Self-adaptive point cloud downsampling method

Abstract

The invention provides a self-adaptive point cloud downsampling method. An original point cloud is obtained based on a laser radar, then a polar coordinate three-dimensional unit is divided according to an azimuth angle, an elevation angle and a distance according to the original point cloud, and then a point cloud distribution condition in each sub-grid is determined based on a grid projection in a two-dimensional direction formed by the azimuth angle and the distance. And if the point cloud distribution condition meets a preset subdivision condition, subdividing the corresponding sub-grids, and adjusting the polar coordinate three-dimensional unit according to the subdivided grids in the two-dimensional direction to generate a target grid unit. According to the method, information loss or redundancy caused by uniform sampling is avoided, and finally, original point clouds in each target grid unit are processed to obtain reduced sampling point clouds, so that key environment characteristics are reserved to the greatest extent while the data volume is reduced, the processing efficiency and precision of a subsequent positioning algorithm are improved, and the requirement of real-time environment perception of the unmanned distribution vehicle is met.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of automatic driving, and in particular to an adaptive point cloud downsampling method. BACKGROUND

[0002] In the operation process of the unmanned delivery vehicle, the laser radar will generate a large amount of point cloud data. These original point cloud data are huge in quantity, which not only occupies a large amount of storage space and transmission bandwidth, but also brings heavy burden to the subsequent data processing and analysis, resulting in low calculation efficiency and failing to meet the real-time requirements of the unmanned delivery vehicle.

[0003] The existing point cloud downsampling methods, such as the general voxel downsampling method, can reduce the number of point clouds to a certain extent, but have obvious limitations. For example, in a specific environment such as the suburbs, when encountering a situation where there are only shrubs or weeds on both sides of the road, the point clouds are usually concentrated in a very small area. At this time, the voxel downsampling method will make the downsampled point clouds too sparse, losing a large amount of key information, thereby seriously affecting the subsequent positioning algorithm. In summary, the voxel downsampling method either leaves too few point clouds resulting in insufficient constraints, or retains too many points resulting in too many constraints in the optimization process, thereby reducing the efficiency. Therefore, in the optimization process, the unmanned delivery vehicle cannot provide accurate and reliable environmental information, affecting its driving safety and task execution efficiency. Therefore, an adaptive point cloud downsampling method is needed to maintain a reasonable number of point clouds after downsampling and ensure the downsampling efficiency, thereby improving the effectiveness and universality of positioning. SUMMARY

[0004] The purpose of the present application is to at least solve one of the above technical defects, in particular the technical defect that the existing point cloud downsampling cannot adaptively determine a reasonable number of point clouds, thereby affecting the effectiveness and universality of subsequent positioning.

[0005] In a first aspect, the present application provides an adaptive point cloud downsampling method, which comprises:

[0006] obtaining original point clouds based on a laser radar;

[0007] dividing a polar coordinate three-dimensional unit according to an azimuth angle, an elevation angle and a distance according to the original point clouds;

[0008] determining the point cloud distribution in each sub-grid based on the grid projection in the two-dimensional direction composed of the azimuth angle and the distance;

[0009] if the point cloud distribution meets a preset subdivision condition, subdividing the corresponding sub-grid, and adjusting the polar coordinate three-dimensional unit according to the subdivided grid in the two-dimensional direction to generate a target grid unit;

[0010] Processing each of the original point clouds in each of the target grid units to obtain a down-sampled point cloud.

[0011] As an optional implementation, the grid projection in the two-dimensional direction formed by the azimuth angle and the distance is used to determine the point cloud distribution in each sub-grid, including:

[0012] For each azimuth angle interval, the number of point clouds in each sub-grid is calculated as a first proportion of the total number of point clouds in the corresponding azimuth angle interval based on the grid projection in the two-dimensional direction formed by the azimuth angle and the distance.

[0013] In addition, the average height of all point clouds in each azimuth angle interval is determined, and the number of point clouds with a height exceeding a preset height threshold is calculated as a second proportion of the total number of point clouds in the corresponding azimuth angle interval.

[0014] If the point cloud distribution meets a preset subdivision condition, the corresponding sub-grid is subdivided, including:

[0015] When the first proportion exceeds a first threshold, each sub-grid corresponding to the azimuth angle interval is subdivided based on the azimuth angle, the elevation angle, and the distance.

[0016] In addition, when the second proportion is lower than a second threshold and / or the average height is lower than a preset height threshold, each sub-grid corresponding to the azimuth angle interval is subdivided based on the elevation angle.

[0017] As an optional implementation, the processing of each of the original point clouds in each of the target grid units to obtain a down-sampled point cloud includes:

[0018] Each of the original point clouds in each of the target grid units is fitted with a plane.

[0019] If the fitting is successful, a first centroid corresponding to the fitted plane is used to replace all original point clouds participating in the plane fitting, and the first centroid is used as a down-sampled point cloud.

[0020] If the fitting fails, a second centroid formed by each of the original point clouds in the corresponding target grid unit is calculated, and a sampling number is determined according to a function relationship positively related to the distance from the second centroid, and a down-sampled point cloud is obtained by sampling from each of the original point clouds in the corresponding target grid unit.

[0021] As an optional implementation, the plane fitting of each of the original point clouds in each of the target grid units includes:

[0022] solving a plane general equation of each of the original point clouds in each of the target grid cells according to a preset plane fitting algorithm, and calculating average residuals of each of the original point clouds in each of the target grid cells to the fitting plane;

[0023] if the average residual is less than a preset distance tolerance, determining that the plane fitting is successful, determining a fitting plane according to the plane general equation, and calculating coordinates of the first gravity center.

[0024] As an optional implementation, after the first gravity center is taken as the down-sampled point cloud, the method further comprises:

[0025] determining a first weight value corresponding to each of the down-sampled point clouds according to a total point cloud quantity of each of the original point clouds in the corresponding target grid cell;

[0026] and after the down-sampled point cloud is obtained by extracting each of the original point clouds in the corresponding target grid cell, the method further comprises:

[0027] determining a second weight value of each of the down-sampled point clouds according to the sampling quantity and a total point cloud quantity of each of the original point clouds in the corresponding target grid cell;

[0028] wherein the first weight value and the second weight value are used to indicate density information of point cloud distribution of each of the target grid cells.

