Intelligent wheelchair ground point cloud segmentation method and system

Abstract

The invention discloses an intelligent wheelchair ground point cloud segmentation method, which comprises the following steps of: acquiring a point cloud through a radar mounted on a wheelchair, preprocessing original point cloud data, performing grid division on the point cloud by using a multi-resolution division method based on a radial interval and an angle interval to obtain sub-grid units, and dividing the sub-grid units into a plurality of sub-grid units; and then classifying the sub-grid units based on height difference and variance information, and finally estimating point clouds in the PG grids by using a Bayesian probability model to obtain ground points in the point clouds, merging all the ground points to obtain an estimated ground point cloud, and outputting the estimated ground point cloud. The method needs few computing resources, can obtain a ground segmentation result with higher precision, can well solve the problems that a traditional method cannot adapt to complex terrains and is poor in sparsity and real-time performance of point clouds, and can improve the efficiency of detecting a driving area of an intelligent wheelchair or other mobile robots.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent wheelchair sensing technology, and in particular relates to an intelligent wheelchair ground point cloud segmentation method and system. Background Art

[0002] In recent years, with the rapid development of intelligent driving technology, autonomous driving systems have gradually moved from theoretical exploration to practical application, and have shown great potential in the fields of automobiles, unmanned delivery, urban security, etc. In these application scenarios, environmental perception technology, as the core module of the intelligent driving system, provides necessary environmental information support for navigation, path planning and obstacle avoidance. As a key environmental perception sensor, LiDAR has become an indispensable tool in the field of intelligent driving with its high precision, strong robustness and all-weather working ability. As a typical application scenario combining mobile robot technology with medical auxiliary equipment, smart wheelchairs have higher requirements for environmental perception technology in autonomous navigation. Unlike smart cars that mainly travel on structured roads, smart wheelchairs need to operate in more complex unstructured environments, such as homes, hospitals, parks, etc. These environments often have challenges such as irregular terrain, blurred road boundaries, and dynamic obstacles, making the ground segmentation task more difficult. Therefore, accurate ground point cloud segmentation is of great significance to smart wheelchairs. It is a key step in detecting safe drivable areas, achieving autonomous navigation, and ensuring the safety of users.

[0003] At present, the methods for ground point cloud segmentation can be mainly divided into four categories: 1) The method based on elevation map, which projects the three-dimensional point cloud onto a two-dimensional grid and extracts grid features for fast processing. However, this method is prone to over-segmentation in complex terrain and is difficult to adapt to steep terrain or overhanging structures; 2) The method based on plane fitting, which regards the ground as a local flat area and segments the ground points by fitting a plane model, but this method has poor adaptability to complex and irregular terrain, especially in actual unstructured environments; 3) The method based on ground plane estimation, which estimates the ground height using statistical models such as Gaussian regression or Markov random fields. Although the segmentation accuracy is high, the computational complexity is too high, which is not conducive to real-time applications; 4) The method based on deep learning, which relies on advanced neural network structures to perform semantic segmentation on point cloud data, has good performance but has problems such as lack of data sets, poor network interpretability and high computing resource requirements. The patent "A three-stage point cloud ground segmentation method based on ground plane fitting (202311711321.8)" proposes a point cloud ground segmentation algorithm, which uses large-resolution infinite polar coordinate grid division and voxel downsampling to fit the ground plane to preliminarily divide the point cloud, and then uses ground plane fitting to extract false negative points and false positive points. After three-order optimization adjustment, it finally achieves accurate ground point cloud segmentation. However, this method has certain shortcomings: the same resolution is used for grid division of near-distance and long-distance point clouds, and the same threshold is used for screening, which easily misses sparse ground point clouds in the distance, and the method of multiple ground plane fitting is time-consuming and lacks real-time performance. Summary of the invention

[0004] To solve the above problems, the present invention discloses a method and system for ground point cloud segmentation of an intelligent wheelchair, which accurately segments ground point clouds based on a multi-resolution grid division method, a grid classification method and a ground point probability estimation method, requires less computing resources and can obtain higher-precision ground segmentation results. It can well solve the problems that traditional methods cannot adapt to complex terrain and the sparsity of point clouds, and have poor real-time performance, and can help intelligent wheelchairs or other mobile robots detect drivable areas and achieve autonomous navigation.

