A sparse vehicle-mounted point cloud pit groove three-dimensional parameter extraction method

Abstract

The application discloses a sparse vehicle-mounted point cloud pit three-dimensional parameter extraction method, and belongs to the technical field of three-dimensional point cloud data processing and geometric shape recovery. In order to solve the problems of unstable surface recovery of sparse point cloud, difficult closure of pit boundary and inconsistent geometric parameter reference, the application fuses multiple frames of point clouds in a pit semantic region of interest and a buffer zone, constructs a local truncated signed distance field, extracts a zero isosurface and resamples; taking a local reference road surface as a reference, a continuous residual depth field is constructed through residual aggregation and smoothing extension, a closed pit region is solved through a region type active contour energy with semantic consistency constraint; and unilateral directional equivalent diameters are calculated in combination with a vehicle advancing direction and left and right wheel track strip domains, and pit area, maximum depth and volume are calculated, so that the continuity of surface recovery can be improved, the influence of boundary discrete error on parameter extraction can be reduced, and the application is suitable for pit disease quantitative characterization and vehicle passing analysis fields.

Description

Technical Field

[0001] This invention belongs to the field of three-dimensional point cloud data processing, computer graphics and image analysis and geometric shape restoration technology, specifically involving the extraction technology of three-dimensional parameters of sparse vehicle-mounted laser point cloud pits. Background Technology

[0002] Potholes alter the contact between vehicle tires and the road surface. Their area, depth, volume, and the scale they span in the direction of vehicle travel provide the geometric information needed for road maintenance assessment and traffic analysis. Vehicle-mounted LiDAR can continuously acquire 3D point clouds of the road as the vehicle moves, and can utilize sensor pose to convert observations from different times to a unified coordinate reference. Therefore, it is suitable for non-contact digital measurement of local road defects. Existing processing methods typically first obtain the region of interest (ROI) of the pothole from the road point cloud, and then directly calculate the elevation, fit the contour, or create a triangulation within that region to obtain the planar and longitudinal scales of the pothole.

[0003] In existing point cloud geometric measurement methods, explicit triangulation relies on the adjacency relationships between discrete points, while implicit distance field methods represent the object surface using isosurfaces of a continuous scalar field. For local road undulations, a normal road surface is typically fitted as a depth reference, and 3D points are projected onto the road plane to form a regular grid. After obtaining the closed contour, the area, maximum depth, and volume can be calculated through boundary integration, residual extrema, or piecewise integration. These techniques provide a basic means for calculating pothole parameters; however, if each processing step uses independent spatial references and computational regions, it is still difficult to guarantee the consistency of the obtained parameters.

[0004] However, vehicle-mounted laser point clouds are affected by the number of line beams, angular resolution, observation distance, vehicle movement, and pothole edge occlusion, exhibiting sparse, non-uniform, and locally missing measurement characteristics in pothole areas. Directly connecting discrete points easily creates cross-gap connections at voids, and direct interpolation of single-frame point clouds makes it difficult to utilize complementary observations from adjacent frames, resulting in unstable continuous surface recovery. Furthermore, road longitudinal slope, cross slope, and gently varying curvature, along with pothole depressions, all contribute to the absolute elevation, making depth definition, boundary location, and volume integration lack a unified relative road surface reference.

[0005] Furthermore, even after reconstructing the surface projection into a 2D raster, there may still be unobserved units. The semantic region boundaries are also affected by projection discretization and residual thresholding. Simply truncating based on gradients or thresholds makes it difficult to stably obtain smooth, closed pothole regions. Commonly used maximum chord lengths or global circumscribed scales do not incorporate vehicle travel direction, wheel positions, and tire contact width. The resulting dimensions cannot accurately represent the actual span of the left and right wheels and cause the scale parameters to become disconnected from the computational domains of area, depth, and volume. Therefore, it is necessary to provide a method for extracting 3D parameters of potholes from sparse vehicle-mounted point clouds. Summary of the Invention

[0006] To address the problems in existing technologies, such as unstable surface recovery of sparse vehicle point clouds, lack of a unified reference standard for depth and volume, difficulty in closing pothole boundaries, and inability of geometric dimensions to reflect the actual range traversed by the wheel, this invention proposes the following solutions: A method for extracting 3D parameters of sparse vehicle-mounted point cloud potholes includes: S1. Obtain the semantic region of interest for the pothole, multi-frame vehicle-mounted laser point cloud, and the corresponding sensor pose for each frame; S2. Within the local voxel computation domain determined according to the semantic region of interest of the potholes, the sensor pose is used to perform truncated symbolic distance fusion on the multi-frame vehicle-mounted laser point cloud to obtain a local truncated symbolic distance field. S3. Obtain a high-resolution surface point set based on the zero isosurface of the locally truncated symbol distance field; S4. Determine the directed residual of the high-resolution surface point set relative to the local reference road surface, and construct a continuous residual depth field based on the directed residual; S5. Based on the continuous residual depth field, perform level set contour evolution with the pit semantic interest region as semantic constraint to obtain the closed pit region. S6. Determine the vehicle's travel direction, left wheel track zone, and right wheel track zone based on the sensor's pose. Determine the directional support width of each intersection area of ​​the closed pothole area with the left wheel track zone and the right wheel track zone along the vehicle's travel direction as the left single-sided directional equivalent diameter and the right single-sided directional equivalent diameter, respectively. S7. Determine the pit area, maximum pit depth, and pit volume based on the closed pit region and the continuous residual depth field, and form a three-dimensional parameter set for the pit together with the left-side unilateral directional equivalent diameter and the right-side unilateral directional equivalent diameter.

