Object labeling method and system based on laser point cloud

Abstract

The invention discloses a laser point cloud-based object labeling method and system, which specifically relates to the technical field of laser point cloud processing, and is used to solve the problem of inaccurate labeling results caused by complex reflection phenomena caused by non-Lambertian surfaces in the prior art; by obtaining original point cloud data collected by a laser radar and performing preprocessing; identifying non-Lambertian reflection areas based on reflection intensity information and geometric topological features and performing filtering correction; correcting false mirror points through multi-layer iterative correction, and reconstructing boundaries to generate corrected point cloud data; grouping and classifying target objects in the corrected point cloud data based on a clustering algorithm combining global features and local features; using an attention mechanism network to finely divide the boundaries of preliminary segmentation results, and evaluating the uncertainty of boundary areas through local mutual information entropy; adjusting boundary division strategies according to uncertainty evaluation results to generate optimized segmentation results.

Description

Technical Field

[0001] The present invention relates to the technical field of laser point cloud processing, and more specifically, to an object labeling method and system based on laser point cloud. Background Art

[0002] As a 3D spatial data acquisition device, LiDAR is widely used in autonomous driving, drone navigation, robot perception and other fields. The point cloud data collected by LiDAR can accurately reflect the geometric shape and surface features of objects. Therefore, the object annotation method based on laser point cloud data has gradually become an important means of 3D environment perception and target recognition. However, due to the diversity and complexity of the surface materials of objects in the environment, the reflection characteristics of different materials have a great impact on the quality of point cloud data, especially when facing targets with non-Lambertian surfaces (such as glass, metal and other highly reflective materials), data distortion problems will occur.

[0003] In the prior art, object labeling methods based on laser point cloud data often cannot effectively handle mirror point and hole problems when faced with complex reflection phenomena caused by non-Lambertian surfaces, which will lead to inaccurate labeling results and affect the reliability of downstream applications. Summary of the invention

[0004] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides an object labeling method and system based on laser point cloud to solve the problems raised in the above-mentioned background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A method for object labeling based on laser point cloud comprises the following steps:

[0007] S1: Obtain the original point cloud data collected by the laser radar, and pre-process the original point cloud data to retain valid point cloud data;

[0008] S2: Based on the reflection intensity information and geometric topological features of the effective point cloud data, the non-Lambertian reflection area is identified and filtered and corrected to generate preliminary processed point cloud data;

[0009] S3: Perform multi-layer iterative correction on the initially processed point cloud data, correct the false mirror points in the non-Lambertian reflection area through deep filtering, and reconstruct the boundary based on the geometric topological characteristics to obtain the corrected point cloud data;

[0010] S4: A clustering algorithm based on global features and local features is used to group and classify the target objects in the corrected point cloud data to generate preliminary segmentation results;

[0011] S5: Use the attention network to finely divide the boundaries of the preliminary segmentation results, and monitor the local mutual information entropy output by the attention network to evaluate the uncertainty of the boundary area;

[0012] S6: Adjust the boundary division strategy according to the uncertainty assessment results of the boundary area to generate an optimized segmentation result.

[0013] In a preferred embodiment, the original point cloud data collected by the laser radar is obtained, and the original point cloud data is preprocessed to retain the valid point cloud data, specifically including:

[0014] The original point cloud data in the three-dimensional space is collected by the laser radar device, and the original point cloud data includes multiple three-dimensional coordinate points and their corresponding reflection intensity information;

[0015] De-noising the collected original point cloud data, and removing isolated points and low-density points in the original point cloud data through a denoising algorithm based on statistical analysis;

[0016] Calibrate the coordinate system of the denoised point cloud data to unify the spatial reference of the point cloud data by aligning the lidar coordinate system with the global reference coordinate system;

[0017] The point cloud data whose distance exceeds the effective range of the LiDAR is eliminated through a filtering algorithm based on a distance threshold;

[0018] The filtered point cloud data is density-smoothed, and the uniformity of the point cloud data is optimized using a spatial interpolation method to retain valid point cloud data for subsequent processing.

[0019] In a preferred embodiment, based on the reflection intensity information and geometric topological features of the effective point cloud data, the non-Lambertian reflection area is identified and a filtering correction process is performed to generate preliminary processed point cloud data, specifically including:

[0020] For each point in the valid point cloud data, the reflection intensity distribution of the points in its neighborhood is calculated, and the points with significantly abnormal reflection intensity are screened using the set intensity threshold, and these points are marked as possible non-Lambertian reflection points;

[0021] Perform geometric topological analysis on the areas marked as possible non-Lambertian reflection points, calculate the local curvature and distribution characteristics of the neighborhood points in the point cloud, and further identify areas with non-Lambertian reflection characteristics;

[0022] Applying filter correction processing to the area with non-Lambertian reflection characteristics, adjusting the points in the area with non-Lambertian reflection characteristics by a weighted mean filtering algorithm based on intensity distribution;

[0023] A filtering algorithm based on geometric consistency is further applied to remove points that deviate too much from the neighborhood geometric characteristics, and the point cloud data after filtering and correction is output as preliminary processed point cloud data.

[0024] In a preferred embodiment, the preliminarily processed point cloud data is subjected to multi-layer iterative correction, false mirror points in the non-Lambertian reflection area are corrected by deep filtering, and the boundary is reconstructed based on the geometric topological characteristics to obtain the corrected point cloud data, specifically including:

[0025] A neighborhood is established for each point in the preliminary processed point cloud data, and the depth value difference between the point and the neighborhood point is calculated, and the point with abnormal depth is marked as a possible false mirror point;

[0026] Apply a filtering algorithm based on depth value statistics to the areas marked as possible false mirror points, and remove points whose depth differences exceed the normal range;

[0027] For the filtered point cloud data, the continuity of the boundary area is analyzed based on the geometric topological characteristics, and the boundary is reconstructed by fitting the curvature change of the point cloud boundary;

[0028] In the process of boundary reconstruction, the points that deviate from the boundary characteristics are removed or reallocated by iteratively adjusting the positions of the points in the boundary area;

[0029] The point cloud data is gradually optimized through multi-layer iterative processing to generate corrected point cloud data.