[0005] To achieve the above object, the technical solution of the present invention is as follows:

[0006] A method for segmenting ground point cloud of an intelligent wheelchair comprises the following steps:

[0007] S1, by installing it on a wheelchair with a height of h sensor The laser radar collects point clouds, pre-processes the raw point cloud data, and filters the point cloud data according to the preset distance and height thresholds;

[0008] S2, using a multi-resolution partitioning method based on radial intervals and angle intervals to perform grid partitioning on the point cloud to obtain sub-grid units, wherein the partitioning method uses different resolutions to partition the point clouds in different radial intervals and angle intervals to adapt to the sparsity of the point cloud, and calculates and stores the attributes of each sub-grid unit during the rasterization process;

[0009] S3. Classify the sub-grid cells based on the height difference and variance information into three categories: G, NG, and PG. The G category is a grid that is all ground points (Ground Cell); the NG category is a grid that is all non-ground points (Potential Ground Cell); and the PG category is a grid that may contain ground points (Non-Ground Cell).

[0010] S4. Calculate the probability of ground points based on the Bayesian probability model, estimate the point cloud in the PG grid to obtain the ground points therein, merge all the ground points, obtain the estimated ground point cloud and output it; the merging method is to take the union of the ground points of category G and the ground points obtained by Bayesian estimation.

[0011] Further, step S1 is explained in detail:

[0012] During the LiDAR acquisition process, laser reflection may cause noise points to appear below the actual ground, causing under-segmentation problems. Therefore, the original point cloud data is pre-processed, and a multi-condition screening method based on height and distance is used to filter the point cloud data. It will reject point clouds with z values ​​below the set threshold within a certain distance, thereby effectively reducing outliers and significantly reducing their interference with ground segmentation results.

[0013] Further, step S2 is explained in detail:

[0014] The fixed interval polar coordinate division method is prone to sparsity problems when processing point clouds in long-distance areas. At the same time, too small grids in close-distance areas may cause feature estimation errors. Therefore, different resolutions are used to divide grids for point clouds in different radial and angular intervals to adapt to the sparsity of point clouds and improve segmentation accuracy. Specifically, M is divided unequally in the radial direction according to the distribution characteristics of the point cloud. z concentric regions, and divide them into M z angle intervals, and obtain the rough division M z ×N zgrid cells; the grid is further subdivided into sub-grid cells using different radial and angular resolutions. The radial and angular resolutions are determined by dividing the radial boundary difference of the concentric areas and the boundary difference of the angular interval by the number of subdivisions, respectively. The number of subdivisions is flexibly adjusted according to the grids at different locations; during the subdivision process, the lowest point, highest point, height variance, and number of point clouds of each sub-grid cell are calculated as its attributes and stored.

[0015] Further, step S3 is explained in detail:

[0016] It is time-consuming to perform ground estimation on a large amount of point cloud data. In order to improve the segmentation efficiency of the algorithm, the sub-grid units are classified based on the height difference and variance information, and the grid point cloud of category G is marked as ground points in advance, reducing the points that require Bayesian probability estimation, greatly improving the segmentation speed. Specifically, all sub-grid units are initialized as grids (PG) that may contain ground points; the distance from the lowest point of each sub-grid unit to the sensor is calculated and its height attribute is updated. When the height of the lowest point of the grid is less than the set lower limit of the height, the lower limit of the height is used as the grid height attribute. When the height of the lowest point of the grid is greater than the set upper limit of the height, the sensor height is used as the grid height attribute. When it is between the upper and lower limits of the height, the height of the lowest point of the grid is used as its height attribute.

[0017] The sub-grid cells that meet the following conditions at the same time are marked as ground grids (G): the number of point clouds in the grid exceeds the set minimum number of points, the height variance of the grid is greater than zero and less than threshold one, and the height attribute of the grid is less than the preset threshold two; threshold one is dynamically adjusted based on the distance of the grid; the sub-grid cells that meet any of the following conditions are marked as non-ground grids (NG): condition one, the distance attribute of the grid is greater than the preset threshold three, and the height attribute of the grid is equal to the sensor height; condition two, the distance attribute of the grid cell is greater than the preset threshold three, and the height difference between the highest point and the lowest point in the grid is greater than the preset threshold four.

