A method and apparatus for generating degraded labels based on point cloud geometric statistics
By constructing local neighborhood and spherical normal histograms to automatically generate degradation supervision labels, the problem of relying on online registration and manual annotation in existing technologies is solved. This achieves efficient and automated generation of degradation category and direction labels, improving the independence and robustness of label generation.
Patent Information
- Application Number
- CN202611264739.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-20
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2046-08-20
AI Technical Summary
Existing technologies struggle to efficiently and automatically generate geometrically consistent degradation category and orientation supervision labels from single-frame LiDAR point clouds, relying on online registration and manual annotation, which are computationally expensive and lack adaptability.
By constructing a local neighborhood and estimating the normal vector, a normal histogram is constructed by mapping it to the spherical orientation space. The dominant normal region is extracted, geometric statistical indices are calculated, and degradation supervision labels, including category and orientation labels, are automatically generated.
It reduces the dependence on the SLAM backend state, provides refined degradation category supervision and structure-related axial direction supervision, reduces data construction costs, and ensures label consistency and reproducibility.
Smart Images

Figure CN122799211B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of lidar point cloud data processing technology, and specifically relates to a method and device for generating degraded labels based on point cloud geometric statistics. Background Technology
[0002] LiDAR (Light Detection and Ranging) can directly acquire 3D geometric information of the environment and has been widely used in Simultaneous Localization and Mapping (SLAM) systems. However, LiDAR SLAM is prone to degradation in scenarios with insufficient geometric constraints, such as long straight alleys, corridors, and open areas. This weakens the constraint capability of point cloud registration in some directions, affecting the accuracy and robustness of localization and mapping.
[0003] Training a degradation recognition network requires a large number of point cloud samples with degradation category and orientation information. Existing degradation label construction methods have the following main shortcomings: they rely on residual or covariance information from registration processes such as Iterative Closest Point (ICP) to statistically determine the degradation state, requiring online matching results or local map support, making it difficult to obtain training labels directly from a single frame of point cloud, resulting in high computational costs and dependence on the SLAM backend state; if only relying on manual experience to label tunnel-like, open-like, and non-degenerate scenes, it is difficult to guarantee the consistency of boundary samples, resulting in poor label consistency and insufficient cross-scene adaptability; some methods statistically determine the degree of degradation through multiple registration experiments, which has high computational overhead and is not suitable for label generation of large-scale training data.
[0004] In summary, existing degradation label construction methods suffer from the aforementioned prominent problems, and there is an urgent need for a technical solution that does not rely on online registration, requires no manual annotation, and can automatically generate degradation supervision labels directly from the geometric structure of a single-frame point cloud. Summary of the Invention
[0005] To address the aforementioned technical problems, this invention proposes a method and device for generating degradation labels based on point cloud geometric statistics. It can automatically generate geometrically consistent degradation category and orientation labels directly from a single frame of LiDAR point cloud, eliminating the need for online registration and manual annotation, thus effectively reducing the cost of training data construction.
[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, this invention proposes a method for generating degenerate labels based on point cloud geometric statistics, comprising the following steps: Local neighborhoods are constructed using each point in a preprocessed single-frame lidar point cloud. The local surface normal vectors of each point are determined based on the spatial distribution of points within each local neighborhood, resulting in a set of normal vectors. Map each normal vector in the set of normal vectors to each interval in the spherical orientation space and count them to construct a spherical normal histogram; In the spherical normal histogram, the histogram interval whose count satisfies the local maximum condition is taken as the main peak, and the histogram intervals located in the neighborhood of the main peak and whose count satisfies the merging condition are merged into the main peak to form at least one dominant normal region. Based on the counts of each dominant normal region and the average direction of each dominant normal region, at least one geometric statistical index is calculated to characterize the degree of concentration of normal vector distribution; based on the comparison between the geometric statistical index and the corresponding threshold, the scene category label is determined; When the scene category label is tunnel-type degradation, the axial degradation direction is determined based on the average direction of at least two dominant normal regions, and the axial degradation direction is quantized into discrete direction interval labels; The scene category label and the direction interval label are output as degradation supervision labels.
[0007] Furthermore, the method also includes: when the scene category label is open, degenerate, or non-degenerate, the direction interval label is set to an invalid value; the degradation supervision label contains identification information used to distinguish whether the direction interval label is valid or invalid.
[0008] Furthermore, the local surface normal vectors of each point are determined based on the spatial distribution of points within each local neighborhood, resulting in a set of normal vectors. Specifically, for any point in the preprocessed single-frame lidar point cloud... ,by Centered on, with Construct a local neighborhood for the radius Calculate the neighborhood covariance matrix: ; in, Point The neighborhood covariance matrix; Point The number of domain points; Representation of domain The first in One point; Represents the average coordinates of neighboring points; right Perform eigenvalue decomposition, and take the unit eigenvector corresponding to the smallest eigenvalue as a point. Local surface normal vector .
[0009] Furthermore, each normal vector in the set of normal vectors is mapped to an interval in the spherical orientation space and counted to construct a spherical normal histogram, specifically as follows: Each normal vector in the set of normal vectors Mapped to a spherical coordinate system, according to the azimuth angle and pitch angle The spherical orientation space is divided into multiple histogram intervals, and the number of normal vectors falling into each interval is counted to construct a spherical normal histogram; where the azimuth resolution is denoted as... The pitch angle resolution is denoted as ,and The spherical intervals corresponding to the positive and negative directions of the normal vector are merged into the same counting unit.