[0030] In a preferred embodiment, a clustering algorithm based on a combination of global features and local features is used to group and classify target objects in the corrected point cloud data to generate a preliminary segmentation result, specifically including:

[0031] Extracting the global features of each point in the corrected point cloud data, the global features include the positional relationship of the points in the corrected point cloud data in the overall point cloud space and the overall distribution of the reflection intensity;

[0032] Extract local features of each point in the corrected point cloud data, including the distance distribution between the point in the corrected point cloud data and the points in its neighborhood, the consistency of the normal vector direction and the curvature information;

[0033] The feature vector of each point is calculated based on the combination of global features and local features. The feature vector is used to characterize the global position, reflection intensity and local geometric relationship of the point.

[0034] Apply a clustering algorithm to the feature vector of each point. The clustering algorithm uses a density-based clustering method to divide the points in the point cloud data into multiple groups according to the similarity of the feature vectors.

[0035] Based on the clustering grouping, the characteristics of each group are further analyzed through the classification algorithm, and the target objects are marked according to the point cloud morphological characteristics;

[0036] The clustering and classification results are combined to generate preliminary segmentation results of the point cloud data.

[0037] In a preferred embodiment, an attention mechanism network is used to finely divide the boundaries of the preliminary segmentation results, and the local mutual information entropy output by the attention mechanism network is monitored to evaluate the uncertainty of the boundary area, specifically including:

[0038] Extract the boundary points in the preliminary segmentation results based on the gradient characteristics of the normal vector direction change rate and geometric distribution in the neighborhood of each point in the point cloud data, and take the points with significant changes as boundary points;

[0039] Construct an input data set for the attention mechanism network. The input data set includes the three-dimensional spatial coordinates of the boundary points, the reflection intensity values, and the geometric topological features of the points in the neighborhood.

[0040] The input data is feature-weighted in the attention mechanism network to generate a feature importance score for each boundary point;

[0041] Update the feature vector of the boundary point according to the feature importance score to characterize the significance of the boundary point in the spatial distribution;

[0042] The local mutual information entropy corresponding to the boundary points is calculated by monitoring the output results of the attention mechanism network. The local mutual information entropy reflects the correlation between the geometric distribution and feature distribution of the boundary points in the neighborhood, and is used to evaluate the uncertainty of the boundary area.

[0043] In a preferred embodiment, the input data is feature-weighted in the attention mechanism network to generate a feature importance score for each boundary point, specifically:

[0044] Construct a deep network architecture based on a multi-head attention mechanism, which consists of an input layer, an attention layer, and an output layer;

[0045] The attention mechanism weights the features by calculating the weights of different features in the input data: ;in, Indicates The current input data instance and The attention weights between reference input data instances, Indicates The query vector of the current input data instance, Indicates a key vector that refers to an instance of the input data, Indicates The key vectors for all other input data instances, represents the index of all other input data instances, Represents the index of the current input data instance, Indicates the index of the reference input data instance.

[0046] In a preferred embodiment, adjusting the boundary division strategy according to the uncertainty assessment result of the boundary area to generate an optimized segmentation result specifically includes:

[0047] The local mutual information entropy of the boundary points calculated by the attention mechanism network is analyzed, and the boundary points are divided into high uncertainty areas and low uncertainty areas according to the local mutual information entropy;

[0048] For the boundary points in the high uncertainty area, the position of the boundary points is adjusted in combination with the local geometric topological characteristics, and the boundary is corrected by recalculating the geometric consistency between the boundary points and other points in the neighborhood, while the existing division of the boundary points in the low uncertainty area is maintained;

[0049] Conduct continuity analysis on the distribution of boundary points in the corrected high uncertainty region to remove isolated and redundant points;

[0050] Through the dynamically updated partitioning strategy, the adjusted boundary points of the high uncertainty region are merged with the boundary points of the low uncertainty region to generate a unified boundary partitioning result;

[0051] The merged boundary division results are segmented and optimized to generate optimized segmentation results for target annotation using point cloud data.

[0052] On the other hand, the present invention provides an object annotation system based on laser point cloud, comprising a point cloud data processing module, a reflection area correction module, a multi-layer iterative correction module, a global local feature module, a boundary fine division module and a boundary division optimization module;

[0053] Point cloud data processing module: obtains the original point cloud data collected by the laser radar, and pre-processes the original point cloud data to retain valid point cloud data;

[0054] Reflection area correction module: Based on the reflection intensity information and geometric topological features of the effective point cloud data, it identifies the non-Lambertian reflection area and performs filtering correction processing to generate preliminary processed point cloud data;

[0055] Multi-layer iterative correction module: Perform multi-layer iterative correction on the initially processed point cloud data, correct the false mirror points in the non-Lambertian reflection area through deep filtering, and reconstruct the boundary based on the geometric topological characteristics to obtain the corrected point cloud data;

[0056] Global and local feature module: Group and classify the target objects in the corrected point cloud data based on a clustering algorithm that combines global features and local features to generate preliminary segmentation results;

[0057] Boundary fine division module: Use the attention mechanism network to finely divide the boundaries of the preliminary segmentation results, and monitor the local mutual information entropy output by the attention mechanism network to evaluate the uncertainty of the boundary area;

[0058] Boundary division optimization module: adjusts the boundary division strategy according to the uncertainty assessment results of the boundary area and generates optimized segmentation results.

[0059] Technical effects and advantages of the object labeling method and system based on laser point cloud of the present invention:

[0060] 1. By introducing a filtering correction step based on reflection intensity information and geometric topological features, the present invention can accurately identify non-Lambertian reflection areas, and correct false mirror points and void data through multi-layer iterative correction, thereby generating high-quality corrected point cloud data. This method not only enhances the adaptability to complex reflection phenomena, but also improves the integrity and reliability of point cloud data, ensuring more accurate annotation results, and laying a solid foundation for downstream applications.

[0061] 2. The target objects are grouped and classified by combining the clustering algorithm of global features and local features, and the attention mechanism network is further introduced to finely divide the segmentation boundary. At the same time, the uncertainty of the boundary area is quantitatively evaluated by dynamically monitoring the local mutual information entropy, making the boundary division more accurate. The dynamic segmentation strategy optimization based on the uncertainty assessment results can effectively avoid the annotation deviation caused by the boundary division error, thereby generating an optimized segmentation result. While improving the annotation accuracy, this method has high versatility and scalability, and is suitable for the laser point cloud processing needs in a variety of complex scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 A schematic diagram of an object labeling method based on laser point cloud according to the present invention;

[0063] Figure 2 It is a structural schematic diagram of an object labeling system based on laser point cloud of the present invention. DETAILED DESCRIPTION

[0064] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0065] Embodiment 1: Figure 1 The present invention provides an object labeling method based on laser point cloud, which comprises the following steps:

[0066] S1: Obtain the original point cloud data collected by the lidar, and preprocess the original point cloud data to retain valid point cloud data.