[0007] Further, in S1, the line-of-sight information corresponding to each frame is also acquired; in S2, the multi-frame vehicle-mounted laser point cloud is converted to a unified local road coordinate system using the sensor pose, the local voxel computation domain is determined according to the region of interest of the pothole semantics and its extended buffer, along the observation ray corresponding to the line-of-sight information, the signed distance of each voxel center relative to the visible surface in the local voxel computation domain is determined according to the distance from the voxel center to the sensor and the effective observation depth corresponding to the point cloud observation point, the signed distance is truncated, and multi-frame recursive fusion is performed according to the observation weight determined by the observation distance from the point cloud observation point to the sensor and the incident geometry relationship to obtain the local truncated signed distance field.

[0008] Furthermore, the multi-frame vehicle-mounted laser point cloud includes a center frame and adjacent frames; in S2, based on the center frame, a local sliding window is constructed according to the acquisition sequence of the center frame and adjacent frames, and only the point cloud observations within the local sliding window and falling into the local voxel calculation domain are used for the multi-frame recursive fusion, and the voxel size of the local voxel calculation domain is determined according to the point spacing of the multi-frame vehicle-mounted laser point cloud and the scale of the region of interest of the pit semantic region.

[0009] Further, in S2, the local surface normal is estimated based on the neighborhood point set of each point in the multi-frame vehicle-mounted laser point cloud, the observation weight is determined based on the observation distance and the incident angle between the observation ray and the local surface normal represented by the line-of-sight information, and the observation weight of point cloud observations that conflict with the current symbol of the local truncated symbol distance field is reduced by symbol consistency discrimination.

[0010] Further, in S3, the zero isosurface is extracted to obtain a continuous triangular mesh surface. The continuous triangular mesh surface is then resampled at high resolution to obtain the high-resolution surface point set. In S4, normal road surface points are selected from the neighborhood of the pothole semantic interest region and fitted to the local reference road surface. The high-resolution surface point set is projected onto the road plane of the local road coordinate system and discretized into regular grids. The directed residuals within each regular grid are aggregated. In the continuous computational domain formed by projecting the pothole semantic interest region onto the road plane, data consistency constraints are applied to the regular grids with valid observations, and the regular grids without observations are smoothly extended to obtain the continuous residual depth field.

[0011] Furthermore, in S4, the median of the directed residuals falling into the same regular grid is determined as the observation residual of the corresponding regular grid, and the continuous residual depth field is solved by a variational model that includes a data consistency term that contains the effective observation location and a field smoothing term in the continuous computation domain.

[0012] Further, in S5, a contour evolution domain is determined based on the projection region of the pothole semantic region of interest on the road plane and a preset residual threshold of the continuous residual depth field. A regional active contour energy is constructed within the contour evolution domain. The regional active contour energy includes the boundary length term of the level set contour, the residual fitting term inside the contour, the residual fitting term outside the contour, and the semantic consistency term between the level set function and the indicator function of the pothole semantic region of interest. The level set function is updated to convergence based on the regional active contour energy, and the region enclosed by the zero level set of the level set function is determined as the closed pothole region.

[0013] Further, in S1, the fixed offsets from the sensor reference point to the left wheel contact point and the right wheel contact point, as well as the effective tire contact width, are obtained; in S6, the short-time pose sequence formed by the sensor poses is projected onto the road plane, and the vehicle's travel direction is determined based on the projection displacement of the sensor poses on the road plane at the current and next moments; each sensor pose is combined with the fixed offset and mapped onto the road plane to obtain the left wheel track centerline and the right wheel track centerline; points whose planar distance from the corresponding wheel track centerline is not greater than half of the effective tire contact width are formed into corresponding wheel track strip regions; the projection range of the corresponding intersection region in the vehicle's travel direction is determined as the corresponding unilateral directional equivalent diameter.

[0014] Further, in S7, the points within the closed pit area are triangulated to obtain an irregular triangular network. Triangles with a maximum side length greater than a preset side length threshold are removed. The pit volume is obtained by piecewise integration based on the area of ​​each retained triangle and the directed residual of each vertex in the continuous residual depth field.

[0015] Based on the same inventive concept, the present invention also proposes a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0016] Compared with the prior art, the present invention has the following beneficial effects: By organizing the point clouds of the central frame and adjacent frames within the semantic region of interest and its buffer, and recursively fusing the local truncated signed distance field based on sensor pose, line of sight, signed distance, and observation weight, the complementary observations of different frames jointly constrain the position of the zero isosurface, thereby improving the continuity and stability of pit surface recovery under sparse, non-uniform, and locally missing measurement conditions.