[0010] Furthermore, in the spherical normal histogram, the histogram intervals whose counts satisfy the local maximum condition are taken as the main peak, and the histogram intervals located in the neighborhood of the main peak and whose counts satisfy the merging condition are merged into the main peak, forming at least one dominant normal region, specifically: In the spherical normal histogram, the histogram interval with a count greater than the counts of all adjacent intervals is selected as the main peak; Centered on the main peak, adjacent histogram intervals located in the neighborhood of the main peak and whose counts reach the preset ratio threshold of the main peak count are incorporated into the main peak; When multiple main peaks are determined, neighborhood merging is performed on each main peak to form multiple dominant normal regions. Histogram intervals within the same dominant normal region share the same main peak direction, while different dominant normal regions are separated from each other in spherical direction space.
[0011] Furthermore, based on the counts of each dominant normal region and the average direction of each dominant normal region, at least one geometrical statistical index is calculated to characterize the degree of concentration of the normal vector distribution, specifically: From the dominant normal region, determine the first dominant normal region with the largest number of normal vectors and the second dominant normal region with the second largest number of normal vectors. The number of normal vectors in the first dominant normal region and the second dominant normal region are respectively... and ,and The corresponding average directions are respectively and ,in, ; The geometric statistical indicators include: normal vector distribution factor and relative peak ratio; Normal vector distribution factor Represented as: ; in, Indicates the first Counting of each residual histogram interval; Represents a very small positive number; This represents the set of residual intervals excluding the first dominant normal region and the second dominant normal region; relative main peak ratio Represented as: ; Vertical angle Represented as: ; in, It is a unit vector in the vertical direction.
[0012] Furthermore, based on the comparison results between the geometric statistical indicators and the corresponding thresholds, the scene category label is determined, specifically as follows: when and At that time, the scene category was determined to be tunnel-type degradation; when , and At that time, the scene category was determined to be open space degradation; In all other cases, the scene category is determined to be non-degenerate; in, This represents the preset threshold for the normal vector distribution; This indicates the preset threshold for the main peak intensity ratio; This indicates the preset vertical angle threshold, and .
[0013] Furthermore, when the scene category label is tunnel-type degradation, the axial degradation direction is determined based on the average direction of at least two dominant normal regions, specifically: From the first dominant method towards the regional average direction With the second dominant normal region average direction The cross product determines the three-dimensional axial degradation direction vector. :
[0014] in, This represents the three-dimensional axial degradation direction vector.
[0015] Furthermore, the axial degradation direction is quantized into discrete direction interval labels; the scene category label and the direction interval labels are output as degradation supervision labels; specifically: Projecting the axial degradation direction onto a horizontal plane yields the horizontal direction; Map the horizontal direction to a preset angle range; The angle range is divided into multiple discrete directional intervals according to a preset angular resolution, and the directional interval number to which the horizontal direction belongs is determined. When the scene category label is open, degenerate, or non-degenerate, the direction interval label is set to an invalid value; The scene category label and the direction interval label are output together as the degradation supervision label.
[0016] Secondly, the present invention proposes a degradation label generation device based on point cloud geometric statistics, comprising at least one processor and a memory, wherein the memory stores a computer program, and the computer program, when executed by the at least one processor, implements the degradation label generation method based on point cloud geometric statistics.
[0017] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. One of the above technical solutions has the following advantages or beneficial effects: (i) Reduce dependence on SLAM backend state Existing degraded label generation methods typically rely on residual or covariance information from registration processes such as Iterative Closest Point (ICP), requiring online matching results or local map support. Label quality is significantly affected by the accuracy of SLAM backend state estimation. This invention directly utilizes the geometric and statistical features of a single-frame LiDAR point cloud to generate degraded labels, without relying on online registration residuals, registration covariance, or local maps. It determines degradation based on the point cloud's own normal vector distribution, effectively reducing the dependence of label generation on SLAM backend state and improving the independence and robustness of the label generation method.
[0018] (ii) Provide detailed monitoring of degradation categories Most existing methods only output a binary determination of whether point clouds exhibit degradation, lacking detailed differentiation of degradation types. This invention extracts the dominant normal peak value through a spherical histogram of normal vectors and combines it with three geometric statistical indicators—normal vector distribution factor, relative dominant peak ratio, and vertical angle—to subdivide scenes into three categories: tunnel-like degradation, open-like degradation, and non-degenerate. Compared to labeling methods that only determine whether degradation occurs, this invention provides richer category supervision information for degradation recognition networks, helping the network learn the intrinsic feature differences of different degradation patterns and improving the network's generalization ability in complex real-world environments.
[0019] (iii) Provide structural axial direction monitoring Most existing methods do not provide degradation direction information, causing degradation recognition networks to be unable to learn the specific direction in which degradation occurs. This invention targets tunnel-type degradation scenarios, using the cross product of the average directions of two dominant normal regions to automatically estimate the axial degradation direction of the tunnel, and quantizing continuous directions into discrete direction interval labels. This provides effective supervision information for subsequent network learning of structure-related degradation directions, which is beneficial for achieving more accurate pose estimation optimization in degradation scenarios.