[0018] Since the height variance increases due to the sparse distribution of point clouds in long-distance areas, in order to maintain a high recognition rate in close-distance areas while having a higher tolerance for point cloud height variance in long-distance areas, it is necessary to adaptively adjust the height variance threshold: the value obtained by multiplying the preset distance scaling factor by the grid distance is used as the height variance threshold of different grids, and adaptive adjustment of the threshold based on distance is achieved, thereby improving the accuracy and robustness of ground grid recognition in long-distance areas.

[0019] Further, step S4 is explained in detail:

[0020] Ground point estimation of PG sub-grid units based on Bayesian probability model: Plane point clouds can usually be distinguished based on the angle features of point clouds, but there are usually non-ground structures such as car hoods or roofs in outdoor environments. In order to reliably distinguish ground points, the joint probability of the angle probability and the height difference probability of each point is calculated to estimate the possibility of it being a ground point; the angle between the normal vector of the point and the vertical direction of the ground is calculated. If the angle is less than the preset boundary angle value, the angle probability of the point is set to 1, indicating that it is more likely to belong to a ground point; otherwise it is set to 0. Different boundary angle values ​​are set in different areas based on experience and point cloud distribution characteristics; the difference between the height of the highest point and the height of the lowest point in the grid is calculated as the grid height difference, and the height difference is subtracted from the threshold of five to obtain a difference. The exponential form of this difference is calculated through an exponential function, and the inverse is taken after adding one to it as the height difference probability; the ground point probability is obtained by multiplying the angle probability with the height difference probability. Points with a ground point probability greater than the preset threshold of six are estimated to be ground points.

[0021] Considering the sparsity of long-distance point clouds, in order to adapt to the requirements of height differences at different distances, the height difference threshold must be adaptively adjusted: Threshold 5 is set as a dynamic function related to the distance from the point to the sensor, and its value is calculated in an exponential form. Specifically, Threshold 5 is equal to an initial value multiplied by the exponential growth of the distance, and the growth rate is controlled by a constant. Among them, the initial value determines the size of Threshold 5 at close distances, and the growth rate controls the growth rate of Threshold 5 as the distance increases. This definition method can dynamically adjust the tolerance range of the height difference according to the distance from the point to the sensor, and adapt to the point cloud characteristics of different distance areas.

[0022] The present invention proposes a smart wheelchair ground point cloud segmentation system, comprising:

[0023] LiDAR scanning module, used to collect 3D point cloud information during the movement of the wheelchair;

[0024] The original point cloud preprocessing module is used to filter the collected point cloud;

[0025] The multi-resolution grid division module is used to divide point clouds in different radial and angular intervals with differentiated resolutions to improve ground segmentation accuracy;

[0026] The grid attribute storage module is used to calculate the lowest point, highest point, height variance and point cloud quantity of the grid when dividing the grid, and store them as grid attributes;

[0027] The grid classification module is used to classify the subdivided subgrid units, first find the grids that are all ground points, and reduce the number of ground points that need to be estimated later;

[0028] The Bayesian probability ground point estimation module is used to perform probability estimation on the potential ground point grids classified in the grid classification module to obtain the ground points therein;

[0029] The estimated ground point merging module is used to merge all estimated ground points to obtain an estimated ground point cloud set.

[0030] The beneficial effects of the present invention are:

[0031] (1) This method uses different resolutions to divide the point clouds in different radial and angular intervals into grids, so that it can adapt to the sparsity problem of point clouds when processing point clouds in long-distance areas, accurately identify distant ground points, and improve the accuracy of ground segmentation.

[0032] (2) It is time-consuming to perform ground estimation on a large amount of point cloud data. In order to improve the segmentation efficiency of the algorithm, this method classifies the sub-grid units based on the height difference and variance information, and marks the grid point cloud of category G as ground points in advance to reduce the points that require Bayesian probability estimation, greatly improving the segmentation speed; the height variance threshold adaptive adjustment strategy is adopted to improve the accuracy and robustness of grid classification in long-distance areas.