[0017] By fitting a local reference road surface with a neighboring normal road surface, high-resolution surface points are converted into directed residuals. Then, observation consistency constraints and smooth extensions are implemented within a unified road plane grid, separating the longitudinal and transverse slopes of the road from the local depressions of potholes. This establishes a consistent relative depth benchmark for contour positioning, maximum depth statistics, and volume integration.

[0018] By using the semantic projection region and residual threshold together to define the contour evolution domain, and combining boundary length, internal and external residual fitting and semantic consistency constraints in the regional active contour energy, the level set can still converge to a semantically reasonable closed region under weak boundary and local empty conditions, thereby reducing the impact of boundary discretization error on area and volume results.

[0019] By determining the vehicle's direction of travel from a short-time pose sequence, and combining the fixed offset from the sensor to the contact points of the left and right wheels with the effective contact width of the tires, left and right wheel tracks are constructed. The directional support width at the intersection of the wheel tracks and the closed pothole area is calculated, so that the dimensional quantities correspond to the actual effective range of a single wheel, thereby forming a three-dimensional pothole parameter set that is uniform in the computational domain and suitable for vehicle traffic analysis.

[0020] This invention has the ability to process multi-frame sparse road point clouds with pose and line-of-sight information and uniformly extract pothole geometric parameters. It can output pothole area, maximum pothole depth, pothole volume, and equivalent diameter of left and right single-sided orientation. It is applicable to fields such as quantitative characterization of pothole defects, vehicle traffic analysis, vehicle traffic simulation, vehicle dynamics analysis, traffic safety assessment, and speed threshold research. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the sparse vehicle-mounted point cloud pothole three-dimensional parameter extraction method described in the implementation method; Figure 2 This is a schematic diagram of the local sliding window truncation symbolic distance field fusion and high-resolution surface resampling process described in the implementation method; Figure 3 This is a schematic diagram of the geometric relationship between the sensor pose, observation ray, visible surface and local voxel as described in the implementation method. Figure 4 This is a schematic diagram of the process for constructing the continuous residual depth field and extracting the closed pit region as described in the implementation method; Figure 5 This is a schematic diagram showing the relationship between the vehicle's direction of travel, the center lines of the left and right wheel tracks, and the wheel track bands as described in the implementation method. Figure 6 This is a schematic diagram illustrating the calculation of the pit area, maximum depth, volume, and unilateral directional equivalent diameter as described in the implementation method. Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention will now be clearly and completely described in conjunction with the accompanying drawings.

[0023] Implementation Method 1 A method for extracting 3D parameters of sparse vehicle-mounted point cloud potholes includes: S1. Obtain the semantic region of interest for the pothole, multi-frame vehicle-mounted laser point cloud, and the corresponding sensor pose for each frame; S2. Within the local voxel computation domain determined according to the semantic region of interest of the potholes, the sensor pose is used to perform truncated symbolic distance fusion on the multi-frame vehicle-mounted laser point cloud to obtain a local truncated symbolic distance field. S3. Obtain a high-resolution surface point set based on the zero isosurface of the locally truncated symbol distance field; S4. Determine the directed residual of the high-resolution surface point set relative to the local reference road surface, and construct a continuous residual depth field based on the directed residual; S5. Based on the continuous residual depth field, perform level set contour evolution with the pit semantic interest region as semantic constraint to obtain the closed pit region. S6. Determine the vehicle's travel direction, left wheel track zone, and right wheel track zone based on the sensor's pose. Determine the directional support width of each intersection area of ​​the closed pothole area with the left wheel track zone and the right wheel track zone along the vehicle's travel direction as the left single-sided directional equivalent diameter and the right single-sided directional equivalent diameter, respectively. S7. Determine the pit area, maximum pit depth, and pit volume based on the closed pit region and the continuous residual depth field, and form a three-dimensional parameter set for the pit together with the left-side unilateral directional equivalent diameter and the right-side unilateral directional equivalent diameter.

[0024] Using the above method, multi-frame point cloud surface reconstruction, residual representation, closed contour and wheel track orientation scale are chained together to a unified road coordinate reference, so that each intermediate result becomes the input for subsequent processing, thereby realizing the integrated extraction of area, depth, volume and left and right single-sided orientation equivalent diameter.

[0025] Further, in S1, the line-of-sight information corresponding to each frame is also acquired; in S2, the multi-frame vehicle-mounted laser point cloud is converted to a unified local road coordinate system using the sensor pose, the local voxel computation domain is determined according to the region of interest of the pothole semantics and its extended buffer, along the observation ray corresponding to the line-of-sight information, the signed distance of each voxel center relative to the visible surface in the local voxel computation domain is determined according to the distance from the voxel center to the sensor and the effective observation depth corresponding to the point cloud observation point, the signed distance is truncated, and multi-frame recursive fusion is performed according to the observation weight determined by the observation distance from the point cloud observation point to the sensor and the incident geometry relationship to obtain the local truncated signed distance field.