[0020] (iv) Reduce data construction costs and ensure label consistency Existing methods that rely solely on manual experience to label degraded tunnel scenes, degraded open scenes, and non-degraded scenes are not only extremely labor-intensive, but also suffer from inconsistent judgment criteria for boundary samples among different annotators, making it difficult to guarantee label quality. This invention eliminates the need for manual frame-by-frame annotation, automating the entire process from point cloud input to label output. This significantly reduces the cost of constructing training data for degradation recognition networks while ensuring the consistency and reproducibility of label generation rules across different samples, providing a feasible technical solution for constructing large-scale training datasets. Attached Figure Description
[0021] Figure 1 The flowchart is shown in Embodiment 1 of the present invention for a method of generating degraded labels based on point cloud geometric statistics. Figure 2 This is a schematic diagram of the framework of a degradation label generation method based on point cloud geometric statistics proposed in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the spherical distribution pattern of normal vectors under different degradation scenarios proposed in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of the single-frame degradation tag generation results for three typical scenarios proposed in Embodiment 1 of the present invention; Figure 5 This is a schematic diagram of the degradation label generation result of the continuous point cloud sequence proposed in Embodiment 1 of the present invention; Figure 6 This is a schematic diagram of a degradation label generation device based on point cloud geometric statistics proposed in Embodiment 2 of the present invention. Detailed Implementation
[0022] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific embodiments and in conjunction with the accompanying drawings. The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure of the invention, components and arrangements of specific examples are described below. Furthermore, reference numerals and / or letters may be repeated in different examples. This repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. It should be noted that the components illustrated in the drawings are not necessarily drawn to scale. Descriptions of well-known components, processing techniques, and processes are omitted in this invention to avoid unnecessarily limiting the invention.
[0023] Example 1 The present invention, in embodiment 1, proposes a degradation label generation method based on point cloud geometric statistics, which addresses the technical problem that existing degradation label construction methods rely on online registration results or manual frame-by-frame annotation, making it difficult to efficiently and consistently generate geometry-driven degradation category and orientation supervision labels from a single frame of point cloud.
[0024] This invention starts with a single frame of LiDAR point cloud and calculates geometric statistical indices reflecting the concentrated characteristics of normal distribution through normal vector estimation, spherical normal histogram construction, dominant normal peak extraction, and neighborhood merging. Based on these indices, it automatically generates three types of scene labels: tunnel-like degeneracy, open-like degeneracy, and non-degenerate. For tunnel-like samples, this invention further estimates the tunnel axial direction and generates direction bin labels.
[0025] Figure 1 The flowchart is shown in Embodiment 1 of the present invention for a method of generating degraded labels based on point cloud geometric statistics. Figure 2 This is a schematic diagram of the framework of a degradation label generation method based on point cloud geometric statistics proposed in Embodiment 1 of the present invention; combined with Figure 1 and Figure 2 The process of implementing this invention will be explained together.
[0026] In step S1, a single frame of lidar point cloud is acquired, and the single frame lidar point cloud is preprocessed to obtain a set of point clouds to be analyzed.
[0027] Specifically, the environment is scanned by a lidar sensor at a certain sampling time to acquire a single frame of 3D point cloud data corresponding to that time. The single frame lidar point cloud contains multiple 3D spatial points, each with 3D coordinate information (x, y, z) in the lidar coordinate system. The acquired single frame lidar point cloud is preprocessed to filter out noise and invalid points, retaining valid environmental geometric information. The preprocessing includes the following sub-steps: (1) Distance filtering: Set the maximum sensing distance threshold and remove outliers that are more than the distance to the center of the lidar. The distant point cloud is sparse and noisy, which may affect the stability of the normal vector estimation.
[0028] (2) Coordinate preservation: The three-dimensional coordinate information of each point in the lidar coordinate system is preserved without coordinate transformation or registration operation, so as to ensure that the subsequent normal vector estimation and histogram construction are based on the original observation data.
[0029] (3) Optional downsampling: Based on actual application requirements, voxel mesh downsampling is performed on the point cloud to reduce the amount of data and improve computational efficiency. The downsampled point cloud can retain the main geometric features of the scene while reducing the consumption of computational resources by redundant points.
[0030] After the above preprocessing, the point cloud set to be analyzed is obtained, denoted as... ,in, This represents the number of points in the preprocessed point cloud. Indicates the first The three-dimensional coordinate vector of a point in the lidar coordinate system.
[0031] It is worth noting that the preprocessing described in this step does not involve the registration of point clouds with local maps, nor does it involve the acquisition of pose estimation between adjacent frames. That is, this method does not rely on any SLAM backend state information and is based entirely on the geometric information of the single-frame point cloud itself for subsequent processing.
[0032] In step S2, local neighborhoods are constructed using each point in the preprocessed single-frame lidar point cloud, and the local surface normal vectors of each point are determined based on the spatial distribution of points within each local neighborhood, thus obtaining a set of normal vectors. Specifically, for any point in the point cloud to be analyzed ,by Centered on, with Given the neighborhood radius, the search distance in three-dimensional space. No more than All neighboring points constitute the local neighborhood set of that point. .
[0033] Neighborhood radius The value is determined based on the point cloud density and the application scenario. Preferably, for common mechanical lidar points... A value of 0.3m to 1.0m can be used. If the neighboring area... Points in If the number of points is less than the preset minimum threshold (e.g., 10 points), it indicates that the point cloud in the area where the point is located is too sparse, and the normal vector estimation is unreliable. The point can be marked as invalid and removed in subsequent processing.
[0034] For points whose number of neighboring points meets the requirement Calculate the average coordinates of all points in its local neighborhood (i.e., the neighborhood centroid). ; Calculate the neighborhood covariance matrix: ; in, Point The neighborhood covariance matrix reflects the spatial distribution of neighborhood points around the centroid: if the variance of the distribution of neighborhood points along a certain direction is minimized, then that direction corresponds to the normal direction of the local plane. Point The number of domain points; Representation of domain The first in One point; Represents the average coordinates of neighboring points; right Perform eigenvalue decomposition, and take the unit eigenvector corresponding to the smallest eigenvalue as a point. Local surface normal vector .