[0033] (3) This method estimates the possibility of each point being a ground point by calculating the joint probability of its angle probability and height difference probability, which greatly reduces the misjudgment of non-ground structures such as the car hood or roof in the PG grid, making the ground segmentation result more reliable. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a flow chart of the ground point cloud segmentation algorithm for the smart wheelchair described in the present invention;

[0035] Figure 2 The multi-resolution point cloud grid division diagram based on radial intervals and angular intervals described in the present invention;

[0036] Figure 3 This is a diagram of the ground point cloud segmentation system of the smart wheelchair described in the present invention. DETAILED DESCRIPTION

[0037] The present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0038] As shown in the figure, a method for segmenting ground point cloud of an intelligent wheelchair according to the present invention comprises the following steps:

[0039] Step 1: Collect point cloud P through the laser radar on the wheelchair and pre-process the raw point cloud data. Specifically, use a multi-condition screening method based on height and distance to filter the point cloud data. It will reject point clouds with z values ​​lower than the set threshold within a certain distance, thereby effectively reducing outliers.

[0040] Use the multi-resolution partitioning method based on radial interval and angle interval to grid the point cloud:

[0041] Step 2: In the radial direction, according to the distribution characteristics of the point cloud around the sensor, M Z concentric radial regions, each with specific minimum and maximum radial boundaries (L min，m , L max，m ), m = 1, 2, 3, ..., M Z , where L max，m =L min，m+1 , that is, the maximum boundary of each radial area is used as the minimum boundary of the next radial area, and the minimum boundary of the first radial area is the shortest distance of the radar scan, and the Mth radial area is used as the minimum boundary of the next radial area. Z The maximum boundary of the radial area is the maximum distance of the radar scan, and L max，MZ =L max , L min，1 =L min .like Figure 2 In the example given, four concentric radial regions S202, S204, S206, and S208 are divided.

[0042] Step 3: In the angular direction, divide N points according to the point cloud distribution characteristics and requirements. Z Angle interval, such as Figure 1 As shown, each interval has minimum and maximum angle boundaries (θ min，n， θ max，n ), n = 1, 2, 3, ..., N Z , where θ max，n =θ min，n+1 , that is, the maximum boundary of each angle region is used as the minimum boundary of the next angle region. Figure 2 In the example given, four angle areas S210, S212, S214, and S216 are divided.

[0043] Step 4: Further subdivide the grid unit: After the division in steps 2 and 3, we can get M z ×N z grid cells, and define the corresponding grid cell set as:

[0044]

[0045] Where Z m,nRepresents the grid cell of the mth radial interval and the nth angular interval. m,n Defined as: Z m,n ={p k ∈

[0046] P|L min，m ≤ρ k ≤L max，m ,θ min，n ≤θ k ≤θ max，n}

[0047] in For point p k The radial distance, θ k =arctan2(y k / x k ). For each grid cell Z m,n The resolutions in the radial and angular directions are further subdivided to obtain their own subgrid cells. The resolutions in the radial and angular directions of each grid cell are defined as:

[0048]

[0049] Among them, N r,m,n and N θ,m,n The corresponding grid cells Z m,n The number of grid cells in radial and angular directions; each subdivided subgrid cell is represented by R i，j,m,n It means that it contains all points p that meet the following conditions k :

[0050] R i,j,m,n

[0051] ={p k ∈Z m,n |ΔL m,n ·(i-1)≤ρ k -L min，m ≤ΔL m,n ·i, Δθ m,n (j-1)≤θ k -θ min,n ≤Δθ m,n ·j}

[0052] where i = 1, 2, 3, ..., N r,m,n , j = 1, 2, 3, ..., N θ,m,n , m=1,2,3,…,(M Z -1), n=1, 2, 3,..., (N Z -1); Figure 2In the example given, the shaded portion S218 in the figure is obtained by combining S206 and S216, and the shaded portion S220 is obtained by combining S208 and S216. In the radial direction, S218 and S220 are subdivided into 2 and 4 radial grids respectively, and in the angular direction, S218 and S220 are subdivided into 6 and 4 angular grids respectively. It can be seen that the resolutions used by the two are different.