[0067] S2: Based on the reflection intensity information and geometric topological features of the valid point cloud data, the non-Lambertian reflection area is identified and filtered and corrected to generate preliminary processed point cloud data.

[0068] S3: Perform multi-layer iterative correction on the initially processed point cloud data, correct the false mirror points in the non-Lambertian reflection area through deep filtering, and reconstruct the boundary based on the geometric topological characteristics to obtain the corrected point cloud data.

[0069] S4: A clustering algorithm based on global and local features is used to group and classify the target objects in the corrected point cloud data to generate preliminary segmentation results.

[0070] S5: Use the attention mechanism network to finely divide the boundaries of the preliminary segmentation results, and monitor the local mutual information entropy output by the attention mechanism network to evaluate the uncertainty of the boundary area.

[0071] S6: Adjust the boundary division strategy according to the uncertainty assessment results of the boundary area to generate an optimized segmentation result.

[0072] Obtain the original point cloud data collected by the LiDAR and preprocess the original point cloud data to retain valid point cloud data, including:

[0073] The original point cloud data in the three-dimensional space is collected by the laser radar device. The original point cloud data includes multiple three-dimensional coordinate points and their corresponding reflection intensity information:

[0074] During the scanning process, the laser radar emits laser pulses at a fixed angular resolution and time interval, and calculates the three-dimensional spatial coordinates and reflection intensity value of each laser point based on the received echo signal.

[0075] The three-dimensional coordinates of each point are calculated based on the horizontal angle, vertical angle and target distance measured by the lidar equipment; the reflection intensity value indicates the target surface's ability to reflect laser signals, which is used for subsequent feature analysis and point cloud data processing.

[0076] The collected original point cloud data is denoised, and isolated points and low-density points in the original point cloud data are removed by a denoising algorithm based on statistical analysis:

[0077] Calculate the number of neighborhood points for each point, determine whether the point is an isolated point based on whether the number of neighborhood points is lower than the set threshold, and remove isolated points to improve the overall accuracy of the point cloud data; set a density lower limit based on the point density distribution in the local area of ​​the point cloud, and remove points with point density lower than the threshold to optimize the consistency of the point cloud distribution.

[0078] Calibrate the coordinate system of the denoised point cloud data to unify the spatial reference of the point cloud data by aligning the lidar coordinate system with the global reference coordinate system:

[0079] The coordinate system of the original point cloud data is usually based on the internal coordinate system of the lidar device; the fixed reference coordinate system required for the application scenario (such as the geographic coordinate system or the vehicle coordinate system); by calculating the pose transformation matrix of the lidar device, the point cloud data is converted from the lidar coordinate system to the global reference coordinate system, so that the spatial position of the point cloud data conforms to the actual scene.

[0080] The point cloud data whose distance exceeds the effective range of the LiDAR is removed through a filtering algorithm based on a distance threshold:

[0081] For the denoised and calibrated point cloud data, a filtering algorithm based on distance threshold is applied to eliminate points whose distance exceeds the effective detection range of the lidar.

[0082] The laser radar device has a certain detection range, including a minimum distance and a maximum distance. For point cloud data with a distance less than the minimum detection range or greater than the maximum detection range, these points are directly eliminated by setting a distance threshold to avoid interference of invalid points in subsequent processing.

[0083] The point cloud data retained after filtering covers the effective area, ensuring that the data used for subsequent processing has practical significance.

[0084] The filtered point cloud data is density-smoothed and the uniformity of the point cloud data is optimized using a spatial interpolation method to retain valid point cloud data for subsequent processing:

[0085] The neighborhood density of each point is calculated. If the point density in a certain area is too low, additional points are generated through interpolation. Based on the three-dimensional spatial coordinates of the points, points are inserted into the low-density areas in the point cloud data to improve the overall spatial uniformity of the point cloud. After density smoothing, the distribution of the point cloud data is more uniform, which is convenient for subsequent algorithms to perform feature extraction and segmentation operations.

[0086] The point cloud data generated after the above steps are processed is effective point cloud data, which has the characteristics of denoising, coordinate calibration, distance filtering and density smoothing optimization.

[0087] Based on the reflection intensity information and geometric topological features of the effective point cloud data, the non-Lambertian reflection area is identified and filtered and corrected to generate preliminary processed point cloud data, including:

[0088] For each point in the valid point cloud data, the reflection intensity distribution of the points in its neighborhood is calculated, and the points with significantly abnormal reflection intensity are screened using the set intensity threshold, and these points are marked as possible non-Lambertian reflection points:

[0089] For each point in the valid point cloud data, a spherical neighborhood with a fixed radius is defined as the center. The reflection intensity values ​​of all points contained in the neighborhood will be used to calculate the reflection intensity distribution characteristics of the center point.

[0090] Assume that the neighborhood of the center point contains points, and the reflection intensity values ​​of these points are The calculation formula for the mean reflection intensity is as follows: ;in, represents the mean reflection intensity, Indicates the neighborhood The reflection intensity value of each point, Indicates the number of the point in the neighborhood.

[0091] Based on the mean reflection intensity of points in the neighborhood, calculate the variance of the reflection intensity: ;in, Represents the variance of the reflection intensity of points in the neighborhood.

[0092] Compare the reflection intensity value of the neighborhood center point with the reflection intensity mean and variance. If the following conditions are met, the point is marked as a possible non-Lambertian reflection point: ;in, Indicates the reflection intensity value of the center point; It represents the coefficient for judging abnormal reflection intensity, and its value is set based on experience (such as 1.5 to 2.5).

[0093] Perform geometric topological analysis on the areas marked as possible non-Lambertian reflection points, calculate the local curvature and distribution characteristics of the neighborhood points in the point cloud, and further identify areas with non-Lambertian reflection characteristics:

[0094] For the point cloud areas marked as possible non-Lambertian reflection points, the point cloud distribution characteristics in its neighborhood are analyzed with each marked point as the center, and the local curvature of the area is calculated. The local curvature indicates the degree of surface curvature of the local area of ​​the point cloud, and is used to identify whether there are abnormal geometric characteristics on the surface of the area.