[0035] It should be noted that, and Both represent the normal direction of the same local plane. In the subsequent spherical histogram statistics, this invention performs equivalent merging of the positive and negative directions to eliminate the influence of uncertainty in the sign of the normal vector.
[0036] In step S3, each normal vector in the set of normal vectors is mapped to each interval of the spherical orientation space and counted to construct a spherical normal histogram; Specifically: each normal vector in the set of normal vectors Mapped to a spherical coordinate system, according to the azimuth angle and pitch angle The spherical orientation space is divided into multiple histogram intervals, and the number of normal vectors falling into each interval is counted to construct a spherical normal histogram; where the azimuth resolution is denoted as... The pitch angle resolution is denoted as The spherical relative intervals corresponding to the positive and negative directions of the normal vector are merged into the same counting unit.
[0037] Let the number of azimuth directions be . The number of pitch angle directions is Then the azimuth resolution is The pitch resolution is ,and Preferably, You can take 60 to 120. The value can be between 30 and 60, and the specific value is determined based on the point cloud density and the desired directional resolution.
[0038] For each normal vector, determine its histogram interval index based on its azimuth and elevation angles, and increment the count for that interval. Iterate through all normal vectors in the set, count the number of normal vectors falling into each interval, and construct a spherical normal histogram.
[0039] Due to the normal vector and This represents the normal direction of the same local plane in space (only the signs are opposite), and they should be treated equally when statistically analyzing the distribution of normal vector directions. Therefore, this invention merges two spherical intervals that are opposite each other passing through the center of the sphere into the same counting unit. That is, the counts of normal vectors falling into the two opposite intervals are merged, and the merged count value is used as the counting result of the same normal vector statistical unit.
[0040] Figure 3 This is a schematic diagram of the spherical distribution pattern of normal vectors under different degradation scenarios in Embodiment 1 of the present invention. Each column corresponds to a scene type: the left columns (a), (d), and (g) correspond to non-degradation scenarios; the middle columns (b), (e), and (h) correspond to open-type degradation scenarios; and the right columns (c), (f), and (i) correspond to tunnel-type degradation scenarios. Each row corresponds to a data format: the first row (a), (b), and (c) are single-frame point clouds for each scene; the second row (d), (e), and (f) are the corresponding spherical normal histograms; and the third row (g), (h), and (i) are the extracted dominant normal regions.
[0041] In the non-degraded scenes (a), (d), and (g) in the left column, the point cloud contains various structures such as building facades, trees, and streetlights, exhibiting rich geometric features. The normal vectors in the spherical normal histogram are dispersed, without forming a clear dominant peak. The number of extracted dominant normal regions is relatively large, and the counts of each region are relatively close. In the open-type degraded scenes (b), (e), and (h) in the middle column, the point cloud is dominated by large areas of horizontal ground, lacking significant vertical structures. The normal vectors in the spherical normal histogram are concentrated near the vertical direction, presenting a single dominant peak. The extracted dominant normal region is mainly a region near the vertical direction. In the tunnel-type degraded scenes (c), (f), and (i) in the right column, the point cloud presents a long, straight tunnel structure, including side walls, ground, and roof. The normal vectors in the spherical normal histogram are concentrated in both the vertical and horizontal directions, presenting a dual dominant peak pattern. The extracted dominant normal regions include two regions with similar intensities in the vertical and horizontal directions.
[0042] Figure 3 The invention visually demonstrates the complete process of extracting different types of scenes from raw point clouds to spherical normal histograms and then to dominant normal regions, verifying the effective ability of the method of the present invention to distinguish different degradation scenes.
[0043] In step S4, in the spherical normal histogram, the histogram interval whose count satisfies the local maximum condition is taken as the main peak, and the histogram intervals located in the neighborhood of the main peak and whose count satisfies the merging condition are merged into the main peak, forming at least one dominant normal region; specifically: In the spherical normal histogram, the histogram intervals with counts greater than those of all adjacent intervals are selected as the main peaks. For each histogram interval, its count is compared with the counts of all its adjacent intervals (including neighboring intervals in the azimuth and pitch dimensions). If the count of this interval is greater than the counts of all its adjacent intervals, then this interval is marked as a local maximum point, i.e., a candidate main peak. To eliminate local maxima caused by noise, the candidate main peaks can be further filtered. Preferably, a number of intervals with the highest count ranking in the spherical normal histogram are retained as a candidate set, for example, the top one-third of the intervals are retained. Then, the intervals with counts greater than those of all adjacent intervals in the candidate set are selected as the main peaks. This filtering strategy can effectively avoid a large number of noisy intervals being misjudged as main peaks, while ensuring that the peaks corresponding to large-area planar structures with physical significance are correctly extracted.
[0044] Centered on the main peak, adjacent histogram intervals located within the main peak's neighborhood and whose counts reach a preset percentage threshold for the main peak's count are merged into the main peak. Specifically, histogram intervals located within the main peak's neighborhood (i.e., intervals adjacent to or within a certain range of the main peak in the azimuth and elevation directions) are checked. If the count of a certain adjacent interval reaches a preset percentage threshold for the main peak's count (preferably, this threshold can be set to 50% of the main peak's count), then that adjacent interval is merged into the main peak. This merging process is repeated until there are no more intervals in the main peak's neighborhood that meet the merging conditions. Through this merging operation, peak splitting caused by normal vector estimation errors and histogram discretization boundary effects on the same physical plane can be effectively avoided, ensuring that each dominant normal region can completely cover its corresponding continuous direction interval.