[0053] Step 5: Record sub-grid unit attributes: During the rasterization process, calculate the lowest point height z of each sub-grid unit min , the highest point height z max , height variance and the number of point clouds N are stored as attributes of the sub-grid unit.

[0054] Classify subgrid cells based on height difference and variance information:

[0055] Step 6: Give the classification definition. All sub-grid cells can be classified into grids that are all ground points (GroundCell, G), grids that may contain ground points (Potential Ground Cell, PG) and grids that are all non-ground (Non-Ground Cell, NG). Initialize all sub-grid cells and mark them as category PG.

[0056] Step 7: Calculate the Euclidean distance d from the lowest point of each subgrid unit to the sensor as the distance attribute of the grid.

[0057] d=dis(min(ρ k ))

[0058] Step 8: Update the height attribute H of the subgrid cell according to the lowest point height c , ensure that the height attributes of the subgrid cells are reasonable, specifically:

[0059]

[0060] where h sensor is the sensor height, h min and h max The minimum and maximum heights of the grid cells are set.

[0061] Step 9: Introducing adaptive variance threshold t v : Variance threshold t v It is defined as an adaptive function related to the distance d:

[0062] t v = k·d

[0063] Where k is the distance scaling factor, which scales the distance to the variance threshold and is uniformly applied to all cells.

[0064] Step 10: Mark the sub-grid cells belonging to category G: Specifically, the sub-grid cells that meet the following three conditions are marked as G: (1) The number of point clouds N in the grid exceeds the set minimum number of points N min ; (2) Grid height variance satisfy (3) Grid height H C Less than the threshold h mid The point clouds in all sub-grid cells of category G are estimated as ground points, and their set is represented as

[0065] Step 11: Mark the sub-grid cells belonging to the category NG: Specifically, the sub-grid cells that meet any of the following two conditions are marked as NG: (1) The distance d of the grid cell is greater than the preset maximum distance threshold L max,3 , and the grid height is equal to the sensor height h sensor ; (2) The grid distance d exceeds L max,3 And the height difference Δz=z max -z min Greater than the height difference threshold Δh max , the grid is also considered a non-ground grid.

[0066] The ground point estimation is performed on the sub-grid cells classified as PG based on the Bayesian probability model:

[0067] Step 12: Give a Bayesian probability model: Assume is the likelihood function of the probability model, where Ω represents the parameter set of all models, is the observed point cloud data. Assuming that the point cloud distribution in each sub-grid unit PG is independent of each other, the overall probability of the model can be expressed as:

[0068]

[0069] Among them, ω i and They represent the parameters and observation values ​​of the i-th point in the grid respectively.

[0070] In each PG grid, the probability of each point belonging to the ground is estimated based on the angle and height difference characteristics of each point. Specifically, the point p PG,i The probability of a ground point is given by its angle probability α(θ i ) and the height difference probability β(Δz i ,d i ) joint estimation gives:

[0071]

[0072] Step 13: Calculate the angle probability α(θ i ): We first calculate the normal vector n for each point i , and define the angle between the vertical direction z=[0,0,1] and the ground as θ i , the normal vector of the ground point cloud is often perpendicular to the XY plane of the sensor frame. The calculation method is as follows:

[0073]

[0074] θ τ is the boundary value, that is, θ τ The larger the value, the more conservative the standard. Different boundary values ​​θ can be set in different areas based on experience and point cloud distribution characteristics. τ , ensuring the reliability of its evaluation results.

[0075] Step 14: Give an adaptive height difference threshold function: We define an adaptive height difference threshold κ(d i ) = γ·exp(δ·d i ), which is used to adapt to the height difference requirements at different distances, where γ and δ are constants for controlling the initial value of the threshold and the growth rate.