[0095] In the calculation, a spherical neighborhood with a fixed radius is defined, and the three-dimensional spatial offset values ​​of all points in the neighborhood relative to the center point are counted. The offset value can reflect the degree of change in the spatial position of the points in the neighborhood and the center point, and the average offset of the entire neighborhood is obtained by the mean or weighted calculation. The more uniform the offset of the points in the neighborhood, the smaller the local curvature and the smoother the surface; conversely, when the offset difference is large, the local curvature is higher, indicating that there is a significant geometric change in the surface.

[0096] Calculate the distribution characteristics of the neighborhood points, including the distribution density and directional characteristics of the points. If the distribution density of the neighborhood points is significantly abnormal and the directional characteristics deviate from the normal range, it is further determined that the area has non-Lambertian reflection characteristics.

[0097] Apply filtering correction processing to areas with non-Lambertian reflection characteristics, and adjust the points in the area with non-Lambertian reflection characteristics through a weighted mean filtering algorithm based on intensity distribution to reduce the impact of reflection anomalies:

[0098] In the non-Lambertian reflection area, the reflection intensity value of the point is adjusted to make it closer to the average value of the neighborhood. The calculation formula for the adjusted reflection intensity value is as follows: ;in, Indicates the corrected center point reflection intensity value; Represents the weight coefficient of the points in the neighborhood, which is usually set to the inverse of the distance from the neighborhood point to the center point.

[0099] The filtering algorithm based on geometric consistency is further applied to remove points that deviate too much from the neighborhood geometric characteristics, and the point cloud data after filtering and correction is output as preliminary processed point cloud data:

[0100] Points that deviate too much from the geometric characteristics of the neighborhood are further eliminated. If the deviation of the average geometric distance of a point from other points in its neighborhood exceeds a preset threshold, the point is directly removed.

[0101] The setting of the preset threshold corresponding to the deviation of the average geometric distance between a point and other points in its neighborhood can be adjusted according to the specific application scenario and the density characteristics of the point cloud data. Usually, it is taken as a fixed multiple of the average geometric distance in the neighborhood (such as 1.5 to 3 times) to tolerate a small amount of normal deviation while effectively eliminating abnormal points with large deviations.

[0102] The point cloud data after filtering and correction is saved as preliminary processed point cloud data and output for subsequent processing.

[0103] The preliminary processed point cloud data is corrected by multi-layer iterative correction, the false mirror points in the non-Lambertian reflection area are corrected by deep filtering, and the boundary is reconstructed based on the geometric topological characteristics to obtain the corrected point cloud data, including:

[0104] A neighborhood is established for each point in the preliminary processed point cloud data. By calculating the depth difference between the point and the neighborhood points, the points with abnormal depth are marked as possible false mirror points:

[0105] Taking the three-dimensional coordinates of each point in the point cloud as the center, a spherical neighborhood radius is set to filter out all neighborhood points within the radius. The size of the neighborhood can be dynamically adjusted according to the distribution density of the point cloud data to ensure that there are enough points in the neighborhood for statistical analysis.

[0106] The difference between the depth value of each point in the neighborhood and the depth value of the center point is calculated to identify depth anomalies. The depth difference is calculated based on the original depth value of the point cloud. The depth value refers to the distance of each point relative to the laser radar when the laser radar device collects data.

[0107] By setting the depth difference threshold, we can filter out the neighborhood points whose depth values ​​deviate significantly from the center point, and mark these points as possible false mirror points. The depth difference threshold is usually set based on experimental data and can be selected in the range of 1.5 to 3 times the standard deviation of the depth difference. The marked false mirror point area will be the focus of subsequent processing.

[0108] Apply a filtering algorithm based on depth value statistics to the areas marked as possible false mirror points, removing points with depth differences exceeding the normal range:

[0109] Count the depth value distribution characteristics of the area where the possible false mirror points are located, and calculate the mean and variance of the depth values ​​of the points in the area;

[0110] The upper and lower limits of the depth value distribution are set to the depth value mean plus or minus several times the variance. The specific multiple is determined according to the distribution density of the point cloud data, and the value range is usually 1.5 to 3 times. Points with depth values ​​outside the upper and lower limits are directly judged as abnormal points and removed from the point cloud data. For points within the critical range, further judgment is made based on their geometric consistency with the distribution of neighboring points, and points that meet the geometric consistency are retained.

[0111] For the filtered point cloud data, the continuity of the boundary area is analyzed based on the geometric topological characteristics, and the boundary is reconstructed by fitting the curvature change of the point cloud boundary:

[0112] After completing the statistical filtering of the depth values, the geometric topological characteristics of the remaining point cloud data are further analyzed, focusing on identifying the continuity of the boundary area.

[0113] Based on the points in the boundary area, local geometric topological features such as curvature and normal vector distribution are calculated. Curvature is used to measure the curvature of the boundary surface, while normal vector distribution reflects the directional consistency of the point cloud boundary.

[0114] By analyzing the spatial distribution of points in the boundary area, it is determined whether there are gaps or irregular areas. If a point does not match the distribution characteristics of the points in its neighborhood, the point is marked as a boundary discontinuity point to provide a basis for subsequent reconstruction.

[0115] During the boundary reconstruction process, the points that deviate from the boundary characteristics are removed or reallocated by iteratively adjusting the positions of the points within the boundary area:

[0116] The curve fitting algorithm based on the least square method is used to fit the continuous boundary curve according to the curvature values ​​of the points in the boundary area. During the fitting process, the abnormal curvature values ​​are smoothed to ensure the continuity and smoothness of the boundary.

[0117] For the boundary points that deviate from the fitting curve, their positions are iteratively adjusted to redistribute them to the fitting boundary, ensuring that the geometric characteristics of the point cloud boundary are more consistent.

[0118] The point cloud data is gradually optimized through multi-layer iterative processing to generate corrected point cloud data:

[0119] After completing a boundary reconstruction, the point cloud data is optimized through multiple layers of iterations. Each iteration reanalyzes the depth value distribution and geometric characteristics of the point cloud data to gradually optimize the removal of false points and the reconstruction of the boundary.