[0029] As an optional implementation, the method further comprises:

[0030] calculating a distance standard deviation of each of the original point clouds in each of the target grid cells to the determined first gravity center or the second gravity center;

[0031] eliminating abnormal point clouds with a distance to the first gravity center or the second gravity center exceeding a preset multiple of the distance standard deviation, and generating an optimized point set;

[0032] recomputing the first gravity center or the second gravity center based on the optimized point set, and re-executing plane fitting according to the optimized point set to determine the down-sampled point cloud.

[0033] In a second aspect, the present application provides an adaptive point cloud down-sampling device, which comprises:

[0034] an acquisition module configured to acquire original point clouds based on a laser radar;

[0035] a processing module configured to divide polar coordinate three-dimensional cells according to azimuth angle, elevation angle and distance according to the original point clouds;

[0036] The processing module is further configured to determine a point cloud distribution in each sub-grid based on a grid projection in a two-dimensional direction formed by the azimuth angle and the distance;

[0037] The processing module is further configured to subdivide the corresponding sub-grid if the point cloud distribution meets a preset subdivision condition, and adjust the polar coordinate three-dimensional unit according to a grid after subdivision in the two-dimensional direction to generate a target grid unit.

[0038] The processing module is further configured to process each original point cloud in each target grid unit to obtain a down-sampled point cloud.

[0039] In a third aspect, the present application provides an unmanned delivery vehicle, comprising:

[0040] A laser radar is configured to collect an original point cloud.

[0041] A processing unit is configured to process the original point cloud according to the method of the first aspect to obtain a down-sampled point cloud.

[0042] A positioning controller is configured to perform pose solving and adjustment according to the down-sampled point cloud.

[0043] In a fourth aspect, the present application provides a computer device, comprising one or more processors and a memory, wherein the memory stores computer readable instructions, and the computer readable instructions are executed by the one or more processors to perform the steps of the method of the first aspect.

[0044] In a fifth aspect, the present application provides a storage medium, wherein the storage medium stores computer readable instructions, and the computer readable instructions are executed by one or more processors to cause the one or more processors to perform the steps of the method of the first aspect.

[0045] As can be seen from the above technical solutions, the embodiments of the present application have the following advantages:

[0046] Based on any of the above embodiments, the application provides a data basis for subsequent downsampling processing by acquiring original point clouds based on a laser radar, divides polar coordinate three-dimensional units according to the azimuth, elevation and distance of the original point clouds, divides the polar coordinate system to better match the data acquisition characteristics of the radar, improves the rationality of space division, analyzes the distribution of each sub-grid point cloud based on the two-dimensional grid projection formed by the azimuth and distance, identifies the dense area of the point cloud through two-dimensional fast scanning, subdivides the corresponding sub-grid and adjusts the polar coordinate unit to generate a target grid unit if the distribution meets the preset subdivision condition, ensures finer processing of high-density areas through adaptive subdivision, avoids information loss or redundancy caused by uniform sampling, performs plane fitting or barycentric sampling processing on the original point cloud in each target grid unit to obtain downsampled point clouds, retains the plane structure and distance distribution characteristics while reducing the data volume through feature extraction and adaptive sampling strategy, retains the point cloud density information through weight assignment during processing, improves the data quality through outlier removal, and finally outputs downsampled point clouds that can accurately represent the original environmental characteristics, greatly reduces the data storage and transmission overhead, improves the processing efficiency of subsequent positioning algorithms, and at the same time, ensures the accuracy and robustness of the pose solution by retaining key environmental features, meets the needs of unmanned delivery vehicles for real-time environmental perception and high-precision positioning in various complex environments, and improves the driving safety and task execution efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0048] Figure 1 The flowchart of the adaptive point cloud downsampling method provided by an embodiment of the present application is shown in the figure.

[0049] Figure 2 The effect diagram of the adaptive point cloud downsampling method provided by an embodiment of the present application is shown in the figure.

[0050] Figure 3 The internal structure diagram of the computer device provided by an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0052] In the operation process of the unmanned delivery vehicle, the laser radar will generate a large amount of point cloud data. These raw point cloud data are huge in quantity, not only occupying a large amount of storage space and transmission bandwidth, but also bringing heavy burden to subsequent data processing and analysis, resulting in low computing efficiency and failing to meet the real-time requirements of the unmanned delivery vehicle.

[0053] The existing point cloud downsampling methods, such as the general voxel downsampling method, can reduce the number of point clouds to a certain extent, but have obvious limitations. In a specific environment such as the suburbs, when encountering the situation that there are only shrubs or weeds on both sides of the road, the point clouds are usually concentrated in a very small area. At this time, the voxel downsampling method will make the down-sampled point cloud too sparse, losing a large amount of key information, thereby seriously affecting the subsequent positioning algorithm. Either too few remaining point clouds result in insufficient constraints, or too many points are retained, resulting in too many constraints in the optimization process, reducing the efficiency. In the optimization process, the unmanned delivery vehicle cannot provide accurate and reliable environmental information, affecting its driving safety and task execution efficiency. Therefore, this paper proposes an adaptive point cloud downsampling method, which can ensure the downsampling efficiency and the number of down-sampled point clouds is small, and realize the subsequent robust positioning.

[0054] The present application provides an efficient and robust point cloud downsampling method for unmanned delivery vehicles, which can reduce the amount of point cloud data while retaining the key information of the point cloud to the greatest extent, improve data processing efficiency, and enhance the environmental perception ability and driving safety of the unmanned delivery vehicle in various environments.