[0026] By defining the voxel computation domain with a semantic region and its extended buffer, the truncated signed distance is obtained by comparing the voxel center distance with the effective observation depth along the observation ray. Then, the distance is recursively fused according to the weights determined by the observation distance and the incident geometry, thereby enabling multi-frame observations to stably constrain the pit surface.

[0027] Furthermore, the multi-frame vehicle-mounted laser point cloud includes a center frame and adjacent frames; in S2, based on the center frame, a local sliding window is constructed according to the acquisition sequence of the center frame and adjacent frames, and only the point cloud observations within the local sliding window and falling into the local voxel calculation domain are used for the multi-frame recursive fusion, and the voxel size of the local voxel calculation domain is determined according to the point spacing of the multi-frame vehicle-mounted laser point cloud and the scale of the region of interest of the pit semantic region.

[0028] By limiting adjacent observations based on the center frame and performing fusion only in the local voxel computation domain, multi-frame data is simultaneously constrained in time and space. The voxel size is then determined based on the point spacing and target scale, thus balancing the geometric details of the pits, the redundancy of cross-frame observations, and computational efficiency.

[0029] Further, in S2, the local surface normal is estimated based on the neighborhood point set of each point in the multi-frame vehicle-mounted laser point cloud, the observation weight is determined based on the observation distance and the incident angle between the observation ray and the local surface normal represented by the line-of-sight information, and the observation weight of point cloud observations that conflict with the current symbol of the local truncated symbol distance field is reduced by symbol consistency discrimination.

[0030] The surface normal is estimated by using a neighborhood point set, and the observation distance and ray incident angle are combined into observation weights, so that stable observations have a greater contribution to fusion. The weight of conflicting observations is reduced by combining sign consistency discrimination, thereby suppressing the zero isosurface drift caused by occlusion boundaries and grazing observations.

[0031] Further, in S3, the zero isosurface is extracted to obtain a continuous triangular mesh surface. The continuous triangular mesh surface is then resampled at high resolution to obtain the high-resolution surface point set. In S4, normal road surface points are selected from the neighborhood of the pothole semantic interest region and fitted to the local reference road surface. The high-resolution surface point set is projected onto the road plane of the local road coordinate system and discretized into regular grids. The directed residuals within each regular grid are aggregated. In the continuous computational domain formed by projecting the pothole semantic interest region onto the road plane, data consistency constraints are applied to the regular grids with valid observations, and the regular grids without observations are smoothly extended to obtain the continuous residual depth field.

[0032] By projecting a high-resolution surface point set onto a road plane grid and aggregating directed residuals, data consistency constraints are imposed on the effective observation grid within the continuous computation domain, and the unobserved grid is smoothly extended, thereby filling in the projection gaps while preserving the observation residuals.

[0033] Furthermore, in S4, the median of the directed residuals falling into the same regular grid is determined as the observation residual of the corresponding regular grid, and the continuous residual depth field is solved by a variational model that includes a data consistency term that contains the effective observation location and a field smoothing term in the continuous computation domain.

[0034] By taking the median of the directed residuals within the same grid, the influence of outliers and density differences is reduced; then, the data consistency term maintains the residuals of the valid observation locations, and the field smoothing term extends to the unobserved areas, thereby obtaining a residual depth field that combines observation reliability and spatial continuity.

[0035] Further, in S5, a contour evolution domain is determined based on the projection region of the pothole semantic region of interest on the road plane and a preset residual threshold of the continuous residual depth field. A regional active contour energy is constructed within the contour evolution domain. The regional active contour energy includes the boundary length term of the level set contour, the residual fitting term inside the contour, the residual fitting term outside the contour, and the semantic consistency term between the level set function and the indicator function of the pothole semantic region of interest. The level set function is updated to convergence based on the regional active contour energy, and the region enclosed by the zero level set of the level set function is determined as the closed pothole region.

[0036] By defining the contour evolution domain with semantic projection region and residual threshold, and setting boundary length term, internal and external residual fitting term and semantic consistency term in contour energy, smoothness constraint, regional difference and semantic prior jointly drive level set update, thereby maintaining contour closure and limiting cross-domain expansion.

[0037] Further, in S1, the fixed offsets from the sensor reference point to the left wheel contact point and the right wheel contact point, as well as the effective tire contact width, are obtained; in S6, the short-time pose sequence formed by the sensor poses is projected onto the road plane, and the vehicle's travel direction is determined based on the projection displacement of the sensor poses on the road plane at the current and next moments; each sensor pose is combined with the fixed offset and mapped onto the road plane to obtain the left wheel track centerline and the right wheel track centerline; points whose planar distance from the corresponding wheel track centerline is not greater than half of the effective tire contact width are formed into corresponding wheel track strip regions; the projection range of the corresponding intersection region in the vehicle's travel direction is determined as the corresponding unilateral directional equivalent diameter.