[0120] The iteration stop condition is set. When the change in the number of false points between two iterations is less than the set threshold, or the boundary reconstruction curvature fitting error is less than the set range, the iteration is stopped. In each iteration, the steps of depth filtering and boundary reconstruction are repeated, and the statistical information of the point cloud data after each iteration is recorded, such as the change in the number of points, the number of boundary point adjustments, etc.

[0121] After completing the multi-layer iterative processing, the optimized point cloud data is saved as the corrected point cloud data. The corrected point cloud data meets the optimization standards in terms of depth value and geometric topological characteristics, has higher accuracy and integrity, and can be used for subsequent grouping and classification of target objects.

[0122] The clustering algorithm based on the combination of global features and local features is used to group and classify the target objects in the corrected point cloud data to generate preliminary segmentation results, including:

[0123] The global features are extracted for each point in the corrected point cloud data. The global features include the positional relationship of the points in the corrected point cloud data in the overall point cloud space and the overall distribution of the reflection intensity:

[0124] For each point in the corrected point cloud data, its 3D spatial coordinates are obtained to represent the spatial position of the point relative to the entire point cloud data. By normalizing the coordinates of all points, the spatial position of the point is mapped to the unit cube range to ensure that point cloud data of different sizes are calculated at a unified scale.

[0125] Calculate the global distribution characteristics of the reflection intensity in the corrected point cloud data. The specific method is to count the reflection intensity values ​​of all points and calculate their mean and standard deviation. By normalizing the reflection intensity of each point to zero mean and unit variance, the distribution characteristics of the reflection intensity can be standardized in subsequent calculations.

[0126] Extract local features from each point in the corrected point cloud data. Local features include the distance distribution between the point in the corrected point cloud data and the points in its neighborhood, the consistency of the normal vector direction, and the curvature information:

[0127] With each point as the center, a spherical neighborhood with a fixed radius is defined. The radius is dynamically adjusted according to the density of the point cloud data to ensure that the number of points in each neighborhood meets the statistical requirements. For dense point cloud data, the neighborhood radius is small; for sparse point cloud data, the neighborhood radius is appropriately enlarged.

[0128] In the neighborhood of each point, the distance distribution characteristics from the neighborhood point to the center point are calculated. The distance distribution characteristics include the average distance from the points in the neighborhood to the center point, the standard deviation, and the maximum and minimum values ​​of the distance distribution. These statistics reflect the local distribution characteristics of the points.

[0129] Based on the geometric distribution of the neighborhood points, the normal vector direction of each point is calculated, and the consistency of the normal vector direction of the points in the neighborhood is analyzed. If the direction consistency of the normal vector in the neighborhood is high, it means that the surface of the area where the point is located is relatively flat; if the direction consistency is low, it means that the point may be located at the edge or in an area with drastic surface changes.

[0130] The local curvature of each point is calculated. The curvature is used to describe the curvature of the surface of the area where the point is located. By analyzing the geometric relationship between the neighborhood points and the center point, the mean and standard deviation of the local curvature are calculated to characterize the curvature characteristics of the area.

[0131] The feature vector of each point is calculated based on the combination of global features and local features. The feature vector is used to characterize the global position, reflection intensity and local geometric relationship of the point:

[0132] The feature vector of each point includes the following information: the three-dimensional coordinates of the point (global position); the normalized reflection intensity value of the point; the average distance and standard deviation of the points in the neighborhood; the consistency of the normal vector direction of the neighborhood points; and the mean and standard deviation of the curvature of the neighborhood points.

[0133] The features of each dimension in the feature vector are standardized to ensure that features of different dimensions have the same weight in subsequent clustering calculations.

[0134] A clustering algorithm is applied to the feature vector of each point. The clustering algorithm uses a density-based clustering method to divide the points in the point cloud data into multiple groups according to the similarity of the feature vectors:

[0135] For each point, the density value of the point in the feature space is calculated based on the similarity of its feature vector with other points. The similarity is calculated using Euclidean distance or cosine similarity.

[0136] The density-based clustering algorithm assigns points with high density similarity to the same group. For points with lower density, the distance between them and other high-density areas is further judged to decide whether to assign them to the nearest group.

[0137] Based on the clustering grouping, the characteristics of each group are further analyzed through the classification algorithm, and the target objects are marked according to the point cloud morphological characteristics:

[0138] The morphological features of each group of point cloud data, such as size, boundary characteristics, surface flatness, etc., are extracted for classification and judgment.

[0139] The features of the groups are matched with the preset category model, and a category label is assigned to each group. For example, point cloud groups are labeled as "vehicles", "pedestrians", "trees" and other categories based on size and morphological features.

[0140] Combine the clustering and classification results to generate preliminary segmentation results for the point cloud data:

[0141] The clustering and classification results are combined to generate the preliminary segmentation results of the point cloud data. The segmentation results include the group number and category label of each point, and keep the corresponding relationship with the corrected point cloud data.

[0142] The attention mechanism network is used to finely divide the boundaries of the preliminary segmentation results, and the local mutual information entropy output by the attention mechanism network is monitored to evaluate the uncertainty of the boundary area, including:

[0143] The boundary points in the preliminary segmentation results are extracted based on the gradient characteristics of the normal vector direction change rate and geometric distribution in the neighborhood of each point in the point cloud data, and the points with significant changes are taken as boundary points:

[0144] For each point in the preliminary segmentation result, a spherical neighborhood is defined with the point as the center. The radius of the neighborhood is dynamically adjusted according to the density and distribution of the point cloud data to ensure that the neighborhood contains a sufficient number of points for subsequent analysis.

[0145] In the defined neighborhood, the normal vector direction of each point is calculated. The normal vector direction is determined by surface fitting of the point cloud data, and its direction is consistent with the surface normal direction. For the center point, the angle change between its normal vector and the normal vector of the neighboring points is analyzed, and points with significant angle changes are considered to be likely to be located in the boundary area.

[0146] Based on the spatial distribution of points in the neighborhood, the geometric distance gradient between each point and its neighboring points is calculated. The larger the gradient value, the more dramatic the change in the geometric characteristics of the area where the point is located. Combining the normal vector direction change rate and the geometric gradient characteristics, the points with significant changes are extracted as boundary points.