[0045] When multiple main peaks are identified, the aforementioned neighborhood merging operation is performed on each main peak to form multiple dominant normal regions. Histogram intervals within the same dominant normal region share the same main peak direction, while different dominant normal regions are separated from each other in spherical spatial orientation.
[0046] When multiple main peaks are determined, neighborhood merging is performed on each main peak to form multiple dominant normal regions. Histogram intervals within the same dominant normal region share the same main peak direction, while different dominant normal regions are separated from each other in spherical direction space.
[0047] In the spherical normal histogram, the top polar region (i.e., the pitch angle is close to 0 or π, corresponding to a normal vector that is close to vertically upward or downward) has a small actual directional distance between adjacent intervals due to the singularity of the spherical coordinate system. This can easily lead to multiple adjacent intervals being detected as the main peak simultaneously. These normal vectors actually correspond to the vertical normal distribution generated by large-area horizontal structures (such as the ground or top surface). Therefore, this invention merges multiple adjacent peaks in the top polar region into a single interval to maintain the stability of the determination of open-type degradation scenes.
[0048] After the above-mentioned main peak detection, neighborhood merging, and top poloidal region processing, at least one dominant normal region is obtained. For ease of subsequent description, the dominant normal region with the most normal vectors is denoted as the first dominant normal region, and the dominant normal region with the second most normal vectors is denoted as the second dominant normal region.
[0049] The constructed spherical normal histogram effectively characterizes the degree of clustering of normal vector directions in a scene: In non-degenerate scenes, the environment contains a variety of planar structures, and the distribution of normal vector directions in the spherical space is relatively dispersed, without forming a clear dominant peak; In open degenerate scenes, the environment is mainly composed of horizontal ground and open sky, and a large number of normal vectors point upwards or downwards, thus exhibiting a single dominant peak near the vertical direction; In tunnel degenerate scenes, the environment includes side walls, ground, and top surface, and the normal vectors are concentrated in both the vertical and horizontal directions, thus exhibiting a dual dominant peak pattern.
[0050] In step S5, at least one geometric statistical index is calculated to characterize the degree of concentration of normal vector distribution based on the count of each dominant normal region and the average direction of each dominant normal region; and the scene category label is determined based on the comparison result between the geometric statistical index and the corresponding threshold. Specifically, from the dominant normal region, a first dominant normal region with the largest number of normal vectors and a second dominant normal region with the second largest number of normal vectors are determined. The number of normal vectors in the first dominant normal region and the second dominant normal region are respectively... and ,and .
[0051] For each dominant normal region, the average direction of all normal vectors within it needs to be calculated. Since histogram intervals within the same dominant normal region share the same dominant peak direction, but the normal vectors within each interval may have slight differences in direction, when calculating the average direction, all normal vectors within the region are first unified to the same sign and direction as the dominant peak direction. Then, all normal vectors are vector-superimposed and normalized to obtain the average direction of the region. Let the average directions of the first and second dominant normal regions be respectively... and ,in, ; This invention defines three geometric statistical indicators to characterize the concentration and directional features of the normal vector distribution from different perspectives. Specifically, these include: normal vector distribution factor, relative peak ratio, and vertical angle.
[0052] Normal vector distribution factor This is used to measure the overall concentration of normal vectors in spherical directional space. The basic idea is to treat all residual regions other than the first and second dominant normal vector regions as background distribution, and determine whether the normal vectors are concentrated in a few dominant directions by comparing the relative proportions between the residual normal vectors and the two dominant normal vectors.
[0053] Normal vector distribution factor Represented as: ; in, Indicates the first Counting of each residual histogram interval; It represents extremely small positive numbers and is used to avoid invalid calculations when the logarithm is zero or negative. This represents the set of residual intervals excluding the first dominant normal region and the second dominant normal region.
[0054] Normal vector distribution factor The formula means that the numerator represents the sum of the logarithms of the residual normal vectors, and the denominator represents the sum of the logarithms of the two dominant normal vectors. The smaller the value, the more concentrated the normal vectors are in a few dominant directions, and the more likely the current point cloud is in a degraded environment (such as a long straight alley or open area); conversely, The larger the value, the more dispersed the normal vector distribution, and the more likely the current point cloud is in a non-degradable environment (such as urban roads or densely built-up areas).
[0055] relative main peak ratio Represented as: ; The relative peak ratio reflects the relative importance between the two main normal aggregation directions. When When the intensity is large, it indicates that the two dominant normal regions are close in strength, and the normal vector is approximately uniformly distributed between the two main peaks. This is consistent with the dual-peak pattern in tunnel-type degradation scenarios (such as the ground / top surface and the two side walls); when When the value is smaller, it indicates that the second peak is significantly weaker than the first peak, and the normal vector is mainly concentrated in a single direction, which is closer to the single peak pattern of open degenerate scenes (such as ground and open sky).
[0056] Vertical angle The degree of deviation of the average direction of the first dominant normal region from the vertical direction is expressed as: ; in, It is a unit vector in the vertical direction.
[0057] when When the value is small, it indicates that the average direction of the first dominant normal region is close to the vertical direction, meaning that there are large areas of horizontal structures (such as the ground or top surface) in the scene, which is a typical feature of open-type degraded scenes.
[0058] Based on the comparison between the geometric statistical indicators and the corresponding thresholds, the scene category label is determined, specifically as follows: when and At that time, the scene category was determined to be tunnel-type degradation; this condition indicates that the overall distribution of normal vectors is concentrated ( Small), and the strengths of the two dominant directions are close ( (Large), conforming to the characteristics of degraded scenarios such as long straight alleys and corridors with two mutually perpendicular dominant plane directions.