[0038] The direction of travel is obtained by projecting displacement of short-term pose sequence, and the wheel track is recovered by fixed offset from the sensor to the contact point of the left and right wheels, so that the wheel track conforms to the vehicle posture and tire width; then the diameter of one side is determined by the directional projection range of the intersection area, thereby reflecting the wheel crossing scale.

[0039] Further, in S7, the points within the closed pit area are triangulated to obtain an irregular triangular network. Triangles with a maximum side length greater than a preset side length threshold are removed. The pit volume is obtained by piecewise integration based on the area of ​​each retained triangle and the directed residual of each vertex in the continuous residual depth field.

[0040] By constructing an irregular triangular network within the closed pit area and deleting triangles with a maximum side length exceeding a threshold, the volume segmentation is restricted to areas with geometric support. The segmented volume is then calculated by combining the area of ​​the retained triangles and the vertex residuals, thereby reducing the impact of connecting the edges of the cross-missile cavity on the volume estimation.

[0041] The method described in this embodiment can be executed by a processor calling a computer program, which can be stored in a computer storage medium. When the computer program is executed by the processor, the above-described method for extracting 3D parameters of sparse vehicle-mounted point cloud potholes can be implemented.

[0042] Implementation Method 2 This implementation method uses a vehicle-mounted LiDAR to collect multi-frame point clouds of road potholes and extract their three-dimensional parameters as a specific example. Figures 1 to 6 The first implementation method will be further explained below.

[0043] like Figure 1 As shown, the method is executed in the following order: semantic region and multi-frame observation acquisition, local truncation symbolic distance field fusion, continuous surface restoration, residual depth field construction, closed contour solution, wheel track orientation scale calculation, and three-dimensional parameter summarization. The coordinates, regions, and intermediate results of each step are passed along the processing chain.

[0044] In S1, the semantic region of interest (ROI) for potholes, the center frame, and adjacent vehicle-mounted laser point cloud frames acquired in sequence are read. Simultaneously, the rotation and translation information of the sensors in the global coordinate system for each frame, as well as the line-of-sight information from the sensor position to the point cloud observation point, are read. The semantic ROI for potholes originates from the preceding pothole semantic segmentation results and is used to define the scope of subsequent geometric processing. The fixed offset from the sensor reference point to the contact points of the left and right wheels is derived from the calibration relationship in the vehicle coordinate system, and the effective tire contact width is derived from the vehicle tire geometry parameters. These data together constitute... Figure 1 The original input for the method shown.

[0045] Combination Figure 2 and Figure 3 In S2, the region of interest (ROI) for pothole semantics is first spatially expanded to obtain a buffer zone covering the normal road surface at the pothole edge, and a voxel mesh is established within the local area formed by the two. A local sliding window is formed by selecting several adjacent frames based on the center frame. The point cloud within the window is transformed to a unified local road coordinate system using the sensor pose of each frame, retaining only observations falling within the local voxel computation domain. The voxel size is determined according to the original point cloud spacing and the scale of the pothole semantic ROI, ensuring that the voxel mesh can represent pothole undulations without expanding to irrelevant road structures. For the first... Frame observation involves comparing the distance from the voxel center to the sensor along the observation ray with the effective observation depth in that direction to obtain the signed distance of the voxel center relative to the visible surface. Optionally, the signed distance can be truncated and normalized according to the following relationship, or other processing relationships capable of truncating the signed distance can be used:

[0046] in, For the first Frame observation at local voxel center The truncated normalized signed distance at the point, For the observation frame number, For local voxel centers within the local voxel computation domain, For the first Frame observation at local voxel center The signed distance relative to the visible surface. For a positive cutoff distance, To restrict the input to a truncation operator between a given lower bound and an upper bound, and These are the lower and upper bounds for truncation, respectively.

[0047] Optionally, when updating the previous fusion result with the current observation, the multi-frame recursive fusion can adopt the following weighted recursive relationship, or other recursive relationships that can fuse the current observation and the previous fusion result according to the observation weights:

[0048] in, To integrate the first Local voxel centers after frame observation The distance value of the truncation symbol at that point. To integrate the first Local voxel centers before frame observation The distance value of the truncation symbol at that point. To integrate the first Local voxel centers before frame observation The cumulative observation weight at each location For the first Frame observation at local voxel center The observation weight at the location, For the first Frame observation at local voxel center The truncated normalized signed distance at the point, For local voxel centers, The observation frame number is used; the weighted recursive relationship represents a weighted average of the distance value before the update and the distance value of the current observation, calculated using the cumulative observation weight and the current observation weight.

[0049] The observation weights are determined by both the observation distance and the incident geometry. Optionally, a principal direction analysis can be performed on the neighborhood set of each point to estimate the local surface normal, or other methods capable of estimating the local surface normal based on the neighborhood set can be used. The closer the observation ray is to the local surface normal and the smaller the observation distance, the greater the observation weight. When the sign of a new observation conflicts with the sign of the current voxel fusion value, the weight of the new observation is reduced. This completes the frame-by-frame update within the local voxel computational domain, outputting a locally truncated symbolic range field.

