidate corresponding points in the target point cloud for each point in the source point cloud. It performs multiple filtering on the candidate corresponding point pairs, constructs a multi-weighted registration error function to solve for the optimal rigid body transformation matrix, applies the optimal rigid body transformation matrix to transform the coordinates of the source point cloud, and calculates the registration error between the transformed point cloud and the target point cloud. Specifically, it adopts a variable-scale neighborhood search algorithm to adaptively adjust the range and shape of the neighborhood search according to the local slope and point cloud density. The multiple filtering adopts a multi-constraint ICP improved algorithm, which includes geometric distance constraints, curvature consistency constraints, and intensity similarity constraints. The multi-weighted registration error function integrates geometric distance, curvature consistency, and intensity similarity. If the registration error is less than a preset threshold or the number of iterations reaches the upper limit, the final registration result is output. Otherwise, the current registration error and curvature consistency index are fed back to dynamically adjust the search radius and the weight coefficients of each constraint in the next iteration. Micro-variation feature enhancement multi-scale analysis module: Receives dynamically optimized parameters and performs multi-scale pyramid decomposition and feature saliency detection on the pure terrain point cloud; Quality assessment and output module: Performs quality assessment on the processed point cloud and outputs clean terrain point cloud, micro-variation feature enhancement results, and quality assessment report.
2. The data acquisition pre-filtering system based on three-dimensional laser scanning according to claim 1, characterized in that, The multimodal data acquisition module specifically integrates a polarization lidar, a multispectral camera, and a GPS / IMU positioning and orientation system; The polarization lidar is used to acquire the three-dimensional coordinates of the point cloud and its corresponding polarization parameters; The multispectral camera is used to simultaneously acquire multispectral intensity information in the near-infrared and short-wave infrared bands; The GPS / IMU positioning and orientation system is used to provide a spatiotemporal reference for each scan point.
3. The data acquisition pre-filtering system based on three-dimensional laser scanning according to claim 1, characterized in that, The multimodal vegetation penetration preprocessing module is based on the VTS-Net model. Its inputs are the point cloud geometric features, multispectral intensity, and polarization parameters provided by the multimodal data acquisition module, and its outputs are vegetation mask, confidence weights, and initial terrain estimation.
4. The data acquisition pre-filtering system based on three-dimensional laser scanning according to claim 3, characterized in that, The VTS-Net model distinguishes between vegetation with random polarization characteristics and soil and rock with stable polarization characteristics by analyzing polarization parameters. It also integrates the high reflectivity of multispectral intensity in the near-infrared band to jointly construct a material classification model. Finally, it identifies and filters out vegetation points through a time-series penetration algorithm, and outputs vegetation mask, confidence weight, and initial terrain estimate.
5. The data acquisition pre-filtering system based on three-dimensional laser scanning according to claim 1, characterized in that, The variable-scale neighborhood search algorithm specifically includes: In steep slope areas, slope-guided neighborhood ellipticization is performed to stretch the search range along the slope direction. Based on the calculated local slope principal direction, the search neighborhood is adjusted from a standard sphere to an ellipse, and the search range is stretched along the slope direction. A density-aware KD-tree is constructed in sparse point cloud regions, automatically expanding the search radius in sparse regions and shrinking the search radius in dense regions.
6. The data acquisition pre-filtering system based on three-dimensional laser scanning according to claim 1, characterized in that, The geometric distance constraint means to remove point pairs whose Euclidean distance is greater than a dynamic threshold. The curvature consistency constraint means to calculate the curvature difference between point pairs and remove point pairs whose curvature difference exceeds a preset tolerance, ensuring that the matching point pairs are located on similar terrain structures. The intensity similarity constraint means to calculate the intensity value difference between point pairs based on the intensity information obtained by the multispectral camera and remove point pairs with excessive differences.
7. The data acquisition pre-filtering system based on three-dimensional laser scanning according to claim 1, characterized in that, The micro-variation feature enhancement multi-scale analysis module includes a crack sensing filtering unit and a curvature-preserving downsampling unit: The crack sensing filtering unit identifies linear crack features based on directional variance analysis, and maintains filtering sharpness in the normal direction of the crack and performs moderate smoothing in the tangential direction. The curvature-preserving downsampling unit retains more sampling points in regions of abrupt curvature change and introduces information entropy constraints to ensure that the loss of geometric information during the downsampling process is minimized.
8. The data acquisition pre-filtering system based on three-dimensional laser scanning according to claim 1, characterized in that, The quality assessment and output module includes a local point cloud consistency detection unit, a confidence-weighted fusion unit, and a differential point cloud analysis unit. The local point cloud consistency detection unit is used to evaluate the signal quality in the loose rock and soil area; The confidence-weighted fusion unit is used to optimize the final terrain estimate by combining the confidence weights output by the front-end module; The differential point cloud analysis unit is used to compare multi-period scan data to extract minute terrain changes at the millimeter to sub-millimeter level.