[0059] when , and When the scene category is determined to be open and degenerate, this condition indicates that the overall distribution of normal vectors is concentrated ( Small), but the second dominant direction is significantly weaker than the first dominant direction ( Small), and the first dominant direction is close to the vertical direction ( (Small) It conforms to the characteristics of degraded scenes such as open areas and squares where the main structure is a horizontal ground.
[0060] In other cases, the scene category is determined to be non-degenerate; that is, the normal vector distribution is relatively dispersed. If the distribution is concentrated but does not meet the characteristic conditions of the above two degradation modes, it indicates that the scene contains a sufficiently rich multi-directional planar structure, and the point cloud registration has good constraints in all directions.
[0061] in, This represents the preset threshold for the normal vector distribution; This indicates the preset threshold for the main peak intensity ratio; This indicates the preset vertical angle threshold, and The three preset thresholds mentioned above can be determined through statistical analysis or experimental verification based on the actual application scenario and the distribution of training data. Typically, A value of 0.3 to 0.6 is acceptable. A value of 0.4 to 0.7 is acceptable. An angle of 15° to 30° is acceptable.
[0062] Figure 4 This is a schematic diagram of the single-frame degradation label generation results for three typical scenarios proposed in Embodiment 1 of the present invention.
[0063] In open environments (non-degraded) Figure 4 a to Figure 4 (c) In this context, pedestrian trajectories are relatively dispersed, with no significant physical boundary constraints. The heatmap distribution of their lateral deflection angles is relatively uniform, covering a wide area; the corresponding geometric statistical indicators show that... 0.65 It is 0.95 and It is 0.34.
[0064] In narrow passage environments (tunnel-like degradation) Figure 4 d to Figure 4 In the f), the statistical indicator changed significantly. The value dropped sharply to 0.27. The value is 0.91, while The value increased significantly to 0.63.
[0065] In long, straight corridor environments (tunnel-like degradation) Figure 4 g to Figure 4 In the i), the statistical indicators further evolved: The value dropped to a minimum of 0.14. It remained at 0.97, while It reached a maximum of 0.87.
[0066] In Embodiment 1 of the present invention, if four symmetrical main peaks are detected, two main peaks in the horizontal direction (or the ground and any wall) are selected for cross product, or multiple possible axial candidates are voted on or clustered to improve the robustness of the direction estimation.
[0067] In step S6, when the scene category label is tunnel-type degradation, the axial degradation direction is determined based on the average direction of at least two dominant normal regions, and the axial degradation direction is quantized into discrete direction interval labels; the scene category label and the direction interval labels are output as degradation supervision labels.
[0068] When the scene category label is open, degenerate, or non-degenerate, the direction interval label is set to an invalid value; the degeneracy supervision label contains identification information used to distinguish whether the direction interval label is valid or invalid.
[0069] Specifically, from the first dominant method towards the regional average direction With the second dominant normal region average direction The cross product determines the three-dimensional axial degradation direction vector. :
[0070] in, This represents the axial degeneracy direction vector in three dimensions. × represents the cross product of vectors. The L2 norm represents the three-dimensional candidate degradation direction, which corresponds to the axial weak constraint direction in a long straight tunnel or similar structure.
[0071] Subsequently, the system performs a direction interval quantization step. The continuously calculated three-dimensional axial degradation directions are then... The data is mapped to the system's preset discrete space division standard (e.g., dividing 360-degree space into several fixed angle intervals) and converted into corresponding discrete direction interval labels.
[0072] Pick The components in the horizontal plane are normalized to obtain the horizontal axial direction. Since the positive and negative directions of the tunnel axis are equivalent in a degenerative sense, the axial angle is... Defined in [0,π), and according to directional resolution Divided into There are 10 bins for each direction. The direction label corresponding to the actual direction is... ,in, Preferably, the first At that time, the angular interval between adjacent bin directions is 15°.
[0073] Next, the system performs special labeling processing for non-tunnel degradation scenarios. When the scenario category label is "open degradation" or "non-degradation", since there are no obvious axial physical boundary constraints in this type of scenario, the system forces the direction interval label to an invalid value (such as NULL or a specific placeholder).
[0074] Finally, the system performs a fusion output of the supervised labels. Scene category labels and orientation interval labels (including valid and invalid states) are coupled to generate the final degradation supervised labels. These degradation supervised labels, in their data structure, not only contain the physical degradation attributes of the scene but also identification information used to distinguish whether the orientation interval labels are valid (tunnel scene) or invalid (open / non-degraded scene). These standardized labels provide ground truth information for subsequent deep learning model training, encompassing both the scene's physical and orientation features.
[0075] Figure 5 This is a schematic diagram of the degradation label generation result of the continuous point cloud sequence proposed in Embodiment 1 of the present invention. Figure 5 (a) and Figure 5 (d) in the figure corresponds to the degradation indicator factor of the two sequences as a function of time, which clearly reflects the transformation process of the scene between non-degradation, open degradation and tunnel-like degradation. When the curve is lower than the preset threshold, it indicates that the tunnel-like degradation scene is triggered. Figure 5 (b) and Figure 5 (e) shows the trend of the relative dominance ratio, indicating that none of the above sequences have met the criteria for dominance degradation. Figure 5 (c) and Figure 5 (f) in the figure represents the statistical results of the number of samples in different scenarios. It intuitively compares the specific distribution ratio of open degradation, non-degradation and tunnel-like degradation frames in each sequence, and verifies the ability of the degradation supervision label generation module to distinguish different scenario categories.