[0050] Combination Figure 2 In S3, optionally, the zero isosurface extraction can employ the moving cube method, locating isosurfaces of adjacent voxels where the sign changes in the locally truncated symbolic distance field and connecting the zero-value intersections on the edges of each voxel to form a continuous triangular mesh surface. Alternatively, other isosurface extraction methods capable of extracting a continuous triangular mesh surface from the locally truncated symbolic distance field can be used. Then, high-resolution surface resampling is performed on the continuous triangular mesh surface to generate a high-resolution surface point set with a more uniform spatial distribution and controllable density. This high-resolution surface point set inherits the surface shape recovered from multiple frames of distance fields and serves as the direct input for calculating the reference road surface residual in S4.

[0051] Combination Figure 4 In S4, points that do not fall into the pothole semantic region and have a continuous road surface distribution are selected from the extended neighborhood of the pothole semantic region of interest as normal road surface points. Optionally, a local reference road surface can be fitted by least squares, and depressed anomalies can be removed based on the directed distance from the point to the initial reference road surface before iterative updating; alternatively, other robust fitting and updating methods that can obtain a local reference road surface based on the normal road surface points can be used. The directed residual is obtained by subtracting the corresponding elevation on the local reference road surface from each point in the high-resolution surface point set; then, the points are projected onto the road plane and discretized into regular grids arranged along two local road directions. The observation residuals in the same grid are aggregated using the following formula:

[0052] in, For the residual function of the raster observation, For regular grid The observation residuals For the grid index along the first road direction, For the grid index along the second road direction, For the high-resolution surface point set The directed residual of a surface point relative to a local reference road surface. For surface point numbers, For the set function of grid point indices, To ensure that the projection falls into a regular grid The set of surface point indices To obtain the statistical operator for the median of a set, The colon indicates a membership relation, and the curly braces represent the set of directed residuals that satisfy the conditions following the colon.

[0053] The projection of the semantic region of interest (GROUP) of potholes onto the road plane and its continuous neighborhood are used as the continuous computational domain. At the effective observation grid, the continuous field is kept consistent with the observation residuals, and spatial smoothing is performed at the unobserved grid locations. Optionally, the continuous residual depth field can be solved using the following variational model, or it can be solved using other variational models that commonly include a data consistency term for the effective observation location and a field smoothing term within the continuous computational domain:

[0054] in, Let be the variational objective function with the continuous residual depth field as the independent variable. For the continuous residual depth field to be determined, For continuous residual depth field in regular grid The value at that location, This represents the value of the continuous residual depth field at the adjacent regular grid along the first road direction. This represents the value of the continuous residual depth field at the adjacent regular grid along the second road direction. For the grid index along the first road direction, For the grid index along the second road direction, A set consisting of regular grids with valid observations. For the residual function of the raster observation, For regular grid The observation residuals To balance the trade-off between the data consistency term and the field smoothing term, For continuous computation domain, The raster resolution along the first road direction. This refers to the raster resolution along the second road direction. This is the summation operator, which represents the summation of all items within the set specified in the summation range. This indicates a membership relation; the first term on the right side of the equals sign is the data consistency term, and the second term is the field smoothing term, making... Get the minimum value It was determined to be a continuous residual depth field.

[0055] Combination Figure 4In S5, the contour evolution domain is determined jointly by the road plane projection of the semantically interesting region of potholes and the residual threshold of the continuous residual depth field, and a level set function is defined within the contour evolution domain. Contour movement is driven by the regional statistical differences of the continuous residual depth field, while the reasonable range of the contour is constrained by the semantically projected region. Optionally, the regional active contour energy can be expressed as the following function, or other regional active contour energy functions that simultaneously include boundary length terms, contour interior residual fitting terms, contour exterior residual fitting terms, and semantic consistency terms can be used:

[0056] in, For regional active contour energy For level set functions, The mean value of the continuous residual depth field within the contour. The mean value of the continuous residual depth field outside the contour. The weight of the boundary length term, Let the level set function be the boundary length function of the zero level set. The weights of the residual fitting term within the contour. The weights of the external residual fitting term are the values ​​of the external residual fitting term. For continuous residual depth field, Let be the Herveside step function. The value is 1 inside the contour and 0 outside the contour. For the indicator function outside the contour, The weight of the semantic consistency term. This is an indicator function for the semantic projection region of the pothole semantic region of interest on the road plane. It takes a value of 1 within the semantic projection region and a value of 0 outside the semantic projection region. For the contour evolution domain, The integral position within the contour evolution domain. The differential symbol, For an integral infinitesimal element, This indicates integration over the contour evolution domain; the right side of the equals sign consists of the boundary length term, the contour interior residual fitting term, the contour exterior residual fitting term, and the semantic consistency term, in order.

[0057] Iteratively update the level set function until convergence; use the converged zero level set as the pit boundary, and use the connected region enclosed by the boundary as the closed pit region and pass it to S6 and S7.

[0058] Combination Figure 5 In S6, the sensor positions in the short-time pose sequence are projected onto the road plane, and the vehicle's direction of travel is determined using the difference between the projected positions at the current moment and the next moment.