[0055] In summary, the technical concept of the present application is that the present application provides a data basis for subsequent downsampling processing by acquiring original point clouds based on a laser radar, divides polar coordinate three-dimensional units according to the original point clouds according to azimuth, elevation and distance, divides the polar coordinate system to better match the data acquisition characteristics of the radar, improves the rationality of space division, analyzes the distribution of each sub-grid point cloud based on the two-dimensional grid projection formed by the azimuth and distance, identifies the point cloud dense area through two-dimensional fast scanning, subdivides the corresponding sub-grid and adjusts the polar coordinate unit to generate a target grid unit if the distribution meets the preset subdivision condition, ensures finer processing of high-density areas through adaptive subdivision, avoids information loss or redundancy caused by uniform sampling, performs plane fitting or barycentric sampling processing on the original point clouds in each target grid unit to obtain downsampled point clouds, retains the plane structure and distance distribution characteristics while reducing the data volume through feature extraction and adaptive sampling strategy, retains the point cloud density information through weight assignment during processing, and improves the data quality through outlier removal, so as to finally output downsampled point clouds that can accurately represent the original environmental characteristics, greatly reduce the data storage and transmission overhead, improve the processing efficiency of subsequent positioning algorithms, and at the same time, ensure the accuracy and robustness of the pose solution by retaining key environmental characteristics, meet the needs of real-time environmental perception and high-precision positioning of unmanned delivery vehicles in various complex environments, and improve the driving safety and task execution efficiency of the unmanned delivery vehicles.

[0056] The method provided by the present application will be described in detail below according to the corresponding embodiments in some practical application scenarios.

[0057] Figure 1 The flowchart of the adaptive point cloud downsampling method provided by an embodiment of the present application is shown in FIG. 1, and the present application provides an adaptive point cloud downsampling method, which comprises the following steps. Figure 1

[0058] S101, acquiring original point clouds based on a laser radar;

[0059] S102, dividing polar coordinate three-dimensional units according to the original point clouds according to azimuth, elevation and distance;

[0060] azimuth, elevation and distance of the original point clouds are used to divide the point clouds by polar coordinates. In the initial division, a fixed resolution is used in the azimuth, elevation and distance dimensions to divide the point cloud space into a plurality of polar coordinate three-dimensional units.

[0061] ​Assuming that the original point cloud data acquired by the laser radar of the unmanned delivery vehicle is P, which contains the azimuth (denoted as a), elevation (denoted as β) and distance (denoted as d) information of each point. Set the fixed resolution in the azimuth dimension as δa, the fixed resolution in the elevation dimension as δβ, and the fixed resolution in the distance dimension (denoted as d) as δd. Taking the position of the laser radar as the origin, the point cloud space is divided into multiple polar coordinate three-dimensional cells in the polar coordinate mode. For any point p∈P, the initial grid cell coordinates (x, y, z) where the point is located can be processed by the floor function, for example, x = floor(a / δα), y = floor(β / δβ), z = floor(d / δd), and all point clouds are divided into corresponding polar coordinate three-dimensional cells.

[0062] S103, based on the grid projection in the two-dimensional direction formed by the azimuth and the distance, determine the point cloud distribution in each sub-grid;

[0063] Specifically, to simplify data processing and ensure as much effective information as possible, an azimuth-distance grid is constructed in a projection manner, using the azimuth and distance two dimensions to construct the grid. Subsequently, the proportion p between the number of point clouds in the sub-grid and the total number of original point clouds in the azimuth direction of the sub-grid can be calculated for the grid projection in the two-dimensional direction, so as to obtain the distribution of the original point clouds in each azimuth direction.

[0064] S104, if the point cloud distribution meets the preset subdivision condition, the corresponding sub-grid is subdivided, and the polar coordinate three-dimensional cell is adjusted according to the subdivided grid in the two-dimensional direction to generate a target grid cell;

[0065] For example, if the p value in a certain grid is too large (for example, greater than 0.3), it means that the original point clouds are concentrated in this grid and occupy a large proportion of the original point clouds in this azimuth direction. At this time, the sub-grid needs to be refined and subdivided using a finer grid. The subdivision ratio can be adaptively set to 1 / p, so that sufficient constraints can be ensured in each horizontal direction and the constraints in some horizontal directions will not be too small due to downsampling.

[0066] In addition, the average height of the original point cloud in each horizontal direction is counted, and if the proportion of the horizontal direction with a height of more than 2 m in each horizontal direction is less than 30%, it is considered that the number of point clouds at a high position is small, and these sub-grids also need to be subdivided to ensure the balance of the constraint between each height.

[0067] S105, processing each original point cloud in each target grid unit to obtain a down-sampled point cloud.

[0068] The processing in this step can include plane fitting processing, and the down-sampling method is determined according to the plane fitting processing.

[0069] Correspondingly, the application provides an unmanned delivery vehicle, comprising:

[0070] A laser radar is configured to collect original point clouds.

[0071] A processing unit is configured to process the original point clouds according to the method provided in any of the embodiments to obtain a down-sampled point cloud.

[0072] A positioning controller is configured to perform pose calculation and adjustment according to the down-sampled point cloud.

[0073] The method provided in the embodiment first acquires original point clouds based on a laser radar to provide a data basis for subsequent processing, and then divides polar coordinate three-dimensional units according to the original point clouds in terms of azimuth angle, elevation angle and distance. The polar coordinate division method can better match the characteristics of radar data collection, improve the rationality of space division, and then determine the point cloud distribution in each sub-grid based on the grid projection in the two-dimensional direction formed by the azimuth angle and the distance. The two-dimensional projection analysis can quickly identify the point cloud concentration area, reduce unnecessary calculation overhead, and if the point cloud distribution meets the preset subdivision condition, the corresponding sub-grid is subdivided, and the polar coordinate three-dimensional unit is adjusted according to the subdivided grid in the two-dimensional direction to generate a target grid unit. Through the adaptive subdivision strategy, the point cloud distribution dense area can be finely processed, avoiding information loss or redundancy caused by uniform sampling. Finally, the original point clouds in each target grid unit are processed to obtain a down-sampled point cloud. Through the partition processing method, the data amount is reduced while the key environmental features are retained to the greatest extent, thereby improving the efficiency and accuracy of subsequent positioning algorithm processing and meeting the real-time environmental perception requirements of the unmanned delivery vehicle.