[0076] pass Figure 4 and Figure 5As shown, the present invention can output degradation labels consistent with scene geometry in single frames and continuous sequences. Open scenes are mainly characterized by a vertical main peak and a weaker second main peak, while alley scenes are characterized by a lower normal vector distribution factor and a higher relative main peak ratio.
[0077] In summary, the degradation label generation method based on point cloud geometric statistics proposed in Embodiment 1 of this invention constructs a local neighborhood of a single-frame LiDAR point cloud and extracts surface normal vectors, then identifies the dominant normal region based on a spherical normal histogram; on this basis, it innovatively constructs a normal vector distribution factor. relative main peak ratio and the vertical angle This method employs multi-dimensional geometric statistical indicators to automate and refine the assessment of three typical scenarios: tunnel-type degradation, open-type degradation, and non-degradation. For tunnel-type degradation scenarios, it accurately calculates the axial degradation direction through cross-product operations on the dominant normal region and quantifies it into discrete direction interval labels using an invalid value mechanism, ultimately outputting structured degradation supervision labels. This method is entirely based on the geometric features of a single-frame point cloud, eliminating the need for pose estimation or local maps in the SLAM backend and avoiding the subjectivity and high costs of manual frame-by-frame annotation. It provides standardized training data with good consistency, low coupling, and rich degradation direction information for deep learning degradation recognition networks, significantly improving the efficiency and quality of constructing large-scale training datasets.
[0078] Example 2 Based on the degradation label generation method based on point cloud geometric statistics proposed in Embodiment 1 of the present invention, Embodiment 2 of the present invention also proposes a degradation label generation device based on point cloud geometric statistics. Figure 6 This is a schematic diagram of a degradation label generation device based on point cloud geometric statistics proposed in Embodiment 2 of the present invention.
[0079] At the hardware level, the electronic device 600 includes a processor 610, and optionally, an internal bus 620, a network interface 630, and memory. The memory may include main memory, such as high-speed random-access memory (RAM), or it may also include non-volatile memory, such as at least one disk drive. Of course, the electronic device may also include other hardware required for other business operations.
[0080] The processor 610, network interface 630, and memory can be interconnected via an internal bus 620. This internal bus 620 can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be categorized as an address bus, data bus, control bus, etc. For ease of illustration, only a single bidirectional arrow is used in this diagram, but this does not imply that there is only one bus or one type of bus. The memory is used to store programs. Specifically, the program can include program code, which includes computer operation instructions. The memory can include main memory 640 and non-volatile memory 650, and provides instructions and data to the processor 610.
[0081] Processor 610 reads the corresponding computer program from non-volatile memory 650 into memory 640 and runs it, forming a degradation tag generation device at the logical level. Processor 610 executes the program stored in memory and specifically performs the following: In step S1, a single frame of lidar point cloud is acquired, and the single frame lidar point cloud is preprocessed to obtain a set of point clouds to be analyzed.
[0082] In step S2, local neighborhoods are constructed using each point in the preprocessed single-frame lidar point cloud, and the local surface normal vectors of each point are determined based on the spatial distribution of points within each local neighborhood, thus obtaining a set of normal vectors. In step S3, each normal vector in the set of normal vectors is mapped to each interval of the spherical orientation space and counted to construct a spherical normal histogram; In step S4, in the spherical normal histogram, the histogram interval whose count satisfies the local maximum condition is taken as the main peak, and the histogram intervals located in the neighborhood of the main peak and whose count satisfies the merging condition are merged into the main peak to form at least one dominant normal region. In step S5, at least one geometric statistical index is calculated to characterize the degree of concentration of normal vector distribution based on the count of each dominant normal region and the average direction of each dominant normal region; and the scene category label is determined based on the comparison result between the geometric statistical index and the corresponding threshold. In step S6, when the scene category label is tunnel-type degradation, the axial degradation direction is determined based on the average direction of at least two dominant normal regions, and the axial degradation direction is quantized into discrete direction interval labels; the scene category label and the direction interval labels are output as degradation supervision labels.
[0083] When the scene category label is open, degenerate, or non-degenerate, the direction interval label is set to an invalid value; the degeneracy supervision label contains identification information used to distinguish whether the direction interval label is valid or invalid.
[0084] Figure 1 It can be applied to processor 610, or implemented by processor 610. The processor may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit in the processor or by instructions in the form of software. The processor mentioned above can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of this application can be directly embodied in the execution of the hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method.
[0085] Of course, in addition to software implementation, the electronic device of this application does not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. In other words, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.
[0086] The description of the relevant parts of the degradation label generation device based on point cloud geometric statistics provided in Embodiment 2 of this application can be found in the detailed description of the corresponding parts of the degradation label generation method based on point cloud geometric statistics provided in Embodiment 1 of this application, and will not be repeated here.
[0087] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that the elements inherent in a process, method, article, or apparatus that includes a list of elements are included. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. Additionally, portions of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of corresponding technical solutions in the prior art have not been described in detail to avoid excessive elaboration.
[0088] While specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art can make other modifications or variations based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.