[0147] Construct the input data set of the attention mechanism network. The input data set includes the three-dimensional spatial coordinates of the boundary points, the reflection intensity values, and the geometric topological features of the points in the neighborhood:

[0148] For each extracted boundary point, an input data set is constructed containing the following: three-dimensional spatial coordinates: recording the three-dimensional position of the boundary point, used to describe the spatial distribution of the point in the point cloud data; reflection intensity value: the original reflection intensity information of the boundary point, used to characterize the material and optical properties of the point; geometric topological features of points in the neighborhood: including the distance distribution between the neighborhood point and the center point, the consistency of the normal vector, and the local curvature value, used to describe the geometric characteristics of the boundary point.

[0149] The above input data is normalized to eliminate the influence of dimension. For example, the three-dimensional space coordinates are mapped to the range of [0,1] and the reflection intensity values ​​are standardized to the form of zero mean and unit variance.

[0150] The input data is feature-weighted in the attention mechanism network to generate feature importance scores for each boundary point:

[0151] A deep network architecture based on the multi-head attention mechanism is constructed. The network consists of the following parts: input layer: receives the constructed input data set, and the data of each boundary point is taken as an input instance; attention layer: the multi-head attention module extracts the key features of the boundary points by calculating the weighted coefficients of different features; output layer: generates the feature importance score of the boundary point, which is used to describe the significance of the point in the refined boundary division.

[0152] The attention mechanism weights the features by calculating the weights of different features in the input data. The weights are calculated as: ;in, Indicates The current input data instance and The attention weights between the reference input data instances are used to measure the The reference input data instance is The importance of the current input data instance; Indicates The query vector of the current input data instance comes from the feature map of the input data and is used to measure the degree of attention paid by the current instance to the features of other instances; Indicates The key vectors of the reference input data instances come from the feature map of the input data and are used to characterize the characteristic attributes of other instances to support attention calculation; Indicates The key vectors for all other input data instances; Represents the index of all other input data instances, that is, all instances other than the current input data instance and the reference input data instance; Represents the index of the current input data instance, which refers to the target point where the attention weight needs to be calculated; Represents the index of the reference input data instance, referring to the current target point Other points (or instances) used to calculate similarity; Indicates The query vector of the current input data instance is The dot product operation of the key vector of the reference input data instance is used to measure the similarity between the two.

[0153] The attention weights are used to perform weighted summation on the features of each dimension of the input data to generate the feature importance score of each boundary point for subsequent boundary division optimization.

[0154] Update the feature vector of the boundary point according to the feature importance score to characterize the significance of the boundary point in the spatial distribution:

[0155] Combined with the feature importance score, the feature vector of the boundary point is updated. The updated feature vector can more accurately represent the spatial distribution and geometric characteristics of the boundary point.

[0156] The updated feature vector includes the following parts: the three-dimensional coordinates and reflection intensity of the point; the distance distribution statistics (mean and standard deviation) of the points in the neighborhood; the consistency statistics (variance) of the neighborhood normal vector; and the curvature value of the boundary point.

[0157] The local mutual information entropy corresponding to the boundary points is calculated by monitoring the output results of the attention mechanism network. The local mutual information entropy reflects the correlation between the geometric distribution and feature distribution of the boundary points in the neighborhood, and is used to evaluate the uncertainty of the boundary area:

[0158] The local mutual information entropy is used to describe the correlation between the feature distribution and geometric distribution of the boundary points, reflecting the feature uncertainty of the boundary area. The calculation formula of the local mutual information entropy is: ;in, It is the local mutual information entropy, which indicates the degree of correlation between the feature distribution and geometric distribution of the current point in the neighborhood, and is used to quantify the uncertainty of point cloud data; Represents all possible eigenvalues ​​in the neighborhood of the current point Carry out statistics and calculations one by one; is the set of all eigenvalues ​​in the neighborhood of the current point, including the value of each feature in the eigenvector dimension of all points; Represents a single feature value of a point in the neighborhood; feature value The specific type depends on the characteristic dimensions of the current input data, such as the reflection intensity value of the point, the geometric distance of the points in the neighborhood, the consistency of the direction of the neighborhood normal vector, and the curvature value; Indicates the eigenvalue in the neighborhood The probability distribution of , that is, the frequency of the eigenvalue in the neighborhood.

[0159] When the local mutual information entropy is high, it means that there is a large uncertainty or unevenness in the feature distribution within the neighborhood; when the local mutual information entropy is low, it means that the feature distribution within the neighborhood is relatively stable and consistent.

[0160] By counting the feature distribution of points in the neighborhood, the feature probability density function is estimated. The local mutual information entropy of the boundary points is calculated using the entropy formula. The larger the value, the higher the uncertainty.

[0161] Adjust the boundary division strategy based on the uncertainty assessment results of the boundary area to generate optimized segmentation results, including:

[0162] The local mutual information entropy of the boundary points calculated by the attention mechanism network is analyzed, and the boundary points are divided into high uncertainty areas and low uncertainty areas according to the local mutual information entropy:

[0163] The local mutual information entropy is used to measure the degree of feature uncertainty of boundary points. The larger its value is, the higher the uncertainty of the boundary points in the geometric distribution or feature distribution is.

[0164] According to the value of local mutual information entropy, the boundary points are divided into high uncertainty areas and low uncertainty areas:

[0165] High uncertainty area: contains boundary points whose local mutual information entropy values ​​exceed the set threshold. The boundary division of these areas needs to be further optimized. Low uncertainty area: contains boundary points whose local mutual information entropy values ​​are lower than the set threshold. The boundary division of these areas is relatively stable.

[0166] The threshold of the local mutual information entropy is dynamically adjusted according to the distribution characteristics of the point cloud data. It is usually 1.5 to 2 times the mean value of the local mutual information entropy of all boundary points to accurately divide the area.

[0167] For the boundary points in the high uncertainty area, the position of the boundary points is adjusted in combination with the local geometric topological characteristics, and the boundary is corrected by recalculating the geometric consistency between the boundary points and other points in the neighborhood. The existing division of the boundary points in the low uncertainty area is maintained:

[0168] For the boundary points in the high uncertainty area, the geometric distance between the boundary points and the points in the neighborhood is calculated, and the distribution density of the points is analyzed; by fitting the local curvature, the position of the boundary points is adjusted to make it conform to the overall characteristics of the geometric distribution in the neighborhood.

[0169] For the boundary points in the low uncertainty area, the existing division results are directly maintained to avoid distortion or redundancy of the boundary points due to repeated adjustments.