[0074] As an optional embodiment, the grid projection in the two-dimensional direction formed by the azimuth angle and the distance is configured to determine the point cloud distribution in each sub-grid, comprising:

[0075] a first proportion of a point cloud quantity of each sub-grid in a corresponding azimuth angle interval to a total point cloud quantity in the corresponding azimuth angle interval is calculated based on a grid projection in a two-dimensional direction formed by the azimuth angle and the distance;

[0076] a second proportion of a point cloud quantity of each sub-grid in a corresponding azimuth angle interval to a total point cloud quantity in the corresponding azimuth angle interval is calculated based on a grid projection in a two-dimensional direction formed by the azimuth angle and the distance;

[0077] wherein, if the point cloud distribution meets a preset subdivision condition, the corresponding sub-grid is subdivided, including:

[0078] when the first proportion exceeds a first threshold, each sub-grid corresponding to the azimuth angle interval is subdivided based on the azimuth angle, the elevation angle and the distance;

[0079] and when the second proportion is lower than a second threshold and / or the average height is lower than a preset height threshold, each sub-grid in the corresponding azimuth angle interval is subdivided based on the elevation angle.

[0080] The method corresponding to the embodiment determines a first proportion of a point cloud quantity of each sub-grid in a corresponding azimuth angle interval to a total point cloud quantity in the corresponding azimuth angle interval based on a grid projection in a two-dimensional direction formed by the azimuth angle and the distance when determining the point cloud distribution, which can quantify the local point cloud density through proportion calculation. Meanwhile, a second proportion of a point cloud quantity of each sub-grid in a corresponding azimuth angle interval to a total point cloud quantity in the corresponding azimuth angle interval is determined based on an average height of all point clouds in the corresponding azimuth angle interval and a height exceeding a preset height threshold, which can evaluate the distribution characteristics of the point cloud in the vertical direction through the height-related index. When the first proportion exceeds a first threshold, each sub-grid corresponding to the azimuth angle interval is subdivided based on the azimuth angle, the elevation angle and the distance, which ensures a finer division in the high-density area through multi-dimensional subdivision. When the second proportion is lower than a second threshold or the average height is lower than a preset height threshold, each sub-grid in the corresponding azimuth angle interval is subdivided based on the elevation angle, which can enhance the sampling granularity in the low-height point cloud area through the targeted adjustment in the elevation direction, thereby improving the distribution balance of the point cloud in the horizontal and vertical directions after down-sampling and avoiding the positioning deviation caused by insufficient constraints in the traditional method in a complex environment.

[0081] As an optional embodiment, the processing of each original point cloud in each target grid unit to obtain a down-sampled point cloud includes:

[0082] plane fitting is performed on each original point cloud in each target grid unit;

[0083] if the fitting is successful, a first barycenter corresponding to the fitting plane is used to replace all original point clouds participating in the plane fitting, and the first barycenter is used as the down-sampled point cloud;

[0084] If the fitting fails, a second gravity center formed by each of the original point clouds in the corresponding target grid cell is calculated, the number of samples is determined according to a function relationship positively related to the distance from the second gravity center, and a down-sampled point cloud is obtained by extracting each of the original point clouds in the corresponding target grid cell.

[0085] The plane fitting processing in the application includes plane fitting for the original point cloud in each target grid cell after division. If the original point cloud can form a plane, all points in the target grid cell are replaced by the gravity center of the plane, and the number of points in this target grid cell is recorded for subsequent weighting of the residual. In this way, one point converges the information of all points in a plane, greatly reducing the number of original point clouds while retaining the plane feature information.

[0086] Non-planar original point cloud processing, if the original point cloud does not form a plane, the n nearest points to the gravity center are extracted from the original point cloud. The value of n is determined by the distance of the target grid cell. Since the original point cloud of the radar has the characteristics of near density and far sparseness, in the down-sampling, the target grid cell retains more points in the far area, thereby ensuring that there are enough original point clouds for environment perception. At the same time, each point is assigned a weight, and the weight value can be the total number of points in the target grid cell / n. In the positioning or map construction process, the point cloud is weighted and processed, which to some extent retains the density information of the original point cloud, and thus helps to improve the robustness of subsequent positioning.

[0087] The embodiment replaces all original point clouds participating in plane fitting with the first gravity center corresponding to the fitted plane and takes the first gravity center as the down-sampled point cloud if the plane fitting is successful. Through plane feature extraction and gravity center replacement, the number of point clouds can be greatly reduced while retaining the plane structure information. If the fitting fails, a second gravity center formed by each of the original point clouds in the corresponding target grid cell is calculated, the number of samples is determined according to a function relationship positively related to the distance from the second gravity center, and a down-sampled point cloud is obtained by extracting each of the original point clouds in the corresponding target grid cell. Through the distance adaptive sampling strategy, it is ensured that the far sparse point cloud area retains a sufficient number of points for environment perception, thereby reducing the data amount while maintaining the effectiveness of point cloud sampling at different distances and improving the stability and adaptability of subsequent positioning algorithms.

[0088] As an optional embodiment, the plane fitting of each of the original point clouds in each of the target grid cells includes:

[0089] According to a preset plane fitting algorithm, a plane general equation of each of the original point clouds in each of the target grid cells is solved, and an average residual of each of the original point clouds in each of the target grid cells to the fitted plane is calculated.

[0090] If the average residual is less than the preset distance tolerance, it is determined that the plane fitting is successful, the fitting plane is determined according to the plane general equation, and the coordinates of the first gravity center are calculated.

[0091] It should be understood that the method provided by the present application can also be used to fit multiple planes in one grid unit using the original point cloud data, thereby preserving multiple first gravity centers.

[0092] In the plane fitting process, the plane general equation of the original point cloud in each target grid unit is solved according to the preset plane fitting algorithm, and the average residual of each point to the fitting plane is calculated. Through residual analysis, the quality of the plane fitting of the point set can be objectively evaluated. If the average residual is less than the preset distance tolerance, it is determined that the plane fitting is successful, the fitting plane is determined according to the plane general equation, and the coordinates of the first gravity center are calculated. Through tolerance control, the high consistency of the fitting plane and the actual point cloud distribution can be ensured, the feature distortion caused by fitting error can be avoided, thereby improving the reliability of the plane replacement operation, making the down-sampled point cloud accurately represent the original plane structure, and enhancing the data reliability of the subsequent processing link.