[0076] Step 15: Calculate the height difference probability β(Δz i ,d i ): Combined with the adaptive height difference threshold function in step 14, the height difference probability is calculated as follows:

[0077]

[0078] Among them, the height difference Δz i Defined as the height z of the current point i The height z of the lowest point in the grid min The difference, d i is the distance attribute of the i-th grid, k(d i ) is the adaptive height difference threshold,

[0079] Step 16: After steps 12 to 15, we can get the estimated ground point clouds in all PG sub-grid units, and their collection is expressed as:

[0080]

[0081] Among them, p PG,i is the i-th point in PG, [·] represents Iversonbracket, returns true when the condition is met, and divides the corresponding point into In the end, Same as in step 10 By merging, we can get the estimated ground point cloud output by our ground point cloud segmentation algorithm.

[0082] It should be noted that the above content only illustrates the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications all fall within the protection scope of the claims of the present invention.

Claims

1. A method for segmenting ground point cloud of an intelligent wheelchair, characterized in that: The method comprises the following steps: S1, by installing it on a wheelchair with a height of h sensor The laser radar collects point clouds, pre-processes the raw point cloud data, and filters the point cloud data according to the preset distance and height thresholds; S2, using a multi-resolution partitioning method based on radial intervals and angle intervals to perform grid partitioning on the point cloud to obtain sub-grid units, wherein the partitioning method uses different resolutions to partition the point clouds in different radial intervals and angle intervals to adapt to the sparsity of the point cloud, and calculates and stores the attributes of each sub-grid unit during the rasterization process; S3, classifying the sub-grid units based on the height difference and variance information into three categories: G, NG, and PG, wherein the G category is a grid that is entirely ground points; the NG category is a grid that is entirely non-ground points; and the PG category is a grid that may contain ground points; S4. Calculate the probability of ground points based on the Bayesian probability model, estimate the point cloud in the PG grid to obtain the ground points therein, merge all the ground points, obtain the estimated ground point cloud and output it; the merging method is to take the union of the ground points of category G and the ground points obtained by Bayesian estimation.

2. The method for ground point cloud segmentation of an intelligent wheelchair according to claim 1, characterized in that: The grid division method described in step S2 specifically includes: S21: In the radial direction, according to the distribution characteristics of the point cloud around the sensor, M Z concentric radial regions, each with a specific minimum radial boundary L min,m and the maximum radial boundary L max,m ,m=1,2,3,…,M Z , where L max,m =L min,m+1 , that is, the maximum boundary of each radial area is used as the minimum boundary of the next radial area, and the minimum boundary of the first radial area is the shortest distance of the radar scan, and the Mth radial area is used as the minimum boundary of the next radial area. Z The maximum boundary of the radial area is the maximum distance scanned by the radar. L min,1 =L min ; S22: In the angle direction, N points are divided according to the point cloud distribution characteristics and requirements. Z angle intervals, each interval has a minimum angle boundary θ min,n , and the maximum angle boundary θ max,n ,n=1,2,3,…,N Z , where θ max,n =θ min,n+1 , that is, the maximum boundary of each angle region is used as the minimum boundary of the next angle region; the radial region and the angle interval are combined to obtain M z ×N z grid cells, and define the corresponding grid cell set as: Where Z m,n Represents the grid cell of the mth radial region and the nth angular interval; Z m,n Defined as: Z m,n ={p k ∈P|L min,m ≤ρ k ≤L max,m ,θ min,n ≤θ k ≤θ max,n } in For point p k The radial distance, θ k =arctan2(y k / x k ); S23: For each grid cell Z m,n The resolutions in the radial and angular directions are further subdivided to obtain their own subgrid cells. The resolutions in the radial and angular directions of each grid cell are defined as: Among them, N r,m,n and N θ,m,n The corresponding grid cells Z m,n The number of grid cells in radial and angular directions; each subdivided subgrid cell is represented by R i,j,m,n It means that it contains all points p that meet the following conditions k : R i,j,m,n =[p k ∈Z m,n |ΔL m,n ·(i-1)≤ρ k -L min,m ≤ΔL m,n ·i,Δθ m,n ·(j-1)≤θ k -θ min,n ≤Δθ m,n ·j} Among them, i=1,2,3,…,N r,m,n , j = 1, 2, 3, …, N θ,m,n , m = 1, 2, 3, …, (M Z -1), n ​​= 1, 2, 3, ..., (N Z -1); S24: During the rasterization process, calculate the lowest point height z of each sub-grid unit min , the highest point height z max , height variance and the number of point clouds N are stored as attributes of the sub-grid unit.