[0059] in, Let be the unit vector representing the direction of vehicle travel within the road plane. For the current moment, This refers to the next time step in the short-time pose sequence that is adjacent to the current time step. Let the sensor position be the projection of the sensor position onto the road plane as a function. Let be the sensor projection position vector at the current moment. This is the sensor projection position vector at the next moment. The expression represents the dot product of vectors. The square root in the denominator represents the length of the displacement vector between the projected positions at two time points. The displacement vector is normalized using this length to obtain the unit vector.

[0060] A fixed offset from the sensor reference point to the left and right wheel contact points is applied to the poses of each sensor, then projected onto the road plane and connected sequentially to obtain the center lines of the left and right wheel tracks. The wheel track bands are determined as follows:

[0061] in, For the side The wheel track band region, For side identification numbering, A value of 1 represents the left side, and a value of 2 represents the right side. Let be the position vector of a point within the road plane. For the side The center line of the wheel track, For planar distance operators, For point To the center line of the wheel track Planar distance, The effective contact width of the tire. Half of the effective contact width of the tire. Indicates not greater than, The colon indicates a membership relationship, and the curly braces represent a set of points that satisfy the conditions following the colon; that is, the wheel track band consists of points whose planar distance from the center line of the corresponding wheel track is no greater than half the effective contact width of the tire.

[0062] Calculate the intersection of the closed pothole region and the left and right wheel track strip regions respectively, and obtain the unilateral directional equivalent diameter from the projection range of each intersection region in the vehicle's direction of travel:

[0063] in, For the side The unilateral directional equivalent diameter, For side identification numbering, A value of 1 represents the left side, and a value of 2 represents the right side. For closed pit area and side wheel track band region The intersection area Let be the position vector of the road plane points within the intersection region. This is the unit vector representing the vehicle's direction of travel. It is the transpose operator. This is the transpose of the unit vector representing the vehicle's direction of travel. For point Projected coordinates in the direction of vehicle travel For the maximum value operator, For the minimum value operator, This indicates a membership relationship; the difference between the maximum and minimum projected coordinates is the projection range of the intersection region along the vehicle's direction of travel.

[0064] Combination Figure 6 In S7, the closed pit region is used as the unified planar computational domain, and the continuous residual depth field is used as the unified longitudinal reference. The pit area is determined by the planar area of ​​the region enclosed by the closed boundary, and the maximum pit depth is determined by the absolute value of the minimum directed residual within the region.

[0065] in, The maximum depth of the pit. and These are the local coordinates within the road plane along the first road direction and the second road direction, respectively. This is a closed pit / groove area. For continuous residual depth field, For the continuous residual depth field at the road plane position The directed residual at the location, For the minimum value operator, This is an absolute value operation used to convert the minimum directed residual of a depression into a non-negative depth. This indicates a belonging relationship.

[0066] Triangulation is performed on points within the closed pit region that have depth field observations, forming an irregular triangular network. First, triangles with a maximum side length exceeding a preset side length threshold are deleted. Then, piecewise integration is performed based on the area of ​​the retained triangles and the directed residuals of the three vertices.

[0067] in, For the volume of the pit, To determine the set of valid triangles remaining after removing triangles with a maximum side length greater than a preset side length threshold, For the first One valid triangle, The triangle number, For the first The area of ​​each effective triangle on the road plane , and The first The directed residuals of the first, second, and third vertices of an effective triangle in the continuous residual depth field. For the first The average of the directed residuals at the three vertices of an effective triangle. This indicates that the larger of 0 and the value within the parentheses is used to retain only the depth contribution of the concave portion. The summation operator represents the summation of all valid triangle segments. This indicates a belonging relationship.

[0068] Optionally, the pit area can be obtained by integrating along the boundary line of the closed pit region, or other methods that can determine the planar area based on the closed pit region can be used. A preset side length threshold is used to remove long-side triangles that cross the local missing measurement area. The pit area, maximum pit depth, pit volume, and equivalent diameters of the left and right single-sided orientations are combined to form a three-dimensional parameter set for the pit and output.

[0069] The embodiments of the present invention have been described above. These embodiments are used to help understand the technical solutions of the present invention and should not be construed as limiting the scope of the technical solutions of the present invention. Equivalent substitutions or conventional modifications made by those skilled in the art to the embodiments without departing from the technical concept of the present invention should all fall within the scope of the technical solutions of the present invention.