[0093] As an optional embodiment, after the first gravity center is taken as the down-sampled point cloud, the method further comprises:

[0094] According to the total number of point clouds in each original point cloud in the corresponding target grid unit, a first weight value corresponding to each down-sampled point cloud is determined;

[0095] After the down-sampled point cloud is obtained by extracting each original point cloud in the corresponding target grid unit, the method further comprises:

[0096] According to the total number of point clouds in each original point cloud in the corresponding target grid unit, a first weight value corresponding to each down-sampled point cloud is determined;

[0097] The first weight value and the second weight value are used to indicate the density information of the point cloud distribution of each target grid unit.

[0098] After the first gravity center is taken as the down-sampled point cloud, the first weight value corresponding to each down-sampled point cloud is determined according to the total number of point clouds in the corresponding target grid unit, and after the down-sampled point cloud is obtained by extracting the original point cloud, the second weight value of each down-sampled point cloud is determined according to the total number of point clouds and the sampling number. Through weight assignment, the density distribution information of the original point cloud can be preserved, so that the down-sampled point cloud not only reduces the data amount but also can reflect the local point cloud density characteristics through the weight value, thereby providing more rich environmental structure information for the subsequent positioning algorithm, improving the adaptability of the algorithm to the uneven point cloud distribution scene in the optimization process, and enhancing the robustness and accuracy of the pose solution.

[0099] As an optional implementation, the method further comprises:

[0100] calculating a distance standard deviation of each of the original point clouds in each of the target grid cells to the determined first gravity center or the second gravity center;

[0101] eliminating abnormal point clouds with a distance to the first gravity center or the second gravity center exceeding a preset multiple of the distance standard deviation, to generate an optimized point set;

[0102] recomputing the first gravity center or the second gravity center based on the optimized point set, and re-executing plane fitting according to the optimized point set, to determine the down-sampled point cloud.

[0103] The embodiment quantifies the dispersion degree of point cloud distribution by calculating the distance standard deviation of the original point clouds in each target grid cell to the first gravity center or the second gravity center, eliminates abnormal point clouds with a distance to the gravity center exceeding a preset multiple of the distance standard deviation to generate an optimized point set, eliminates the interference of noise and outliers on gravity center calculation and plane fitting through abnormal point elimination, recomputes the gravity center or re-executes plane fitting based on the optimized point set to determine the down-sampled point cloud, further improves the quality of point cloud processing through an iterative optimization process, ensures that the down-sampling result more accurately represents the original environmental features, thereby enhancing the reliability of input data of subsequent positioning algorithms, avoiding pose estimation deviation caused by point cloud abnormalities, and improving the driving safety of the unmanned delivery vehicle in complex environments.

[0104] Each of the embodiments provided in the present application can be combined for implementation. In a specific application scenario, it can include:

[0105] Construct an azimuth-distance grid: on the basis of each polar coordinate three-dimensional cell, further construct an azimuth-distance two-dimensional statistical grid. Assuming that in a certain azimuth direction, the distance range is [d1, d2], it is divided into multiple sub-grids according to the distance resolution δd. For each sub-grid, the number of original point clouds Nij (i represents the index in the azimuth direction, and j represents the index of the distance sub-grid) and the total number of original point clouds Ni in the azimuth direction are counted. Calculate the proportion pij = Nij / Ni.

[0106] Grid refinement: Set a scale threshold Tp, when pij of a certain sub-grid > Tp, it means that the original point cloud is concentrated in the sub-grid, which needs to be refined. The refinement method is to redivide the sub-grid in the azimuth, elevation and distance dimensions according to finer resolution. At the same time, the average height of the original point cloud in each horizontal direction (reference to the azimuth dimension, divide multiple horizontal intervals) is counted. Set the horizontal direction interval as [βk, βk + 1], the average height of the original point cloud in the interval is Hk, and the count Mk of the horizontal direction whose height Hk exceeds the threshold Th (2m) is counted. When Mk / ∑Mk< 0.3, the target grid unit corresponding to the horizontal direction is subdivided.

[0107] Reduce information loss and improve efficiency:

[0108] Plane fitting processing: For the original point cloud in each target grid unit after the above refinement, assume that the point set in the target grid unit is S. First, determine whether the point set S can form a plane by using a plane fitting algorithm such as least squares. If it can form a plane, calculate the equation of the plane Ax + By + Cz + D = 0 and the center of gravity G (x0, y0, z0) of the plane. Replace all points in the target grid unit with the center of gravity G, and record the number of points Ns in the target grid unit.

[0109] Non-planar original point cloud processing: If the point set S cannot form a plane, calculate the center of gravity G of the point set S. According to the distance value γ of the target grid unit, determine the number of extracted points n. For example, set a function relationship n =avg_dist / 5, so that the farther the distance, the more points are extracted. Extract the n closest points to the center of gravity G from the point set S, and assign each point a weight w = Ns / n, Ns is the total number of points in the target grid unit. Finally, the processed original point cloud is integrated to obtain the down-sampled original point cloud data.

[0110] Through the above specific embodiments, efficient and robust original point cloud down-sampling for unmanned delivery vehicles can be effectively realized, providing stable and efficient original point cloud data for positioning of unmanned delivery vehicles, thereby improving the efficiency and accuracy of subsequent positioning algorithm processing and meeting the real-time environmental perception needs of unmanned delivery vehicles.

[0111] Figure 2 The effect diagram of the adaptive point cloud down-sampling method provided by an embodiment of the present application is as follows: Figure 2As shown, the present application has carried out relevant experiments in actual block and narrow road scenes, and the traditional downsampling has appeared positioning drift in some positions of narrow or small roads, while the corresponding method of the present application can smoothly pass. The yellow annular is the original point cloud on the ground hit by the laser radar, and since the azumth elvation distance is divided at a fixed resolution, the ground point cloud is sparse. The red point cloud is the prior positioning map used for positioning. The blue is the trajectory of the unmanned vehicle, the yellow is the non-planar point, and the purple is the planar point. Because it is a bush scene, only a small number of voxels meet the planar requirements (purple points), and most of them are non-planar points (yellow points). The coordinate system on the blue trajectory (slim coordinate system) is the position obtained by using the downsampling of the present application, and the other short and thick one is the trajectory obtained by the traditional method (the positioning has drifted and made an error at this time). In the bush scene, the adaptive downsampling of the present application is more robust than the fixed resolution voxel downsampling.