[0170] Perform continuity analysis on the distribution of boundary points in the corrected high uncertainty region and remove isolated and redundant points:

[0171] Continuity analysis is performed in the following ways: checking the connectivity between boundary points and neighboring points; calculating the distance between boundary points and their nearest neighbor boundary points, and determining whether there are isolated points or overly dense redundant points.

[0172] For isolated points (points whose distance to the nearest neighbor exceeds the set threshold), they are directly removed from the boundary point set; for redundant points (points that are too close to neighboring points and have no significant difference in geometric characteristics), they are selectively removed to optimize the uniformity of the boundary point distribution.

[0173] Through the dynamically updated partitioning strategy, the adjusted boundary points of the high uncertainty area are merged with the boundary points of the low uncertainty area to generate a unified boundary partitioning result:

[0174] According to the correction results of the boundary points in the high uncertainty area, the boundary division strategy is dynamically adjusted. The division strategy includes: updating the boundary point set, removing the isolated points and redundant points from the boundary point set; recording the corrected positions to generate the final unified division result.

[0175] The corrected boundary points of the high uncertainty region are merged with the boundary points of the low uncertainty region to generate a unified boundary division result. During the merging process, the spatial continuity and geometric consistency of the boundary points are ensured.

[0176] Perform segmentation optimization on the merged boundary segmentation results to generate optimized segmentation results for target annotation with point cloud data:

[0177] Smooth boundary lines to eliminate excessive changes between boundary points; adjust the final position of boundary points based on local geometric characteristics to ensure that the boundary points are consistent with the overall geometric distribution of the point cloud.

[0178] Based on the optimized boundary division results, the point cloud data is segmented to generate optimized segmentation results. The segmentation results include the segmentation category and boundary point marking of each point to ensure the integrity and accuracy of the point cloud segmentation.

[0179] Embodiment 2: The difference between Embodiment 2 of the present invention and Embodiment 1 is that this embodiment introduces an object labeling system based on laser point cloud.

[0180] Figure 2 A structural schematic diagram of an object labeling system based on laser point cloud of the present invention is given, which includes a point cloud data processing module, a reflection area correction module, a multi-layer iterative correction module, a global-local feature module, a boundary fine division module and a boundary division optimization module.

[0181] Point cloud data processing module: obtains the original point cloud data collected by the lidar and preprocesses the original point cloud data to retain valid point cloud data.

[0182] Reflection area correction module: Based on the reflection intensity information and geometric topological features of the effective point cloud data, the non-Lambertian reflection area is identified and filtered and corrected to generate preliminary processed point cloud data.

[0183] Multi-layer iterative correction module: Perform multi-layer iterative correction on the initially processed point cloud data, correct the false mirror points in the non-Lambertian reflection area through deep filtering, and reconstruct the boundaries based on the geometric topological characteristics to obtain the corrected point cloud data.

[0184] Global and local feature module: A clustering algorithm based on the combination of global and local features is used to group and classify the target objects in the corrected point cloud data to generate preliminary segmentation results.

[0185] Boundary fine division module: The attention mechanism network is used to finely divide the boundaries of the preliminary segmentation results, and the local mutual information entropy output by the attention mechanism network is monitored to evaluate the uncertainty of the boundary area.

[0186] Boundary division optimization module: adjusts the boundary division strategy according to the uncertainty assessment results of the boundary area and generates optimized segmentation results.

[0187] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters and thresholds in the formula are set by technicians in this field according to actual conditions.

[0188] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination thereof. When implemented by software, the above embodiments may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium, or may be transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions may be transmitted from one website, computer, server or data center to another website, computer, server or data center by wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can access or a data storage device such as a server or data center that contains one or more available media sets. The available medium may be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a DVD), or a semiconductor medium. The semiconductor medium may be a solid-state hard disk.

[0189] Those of ordinary skill in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0190] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and modules described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0191] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the modules is only a logical function division. There may be other division methods in actual implementation, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0192] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical modules, and may be located in one place or distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0193] In addition, each functional module in each embodiment of the present application may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0194] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program codes.

[0195] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

[0196] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for object labeling based on laser point cloud, characterized in that: The steps include: S1: Obtain the original point cloud data collected by the laser radar, and pre-process the original point cloud data to retain valid point cloud data; S2: Based on the reflection intensity information and geometric topological features of the valid point cloud data, identify the non-Lambertian reflection area and perform filtering and correction processing to generate preliminary processed point cloud data, including: For each point in the valid point cloud data, the reflection intensity distribution of the points in its neighborhood is calculated, and the points with significantly abnormal reflection intensity are screened using the set intensity threshold, and these points are marked as possible non-Lambertian reflection points; Perform geometric topological analysis on the areas marked as possible non-Lambertian reflection points, calculate the local curvature and distribution characteristics of the neighborhood points in the point cloud, and further identify areas with non-Lambertian reflection characteristics; Applying filter correction processing to the area with non-Lambertian reflection characteristics, adjusting the points in the area with non-Lambertian reflection characteristics by a weighted mean filtering algorithm based on intensity distribution; The filtering algorithm based on geometric consistency is further applied to remove points that deviate too much from the geometric characteristics of the neighborhood, and the point cloud data after filtering and correction is output as preliminary processed point cloud data; S3: Perform multi-layer iterative correction on the initially processed point cloud data, correct the false mirror points in the non-Lambertian reflection area through deep filtering, and reconstruct the boundary based on the geometric topological characteristics to obtain the corrected point cloud data; S4: A clustering algorithm based on global and local features is used to group and classify the target objects in the corrected point cloud data to generate preliminary segmentation results, including: Extracting the global features of each point in the corrected point cloud data, the global features include the positional relationship of the points in the corrected point cloud data in the overall point cloud space and the overall distribution of the reflection intensity; Extract local features of each point in the corrected point cloud data, including the distance distribution between the point in the corrected point cloud data and the points in its neighborhood, the consistency of the normal vector direction and the curvature information; The feature vector of each point is calculated based on the combination of global features and local features. The feature vector is used to characterize the global position, reflection intensity and local geometric relationship of the point. Apply a clustering algorithm to the feature vector of each point. The clustering algorithm uses a density-based clustering method to divide the points in the point cloud data into multiple groups according to the similarity of the feature vectors. Based on the clustering grouping, the characteristics of each group are further analyzed through the classification algorithm, and the target objects are marked according to the point cloud morphological characteristics; Combine the clustering and classification results to generate preliminary segmentation results of the point cloud data; S5: Use the attention network to finely divide the boundaries of the preliminary segmentation results, and monitor the local mutual information entropy output by the attention network to evaluate the uncertainty of the boundary area; S6: Adjust the boundary division strategy according to the uncertainty assessment results of the boundary area to generate an optimized segmentation result.