3. The method for ground point cloud segmentation of an intelligent wheelchair according to claim 1, characterized in that: The classification method described in step S3 specifically includes: S31: Sub-grid unit initialization; the sub-grid units are first marked as PG category; S32: Calculate the Euclidean distance d from the lowest point of each sub-grid unit to the sensor as the distance attribute of the grid; d=dis(min(ρ k )) S33: Update the height attribute H of the subgrid unit according to the lowest point height C , ensure that the height attributes of the subgrid cells are reasonable, specifically: where h sensor is the sensor height, h min and h max The minimum and maximum heights of the grid cells are set; S34: Mark the sub-grid cells belonging to category G; specifically, mark the sub-grid cells that meet the following three conditions as G: (1) The number of point clouds N in the grid exceeds the set minimum number of points N min ; (2) Grid height variance satisfy (3) Grid height H C Less than the threshold h mid ; The point clouds in all sub-grid cells of category G are estimated as ground points, and their set is represented as where t v is the adaptive variance threshold, t v = k·d, where k is the distance scaling factor and d is the grid distance attribute; S35: Mark the sub-grid cells belonging to the category NG; specifically, mark the sub-grid cells that meet any of the following two conditions as NG: (1) The distance d of the grid cell is greater than the preset maximum distance threshold L max,3 , and the grid height is equal to the sensor height h sensor ; (2) The grid distance d exceeds L max,3 And the height difference Δz=z max -z min Greater than the height difference threshold Δh max , the grid is also considered a non-ground grid.

4. The method for ground point cloud segmentation of an intelligent wheelchair according to claim 1, characterized in that: The calculation of ground point probability in step S4 specifically includes: S41: Calculate the angle probability α(θ i ); Specifically, first calculate the normal vector n of each point i , and define the angle between the vertical direction z=[0,0,1] and the ground as θ i , the normal vector of the ground point cloud is often perpendicular to the XY plane of the sensor frame; the angle probability is calculated as follows: where θ τ is the angle boundary value, and different boundary values ​​θ are set in different areas τ , to ensure the reliability of its evaluation results; S42: Calculate the height difference probability β(Δz based on the height difference and distance attributes i ,d i ), calculated as follows: Among them, the height difference Δz i Defined as the height z of the current point i The height z of the lowest point in the grid min The difference, d i is the distance attribute of the i-th grid, κ(d i ) is the adaptive height difference threshold, and its value is calculated in an exponential form: k(d i )=γ·exp(δ·d i ) Where γ and δ are constants that control the initial value of the threshold and the growth rate; S43: According to the angle probability α(θ i ) and the height difference probability β(Δz i ,d i ), jointly estimate the ground point probability For example, point p PG,i If the ground point probability of is greater than the preset probability threshold, the point is estimated to be a ground point, and the corresponding ground point set is expressed as Among them, p PG,i is the i-th point in PG, [·] represents Iversonbracket, returns true when the condition is met, and divides the corresponding point into middle; S44: Output all ground point clouds Specifically, For S43 Compared with S34 The union of; 5. An intelligent wheelchair ground point cloud segmentation system, characterized by: include: LiDAR scanning module, used to collect 3D point cloud information during the movement of the wheelchair; The original point cloud preprocessing module is used to filter the collected point cloud; The multi-resolution grid division module is used to divide point clouds in different radial and angular intervals with differentiated resolutions to improve ground segmentation accuracy; The grid attribute storage module is used to calculate the lowest point, highest point, height variance and point cloud quantity of the grid when dividing the grid, and store them as grid attributes; The grid classification module is used to classify the subdivided subgrid units, first find the grids that are all ground points, and reduce the number of ground points that need to be estimated later; The Bayesian probability ground point estimation module is used to perform probability estimation on the potential ground point grids classified in the grid classification module to obtain the ground points therein; The estimated ground point merging module is used to merge all estimated ground points to obtain an estimated ground point cloud set.

Images