[0112] In the present application, through polar coordinate division and targeted grid refinement strategy, the area that needs to be focused on can be quickly determined, and unnecessary calculation amount is reduced. At the same time, the plane fitting and the distance-based point extraction method greatly reduce the number of point clouds and improve the efficiency of data processing on the premise of ensuring that information is not lost, meeting the real-time requirements of unmanned delivery vehicles. At the same time, the distribution characteristics of point clouds in different environments are fully considered, especially for the concentrated distribution of point clouds that may occur in complex environments such as suburban areas. Through reasonable grid refinement and point cloud processing methods, the key information of point clouds can be retained to the greatest extent, enhancing the environmental perception ability of unmanned delivery vehicles in various environments and improving their robustness and safety in driving.

[0113] The present application also provides an adaptive point cloud downsampling device, which comprises:

[0114] The acquisition module is configured to acquire an original point cloud based on a laser radar.

[0115] The processing module is configured to divide a polar coordinate three-dimensional unit according to an azimuth angle, an elevation angle and a distance based on the original point cloud.

[0116] The processing module is further configured to determine a point cloud distribution in each sub-grid based on a grid projection in a two-dimensional direction formed by the azimuth angle and the distance.

[0117] The processing module is further configured to subdivide the corresponding sub-grid if the point cloud distribution meets a preset subdivision condition, and adjust the polar coordinate three-dimensional unit based on the subdivided grid in the two-dimensional direction to generate a target grid unit.

[0118] The processing module is further configured to process each of the original point clouds in each of the target grid units to obtain a down-sampled point cloud.

[0119] The method provided by the embodiment first acquires an original point cloud based on a laser radar, thereby providing a data basis for subsequent processing, and then divides polar coordinate three-dimensional units according to an azimuth angle, an elevation angle and a distance. The polar coordinate division manner can better match the collection characteristics of the radar data, thereby improving the rationality of the spatial division. Then, the point cloud distribution in each sub-grid is determined based on a grid projection in a two-dimensional direction formed by the azimuth angle and the distance. The two-dimensional projection analysis can quickly identify a point cloud concentrated area, thereby reducing unnecessary calculation overhead. If the point cloud distribution meets a preset subdivision condition, the corresponding sub-grid is subdivided. The polar coordinate three-dimensional units are adjusted according to the subdivided grid in the two-dimensional direction to generate a target grid unit. The adaptive subdivision strategy can perform fine processing on a point cloud dense area, thereby avoiding information loss or redundancy caused by uniform sampling. Finally, the original point cloud in each target grid unit is processed to obtain a down-sampled point cloud. Through the partition processing manner, the data amount is reduced while the key environmental features are retained to the greatest extent, thereby improving the efficiency and accuracy of subsequent positioning algorithm processing and meeting the real-time environmental perception requirement of the unmanned delivery vehicle.

[0120] As an optional implementation manner, the processing module determines the specific manner of determining the point cloud distribution in each sub-grid based on the grid projection in the two-dimensional direction formed by the azimuth angle and the distance, which includes:

[0121] For each azimuth angle interval, the processing module calculates a first proportion of the point cloud quantity of each sub-grid to the total point cloud quantity of the azimuth angle interval to which the sub-grid belongs based on the grid projection in the two-dimensional direction formed by the azimuth angle and the distance.

[0122] In addition, the processing module determines the average height of all point clouds in each azimuth angle interval and a second proportion of the number of point clouds with a height exceeding a preset height threshold to the total point cloud quantity of the azimuth angle interval to which the point clouds belong.

[0123] If the point cloud distribution meets a preset subdivision condition, the processing module subdivides the corresponding sub-grid in the specific manner, which includes:

[0124] When the first proportion exceeds a first threshold, each sub-grid corresponding to the azimuth angle interval is subdivided based on the azimuth angle, the elevation angle and the distance.

[0125] In addition, when the second proportion is lower than a second threshold and / or the average height is lower than a preset height threshold, each sub-grid corresponding to the azimuth angle interval is subdivided based on the elevation angle.

[0126] The method corresponding to the embodiment determines the point cloud distribution, calculates a first proportion of the number of points in each sub-grid in a corresponding azimuth angle interval to the total number of points in the azimuth angle interval based on the grid projection in the two-dimensional direction formed by the azimuth angle and the distance, quantifies the local point cloud density through the proportion calculation, determines the average height of all points in each azimuth angle interval and a second proportion of the number of points whose height exceeds a preset height threshold to the total number of points in the azimuth angle interval, evaluates the distribution characteristics of the point cloud in the vertical direction through the height-related indicators, subdivides each sub-grid in the corresponding azimuth angle interval based on the azimuth angle, the elevation angle and the distance when the first proportion exceeds a first threshold, ensures finer division in the high-density area through multi-dimensional subdivision, subdivides each sub-grid in the corresponding azimuth angle interval based on the elevation angle when the second proportion is lower than a second threshold or the average height is lower than a preset height threshold, and enhances the sampling granularity of the low-height point cloud area through the targeted adjustment in the elevation direction, thereby improving the balance of the distribution of the point cloud in the horizontal and vertical directions after down-sampling, and avoiding the positioning deviation caused by insufficient constraints in the traditional method in a complex environment.

[0127] As an optional embodiment, the processing module processes each of the original point clouds in each of the target grid units to obtain the specific manner of the down-sampled point cloud, including:

[0128] plane fitting is performed on each of the original point clouds in each of the target grid units;

[0129] if the plane fitting is successful, a first gravity center corresponding to the fitted plane is used to replace all the original point clouds participating in the plane fitting, and the first gravity center is taken as the down-sampled point cloud;

[0130] if the plane fitting fails, a second gravity center formed by each of the original point clouds in the corresponding target grid unit is calculated, a sampling number is determined according to a function relationship positively related to the distance from the second gravity center, and the down-sampled point cloud is obtained by extracting from each of the original point clouds in the corresponding target grid unit.