2. The object labeling method based on laser point cloud according to claim 1, characterized in that: Obtain the original point cloud data collected by the LiDAR and preprocess the original point cloud data to retain valid point cloud data, including: The original point cloud data in the three-dimensional space is collected by the laser radar device, and the original point cloud data includes multiple three-dimensional coordinate points and their corresponding reflection intensity information; De-noising the collected original point cloud data, and removing isolated points and low-density points in the original point cloud data through a denoising algorithm based on statistical analysis; Calibrate the coordinate system of the denoised point cloud data to unify the spatial reference of the point cloud data by aligning the lidar coordinate system with the global reference coordinate system; The point cloud data whose distance exceeds the effective range of the LiDAR is eliminated through a filtering algorithm based on a distance threshold; The filtered point cloud data is density-smoothed, and the uniformity of the point cloud data is optimized using a spatial interpolation method to retain valid point cloud data for subsequent processing.

3. The object labeling method based on laser point cloud according to claim 1, characterized in that: The preliminary processed point cloud data is corrected by multi-layer iterative correction, the false mirror points in the non-Lambertian reflection area are corrected by deep filtering, and the boundary is reconstructed based on the geometric topological characteristics to obtain the corrected point cloud data, including: A neighborhood is established for each point in the preliminary processed point cloud data, and the depth value difference between the point and the neighborhood point is calculated, and the point with abnormal depth is marked as a possible false mirror point; Apply a filtering algorithm based on depth value statistics to the areas marked as possible false mirror points, and remove points whose depth differences exceed the normal range; For the filtered point cloud data, the continuity of the boundary area is analyzed based on the geometric topological characteristics, and the boundary is reconstructed by fitting the curvature change of the point cloud boundary; In the process of boundary reconstruction, the points that deviate from the boundary characteristics are removed or reallocated by iteratively adjusting the positions of the points in the boundary area; The point cloud data is gradually optimized through multi-layer iterative processing to generate corrected point cloud data.

4. The object labeling method based on laser point cloud according to claim 1, characterized in that: The attention mechanism network is used to finely divide the boundaries of the preliminary segmentation results, and the local mutual information entropy output by the attention mechanism network is monitored to evaluate the uncertainty of the boundary area, including: The boundary points in the preliminary segmentation results are extracted based on the gradient characteristics of the normal vector direction change rate and geometric distribution in the neighborhood of each point in the point cloud data, and the points with significant changes are taken as boundary points; Construct an input data set for the attention mechanism network. The input data set includes the three-dimensional spatial coordinates of the boundary points, the reflection intensity values, and the geometric topological features of the points in the neighborhood. The input data is feature-weighted in the attention mechanism network to generate a feature importance score for each boundary point; Update the feature vector of the boundary point according to the feature importance score to characterize the significance of the boundary point in the spatial distribution; The local mutual information entropy corresponding to the boundary points is calculated by monitoring the output results of the attention mechanism network. The local mutual information entropy reflects the correlation between the geometric distribution and feature distribution of the boundary points in the neighborhood, and is used to evaluate the uncertainty of the boundary area.

5. The object labeling method based on laser point cloud according to claim 4, characterized in that: The input data is feature-weighted in the attention mechanism network to generate the feature importance score of each boundary point, specifically: Construct a deep network architecture based on a multi-head attention mechanism, which consists of an input layer, an attention layer, and an output layer; The attention mechanism weights the features by calculating the weights of different features in the input data: ;in, Indicates The current input data instance and The attention weights between reference input data instances, Indicates The query vector of the current input data instance, Indicates a key vector that refers to an instance of the input data, Indicates The key vectors for all other input data instances, represents the index of all other input data instances, Represents the index of the current input data instance, Indicates the index of the reference input data instance.

6. The object labeling method based on laser point cloud according to claim 1, characterized in that: Adjust the boundary division strategy based on the uncertainty assessment results of the boundary area to generate optimized segmentation results, including: The local mutual information entropy of the boundary points calculated by the attention mechanism network is analyzed, and the boundary points are divided into high uncertainty areas and low uncertainty areas according to the local mutual information entropy; For the boundary points in the high uncertainty area, the position of the boundary points is adjusted in combination with the local geometric topological characteristics, and the boundary is corrected by recalculating the geometric consistency between the boundary points and other points in the neighborhood, while the existing division of the boundary points in the low uncertainty area is maintained; Conduct continuity analysis on the distribution of boundary points in the corrected high uncertainty region to remove isolated and redundant points; Through the dynamically updated partitioning strategy, the adjusted boundary points of the high uncertainty region are merged with the boundary points of the low uncertainty region to generate a unified boundary partitioning result; The merged boundary division results are segmented and optimized to generate optimized segmentation results for target annotation using point cloud data.

7. A laser point cloud-based object labeling system, used to implement the laser point cloud-based object labeling method according to any one of claims 1 to 6, characterized in that: It includes point cloud data processing module, reflection area correction module, multi-layer iterative correction module, global and local feature module, boundary fine division module and boundary division optimization module; Point cloud data processing module: obtains the original point cloud data collected by the laser radar, and pre-processes the original point cloud data to retain valid point cloud data; Reflection area correction module: Based on the reflection intensity information and geometric topological features of the effective point cloud data, it identifies the non-Lambertian reflection area and performs filtering correction processing to generate preliminary processed point cloud data; Multi-layer iterative correction module: Perform multi-layer iterative correction on the initially processed point cloud data, correct the false mirror points in the non-Lambertian reflection area through deep filtering, and reconstruct the boundary based on the geometric topological characteristics to obtain the corrected point cloud data; Global and local feature module: Group and classify the target objects in the corrected point cloud data based on a clustering algorithm that combines global features and local features to generate preliminary segmentation results; Boundary fine division module: Use the attention mechanism network to finely divide the boundaries of the preliminary segmentation results, and monitor the local mutual information entropy output by the attention mechanism network to evaluate the uncertainty of the boundary area; Boundary division optimization module: adjusts the boundary division strategy according to the uncertainty assessment results of the boundary area and generates optimized segmentation results.

Images