Claims
1. A method for generating degradation labels based on point cloud geometry statistics, characterized in that, Includes the following steps: Local neighborhoods are constructed using each point in a preprocessed single-frame lidar point cloud. The local surface normal vectors of each point are determined based on the spatial distribution of points within each local neighborhood, resulting in a set of normal vectors. Map each normal vector in the set of normal vectors to each interval in the spherical orientation space and count them to construct a spherical normal histogram; In the spherical normal histogram, the histogram interval whose count satisfies the local maximum condition is taken as the main peak, and the histogram intervals located in the neighborhood of the main peak and whose count satisfies the merging condition are merged into the main peak to form at least one dominant normal region. Based on the counts of each dominant normal region and the average direction of each dominant normal region, at least one geometric statistical index is calculated to characterize the degree of concentration of normal vector distribution; based on the comparison between the geometric statistical index and the corresponding threshold, the scene category label is determined; When the scene category label is tunnel-type degradation, the axial degradation direction is determined based on the average direction of at least two dominant normal regions, and the axial degradation direction is quantized into discrete direction interval labels; The scene category label and the direction interval label are output as degradation supervision labels.
2. The method of claim 1, wherein, The method further includes: when the scene category label is open, degenerate, or non-degenerate, the direction interval label is set to an invalid value; the degradation supervision label contains identification information used to distinguish whether the direction interval label is valid or invalid.
3. The method according to claim 1, characterized in that, A local surface normal vector of each point is determined according to a spatial distribution of points in each local neighborhood, to obtain a normal vector set, specifically: for any point in the single-frame laser radar point cloud after preprocessing , a local neighborhood is constructed with as the center and as the radius , and a neighborhood covariance matrix is calculated ; in, Point The neighborhood covariance matrix; Point The number of domain points; Representation of domain The first in One point; Represents the average coordinates of neighboring points; right Perform eigenvalue decomposition, and take the unit eigenvector corresponding to the smallest eigenvalue as a point. Local surface normal vector .
4. The method according to claim 1, characterized in that, Each normal vector in the set of normal vectors is mapped to an interval in the spherical orientation space and counted to construct a spherical normal histogram, specifically as follows: Each normal vector in the set of normal vectors Mapped to a spherical coordinate system, according to the azimuth angle and pitch angle The spherical orientation space is divided into multiple histogram intervals, and the number of normal vectors falling into each interval is counted to construct a spherical normal histogram; where the azimuth resolution is denoted as... The pitch angle resolution is denoted as ,and The spherical intervals corresponding to the positive and negative directions of the normal vector are merged into the same counting unit.
5. The method according to claim 1, characterized in that, In the spherical normal histogram, the histogram intervals whose counts satisfy the local maximum condition are taken as the main peaks. The histogram intervals located in the neighborhood of the main peaks and whose counts satisfy the merging condition are merged into the main peaks to form at least one dominant normal region, specifically: In the spherical normal histogram, the histogram interval with a count greater than the counts of all adjacent intervals is selected as the main peak; Centered on the main peak, adjacent histogram intervals located in the neighborhood of the main peak and whose counts reach the preset ratio threshold of the main peak count are incorporated into the main peak; When multiple main peaks are determined, neighborhood merging is performed on each main peak to form multiple dominant normal regions. Histogram intervals within the same dominant normal region share the same main peak direction, while different dominant normal regions are separated from each other in spherical direction space.
6. The method according to claim 1, characterized in that, Based on the counts of each dominant normal region and the average direction of each dominant normal region, at least one geometrical statistical index is calculated to characterize the degree of concentration of the normal vector distribution, specifically: From the dominant normal region, determine the first dominant normal region with the largest number of normal vectors and the second dominant normal region with the second largest number of normal vectors. The number of normal vectors in the first dominant normal region and the second dominant normal region are respectively... and ,and The corresponding average directions are respectively and ,in, ; The geometric statistical indicators include: normal vector distribution factor and relative peak ratio; Normal vector distribution factor Represented as: ; in, Indicates the first Counting of each residual histogram interval; Represents a very small positive number; This represents the set of residual intervals excluding the first dominant normal region and the second dominant normal region; relative main peak ratio Represented as: ; Vertical angle Represented as: ; in, It is a unit vector in the vertical direction.
7. The method according to claim 6, characterized in that, Based on the comparison between the geometric statistical indicators and the corresponding thresholds, the scene category label is determined, specifically as follows: when and At that time, the scene category was determined to be tunnel-type degradation; when , and At that time, the scene category was determined to be open space degradation; In all other cases, the scene category is determined to be non-degenerate; in, This represents the preset threshold for the normal vector distribution; This indicates the preset threshold for the main peak intensity ratio; This indicates the preset vertical angle threshold, and .
8. The method according to claim 7, characterized in that, When the scene category label is tunnel-type degradation, the axial degradation direction is determined based on the average direction of at least two dominant normal regions, specifically: From the first dominant method towards the regional average direction With the second dominant normal region average direction The cross product determines the three-dimensional axial degradation direction vector. : in, This represents the three-dimensional axial degradation direction vector.
9. The method according to claim 1, characterized in that, The axial degradation direction is quantized into discrete direction interval labels; the scene category label and the direction interval label are output as degradation supervision labels; specifically: Projecting the axial degradation direction onto a horizontal plane yields the horizontal direction; Map the horizontal direction to a preset angle range; The angle range is divided into multiple discrete directional intervals according to a preset angular resolution, and the directional interval number to which the horizontal direction belongs is determined. When the scene category label is open, degenerate, or non-degenerate, the direction interval label is set to an invalid value; The scene category label and the direction interval label are output together as the degradation supervision label.
10. A degradation label generation device based on point cloud geometric statistics, comprising at least one processor and a memory, wherein the memory stores a computer program, characterized in that, When the computer program is executed by the at least one processor, it implements a degradation label generation method based on point cloud geometric statistics as described in any one of claims 1 to 9.
Citation Information
Patent Citations
Degradation detection method, system and equipment for underground tunnel environment
CN120254889A
Multi-sensor fusion anti-degradation SLAM mapping method and system
CN121353569A