[0131] In the embodiment, plane fitting is performed on the original point clouds in each target grid unit, if the plane fitting is successful, a first gravity center corresponding to the fitted plane is used to replace all the original point clouds participating in the plane fitting, and the first gravity center is taken as the down-sampled point cloud, the plane feature extraction and gravity center replacement can greatly reduce the number of point clouds while retaining the plane structure information, if the plane fitting fails, a second gravity center formed by each of the original point clouds in the corresponding target grid unit is calculated, a sampling number is determined according to a function relationship positively related to the distance from the second gravity center, and the down-sampled point cloud is obtained by extracting from each of the original point clouds in the corresponding target grid unit, the distance adaptive sampling strategy ensures that a sufficient number of points are retained in the distant sparse point cloud area for environment perception, thereby maintaining the effectiveness of point cloud sampling at different distances while reducing the data volume, and improving the stability and adaptability of the subsequent positioning algorithm.

[0132] As an optional implementation, the specific manner of the processing module performing plane fitting on each of the original point clouds in each of the target grid cells comprises:

[0133] According to a preset plane fitting algorithm, a plane general equation of each of the original point clouds in each of the target grid cells is solved, and an average residual of each of the original point clouds in each of the target grid cells to the fitted plane is calculated.

[0134] If the average residual is less than a preset distance tolerance, it is determined that the plane fitting is successful, a fitted plane is determined according to the plane general equation, and the coordinates of the first gravity center are calculated.

[0135] In the plane fitting process, the present embodiment solves a plane general equation of the original point cloud in each target grid cell according to a preset plane fitting algorithm, and calculates an average residual of each point to the fitted plane. Through residual analysis, the quality of plane fitting of the point set can be objectively evaluated. If the average residual is less than a preset distance tolerance, it is determined that the plane fitting is successful, a fitted plane is determined according to the plane general equation, and the coordinates of the first gravity center are calculated. Through tolerance control, the high consistency of the fitted plane and the actual point cloud distribution can be ensured, the feature distortion caused by fitting error can be avoided, and thus the reliability of the plane replacement operation is improved, so that the down-sampled point cloud can accurately represent the original plane structure, and the data reliability of the subsequent processing link is enhanced.

[0136] As an optional implementation, the processing module is further configured to, after obtaining the first gravity center as the down-sampled point cloud, determine a first weight value corresponding to each of the down-sampled point clouds according to a total point cloud quantity of each of the original point clouds in the corresponding target grid cell.

[0137] In addition, the processing module is further configured to, after extracting the down-sampled point cloud from each of the original point clouds in the corresponding target grid cell, determine a second weight value of each of the down-sampled point clouds according to the sampling quantity and the total point cloud quantity of each of the original point clouds in the corresponding target grid cell.

[0138] The first weight value and the second weight value are used to indicate the density information of the point cloud distribution of each of the target grid cells.

[0139] The first weight value corresponding to each down-sampling point cloud is determined according to the total number of point clouds in the corresponding target grid cell after the first gravity center is taken as the down-sampling point cloud, and the second weight value of each down-sampling point cloud is determined according to the sampling number and the total number of point clouds after the down-sampling point cloud is obtained by sampling from the original point cloud. The density distribution information of the original point cloud can be retained through weight assignment, so that the down-sampled point cloud not only reduces the data amount, but also reflects the local point cloud density characteristics through the weight value, thereby providing more abundant environmental structure information for subsequent positioning algorithms, improving the adaptability of the algorithm to the point cloud distribution uneven scene in the optimization process, and enhancing the robustness and accuracy of the pose solution.

[0140] As an optional implementation, the processing module is further configured to:

[0141] calculate the distance standard deviation of each original point cloud in each target grid cell to the determined first gravity center or second gravity center;

[0142] remove abnormal point clouds with a distance to the first gravity center or the second gravity center exceeding a preset multiple of the distance standard deviation, and generate an optimized point set;

[0143] re-calculate the first gravity center or the second gravity center based on the optimized point set, and re-perform plane fitting according to the optimized point set to determine the down-sampling point cloud.

[0144] The distance standard deviation of the original point cloud in each target grid cell to the first gravity center or the second gravity center is calculated to quantify the dispersion degree of the point cloud distribution, and the abnormal point clouds with a distance to the gravity center exceeding a preset multiple of the distance standard deviation are removed to generate an optimized point set. Through the removal of abnormal points, the interference of noise and outliers on gravity center calculation and plane fitting can be eliminated. The gravity center is re-calculated or the plane fitting is re-performed based on the optimized point set to determine the down-sampling point cloud. Through the iterative optimization process, the quality of the point cloud processing is further improved, and the down-sampling result is ensured to more accurately represent the original environmental characteristics, thereby enhancing the reliability of the input data of the subsequent positioning algorithm, avoiding the pose estimation deviation caused by point cloud abnormalities, and improving the driving safety of the unmanned delivery vehicle in complex environments.

[0145] It should be noted that the division of each module of the above apparatus is only a logical function division, and all or part of the actual implementation can be integrated into a physical entity, or can be physically separated. And these modules can all be realized in the form of software called by a processing element; all can be realized in the form of hardware; some modules can be realized in the form of software called by a processing element, and some modules can be realized in the form of hardware. For example, the processing module can be a separately established processing element, or can be integrated in a chip of the above apparatus, in addition, it can also be stored in the form of program code in the memory of the above apparatus, and the function of the above determination module is called and executed by a processing element of the above apparatus. The implementation of other modules is similar. In addition, all or part of these modules can be integrated together, or can be independently implemented. The processing element here can be an integrated circuit with signal processing capability. In the implementation process, each step of the above method or each module can be completed by the integrated logic circuit of hardware in the processor element or the instruction in the form of software.

[0146] As shown in Figure 3 , Figure 3 An internal structure schematic diagram of a computer device provided by an embodiment of the present application is shown in the figure. The computer device 300 can be provided as a server. Referring to Figure 3 , the computer device 300 includes a processing component 302, which further includes one or more processors, and a memory resource represented by a memory 301, for storing instructions executable by the processing component 302, such as an application program. The application program stored in the memory 301 can include one or more than one module each corresponding to a set of instructions. In addition, the processing component 302 is configured to execute the instructions to perform the method of any of the above embodiments.