Claims

1. A method for extracting three-dimensional parameters of sparse vehicle-mounted point cloud potholes, characterized in that, include: S1. Obtain the semantic region of interest for the pothole, multi-frame vehicle-mounted laser point cloud, and the corresponding sensor pose for each frame; S2. Within the local voxel computation domain determined according to the semantic region of interest of the potholes, the sensor pose is used to perform truncated symbolic distance fusion on the multi-frame vehicle-mounted laser point cloud to obtain a local truncated symbolic distance field. S3. Obtain a high-resolution surface point set based on the zero isosurface of the locally truncated symbol distance field; S4. Determine the directed residual of the high-resolution surface point set relative to the local reference road surface, and construct a continuous residual depth field based on the directed residual; S5. Based on the continuous residual depth field, perform level set contour evolution with the pit semantic interest region as semantic constraint to obtain the closed pit region. S6. Determine the vehicle's travel direction, left wheel track zone, and right wheel track zone based on the sensor's pose. Determine the directional support width of each intersection area of ​​the closed pothole area with the left wheel track zone and the right wheel track zone along the vehicle's travel direction as the left single-sided directional equivalent diameter and the right single-sided directional equivalent diameter, respectively. S7. Determine the pit area, maximum pit depth, and pit volume based on the closed pit region and the continuous residual depth field, and form a three-dimensional parameter set for the pit together with the left-side unilateral directional equivalent diameter and the right-side unilateral directional equivalent diameter.

2. The method according to claim 1, characterized in that, In S1, the line-of-sight information corresponding to each frame is also acquired; in S2, the multi-frame vehicle-mounted laser point cloud is converted to a unified local road coordinate system using the sensor pose, the local voxel computation domain is determined according to the region of interest of the pothole semantics and its extended buffer, the signed distance of each voxel center relative to the visible surface in the local voxel computation domain is determined along the observation ray corresponding to the line-of-sight information based on the distance from the voxel center to the sensor and the effective observation depth corresponding to the point cloud observation point, the signed distance is truncated, and multi-frame recursive fusion is performed according to the observation weight determined by the observation distance from the point cloud observation point to the sensor and the incident geometry relationship to obtain the local truncated signed distance field.

3. The method according to claim 2, characterized in that, The multi-frame vehicle-mounted laser point cloud includes a center frame and adjacent frames; in S2, based on the center frame, a local sliding window is constructed according to the acquisition time sequence of the center frame and adjacent frames. Only the point cloud observations within the local sliding window and falling into the local voxel calculation domain are used for the multi-frame recursive fusion, and the voxel size of the local voxel calculation domain is determined according to the point spacing of the multi-frame vehicle-mounted laser point cloud and the scale of the region of interest of the pit semantic region.

4. The method according to claim 2, characterized in that, In S2, the local surface normal is estimated based on the neighborhood point set of each point in the multi-frame vehicle-mounted laser point cloud. The observation weight is determined based on the observation distance and the incident angle between the observation ray and the local surface normal represented by the line-of-sight information. The observation weight of point cloud observations that conflict with the current symbol of the local truncated symbol distance field is reduced by the symbol consistency judgment.

5. The method according to claim 2, characterized in that, In S3, the zero isosurface is extracted to obtain a continuous triangular mesh surface. The continuous triangular mesh surface is then resampled at high resolution to obtain the high-resolution surface point set. In S4, normal road surface points are selected from the neighborhood of the pothole semantic interest region and fitted to the local reference road surface. The high-resolution surface point set is projected onto the road plane of the local road coordinate system and discretized into regular grids. The directed residuals in each regular grid are aggregated. In the continuous computation domain formed by projecting the pothole semantic interest region onto the road plane, data consistency constraints are applied to the regular grids with effective observations, and the regular grids without observations are smoothly extended to obtain the continuous residual depth field.

6. The method according to claim 5, characterized in that, In S4, the median of the directed residuals falling into the same regular grid is determined as the observation residual of the corresponding regular grid, and the continuous residual depth field is solved by a variational model that includes a data consistency term that contains the effective observation location and a field smoothing term in the continuous computation domain.

7. The method according to claim 5, characterized in that, In S5, the contour evolution domain is determined based on the projection region of the pothole semantic interest region on the road plane and the preset residual threshold of the continuous residual depth field. A regional active contour energy is constructed within the contour evolution domain. The regional active contour energy includes the boundary length term of the level set contour, the residual fitting term inside the contour, the residual fitting term outside the contour, and the semantic consistency term between the level set function and the indicator function of the pothole semantic interest region. The level set function is updated to convergence based on the regional active contour energy, and the region enclosed by the zero level set of the level set function is determined as the closed pothole region.

8. The method according to claim 5, characterized in that, In S1, the fixed offsets from the sensor reference point to the left wheel contact point and the right wheel contact point, as well as the effective tire contact width, are also obtained. In S6, the short-time pose sequence formed by the sensor poses is projected onto the road plane, and the vehicle's travel direction is determined based on the projection displacement of the sensor poses on the road plane at the current and next moments. Each sensor pose is combined with the fixed offset and mapped onto the road plane to obtain the left wheel track centerline and the right wheel track centerline. Points whose planar distance from the corresponding wheel track centerline is not greater than half of the effective tire contact width are used to form the corresponding wheel track strip domain. The projection range of the corresponding intersection region in the vehicle's travel direction is determined as the corresponding unilateral directional equivalent diameter.

9. The method according to claim 1, characterized in that, In S7, the points within the closed pit area are triangulated to obtain an irregular triangular network. Triangles with a maximum side length greater than a preset side length threshold are removed. The pit volume is obtained by piecewise integration based on the area of ​​each retained triangle and the directed residual of each vertex in the continuous residual depth field.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 9.

Images