[0147] The computer device 300 can also include a power supply component 303 configured to perform power management of the computer device 300, a wired or wireless network interface 304 configured to connect the computer device 300 to a network, and an input / output (I / O) interface 305. The computer device 300 can operate based on an operating system stored in the memory 301, such as Windows Server TM, Mac OS X TM, Unix TM, Linux TM, Free BSD TM or the like.

[0148] Those skilled in the art can understand that Figure 3 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0149] The embodiments of the present application provide a storage medium, the storage medium stores computer readable instructions, the computer readable instructions are executed by one or more processors, so that the one or more processors execute the method provided by any one of the embodiments.

[0150] Finally, it should be noted that in this document, the terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such process, method, article or device. Without more limitations, the element defined by the statement "including a" does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0151] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The various embodiments can be combined as needed, and the same and similar parts refer to each other.

[0152] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to the embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An adaptive point cloud down-sampling method, characterized in that, The method comprises: obtaining an original point cloud based on a laser radar; dividing a polar coordinate three-dimensional unit according to an azimuth angle, an elevation angle and a distance according to the original point cloud; determining a point cloud distribution in each sub-grid based on a grid projection in a two-dimensional direction formed by the azimuth angle and the distance; if the point cloud distribution meets a preset subdivision condition, subdividing the corresponding sub-grid, and adjusting the polar coordinate three-dimensional unit according to the grid after subdivision in the two-dimensional direction to generate a target grid unit; processing each original point cloud in each target grid unit to obtain a down-sampling point cloud.

2. The method of claim 1, wherein, The method for determining the point cloud distribution in each sub-grid based on the grid projection in the two-dimensional direction formed by the azimuth angle and the distance comprises: based on the grid projection in the two-dimensional direction formed by the azimuth angle and the distance, calculating, for each azimuth angle interval, a first proportion of the number of point clouds in each sub-grid to the total number of point clouds in the azimuth angle interval to which the sub-grid belongs; determining an average height of all point clouds in each azimuth angle interval and a second proportion of the number of point clouds with a height exceeding a preset height threshold to the total number of point clouds in the azimuth angle interval to which the point clouds belong; wherein, if the point cloud distribution meets the preset subdivision condition, the corresponding sub-grid is subdivided, comprising: when the first proportion exceeds a first threshold, each sub-grid corresponding to the azimuth angle interval is subdivided based on the azimuth angle, the elevation angle and the distance; and when the second proportion is lower than a second threshold and / or the average height is lower than a preset height threshold, each sub-grid corresponding to the azimuth angle interval is subdivided based on the elevation angle.

3. The method of claim 1, wherein, The method for processing each original point cloud in each target grid unit to obtain a down-sampling point cloud comprises: plane fitting is performed on each original point cloud in each target grid unit; if the fitting is successful, a first barycenter corresponding to the fitting plane is used to replace all original point clouds participating in the plane fitting, and the first barycenter is taken as a down-sampling point cloud; if the fitting fails, a second barycenter formed by each original point cloud in the corresponding target grid unit is calculated, a sampling number is determined according to a function relationship positively related to the distance from the second barycenter, and a down-sampling point cloud is obtained by sampling from each original point cloud in the corresponding target grid unit.

4. The method of claim 3, wherein, The method for performing plane fitting on each original point cloud in each target grid unit comprises: a plane general equation of each original point cloud in each target grid unit is solved according to a preset plane fitting algorithm, and an average residual error of each original point cloud in each target grid unit to the fitting plane is calculated; if the average residual error is less than a preset distance tolerance, it is determined that the plane fitting is successful, a fitting plane is determined according to the plane general equation, and the coordinates of the first barycenter are calculated.

5. The method of claim 3, wherein, After the first barycenter is taken as a down-sampling point cloud, the method further comprises: determining a first weight value corresponding to each down-sampling point cloud according to the total number of point clouds in each original point cloud in the corresponding target grid unit; And, after the down-sampling point cloud is obtained by extracting from each of the original point clouds in the corresponding target grid unit, the method further comprises: determining a second weight value of each of the down-sampling point clouds according to the sampling number and a total point cloud number of each of the original point clouds in the corresponding target grid unit; wherein the first weight value and the second weight value are used to indicate the density information of the point cloud distribution of each of the target grid units.

6. The method according to any one of claims 3-5, characterized in that, The method further comprises: calculating a distance standard deviation of each of the original point clouds in each of the target grid units to the determined first barycenter or the second barycenter; eliminating abnormal point clouds whose distance to the first barycenter or the second barycenter exceeds a preset multiple of the distance standard deviation, to generate an optimized point set; recomputing the first barycenter or the second barycenter based on the optimized point set, and re-executing plane fitting according to the optimized point set to determine the down-sampling point cloud.

7. An apparatus for adaptive point cloud downsampling, the apparatus comprising: The device comprises: an acquisition module configured to acquire original point clouds based on a laser radar; a processing module configured to divide polar coordinate three-dimensional units according to azimuth, elevation and distance based on the original point clouds; the processing module is further configured to determine point cloud distribution in each sub-grid based on grid projection in a two-dimensional direction formed by the azimuth and the distance; the processing module is further configured to subdivide the corresponding sub-grid if the point cloud distribution meets a preset subdivision condition, and adjust the polar coordinate three-dimensional units based on the subdivided grid in the two-dimensional direction to generate target grid units; the processing module is further configured to process each of the original point clouds in each of the target grid units to obtain a down-sampling point cloud.

8. An unmanned delivery vehicle, characterized by comprises: a laser radar configured to collect original point clouds; a processing unit configured to process the original point clouds according to the method of any one of claims 1-6 to obtain a down-sampling point cloud; a positioning controller configured to perform pose solving and adjustment based on the down-sampling point cloud.

9. A computer device, comprising: comprises one or more processors and a memory having computer readable instructions stored therein, which, when executed by the one or more processors, perform the steps of the method of any one of claims 1-6.

10. A storage medium, characterized by The storage medium stores computer readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the method of any one of claims 1-6.

Images