Stair environment lightweight map construction method for unmanned platform
By extracting the feasible area and Bayesian classifier labeling technology, the problem of large amount of map data and low utilization in ground unmanned platform navigation is solved, navigation efficiency and accuracy are improved, and a clear and accurate stair environment is built. map.
Patent Information
- Application Number
- CN202411121154.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-15
- Publication Date
- 2024-11-26
AI Technical Summary
The existing technology is difficult to effectively distinguish special terrains such as stairs without relying on visual images or manual annotations, resulting in large amounts of map data, low utilization rate, and low path planning efficiency in ground unmanned platform navigation. .
Acquire point cloud data inside the building through lidar and inertial navigation, extract the travelable areas, remove discrete point clouds, establish multi-layer raster maps, and use Bayesian classifiers to label the terrain to reduce redundant information and improve Path search efficiency.
It significantly reduces the redundant information in the map data, improves the path search efficiency and navigation accuracy of ground unmanned platform navigation, avoids searching in unmovable areas, and builds a stair area map with more accurate map boundaries and clear map boundaries.
Smart Images

Figure CN119022914A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of map construction for unmanned platform navigation, and in particular relates to a lightweight map construction method for a stair environment of an unmanned platform. Background Art
[0002] Building an environmental map is a key prerequisite for unmanned platform navigation in buildings. A complete and detailed map can effectively assist unmanned platforms in planning the best movement path. However, building a map faces the problem of balancing resolution and storage space. High-resolution maps can improve the accuracy of navigation and obstacle avoidance, but they consume a lot of computing resources and memory.
[0003] For complex interior environments of buildings, the complexity of map environment modeling increases with the dense composition of large-area planes such as walls, floors, and ceilings. Traditional map construction methods require a large number of points, voxels, and three-dimensional surface grids to describe these planes, so the amount of map data is much larger than other environmental data. This makes map loading and updating difficult, and also places higher requirements on the storage and processing capabilities of the device.
[0004] Although the octree map reduces the storage space occupied by the map by dividing the entire environment into hierarchical three-dimensional sub-regions, its data access efficiency is not as efficient as other map models. In addition, ground unmanned platforms pay more attention to the situation of drivable planes when planning paths, so the octree map is more suitable for aircraft path planning, but not for ground unmanned platforms.
[0005] In addition, the path planning of ground unmanned platforms only requires information about the drivable area on the horizontal plane. Unnecessary information such as ceilings and walls does not contribute to path planning, but consumes a lot of storage space. This results in a small proportion of data available for navigation in the overall map data, resulting in serious waste of resources.
[0006] In addition, most of the current domestic ground unmanned platform path planning decomposes the navigation in the building into the type of finite state machine to execute, and adopts different path planning methods according to different terrains such as planes and stairs. The premise of adopting this path planning method is to obtain the terrain classification labels of the map used for path planning, but it is currently difficult to distinguish special terrains such as stairs and turning platforms only by relying on the point cloud map established by laser without relying on visual images and manual annotation. In addition, the method relying on visual images is greatly affected by lighting conditions, and the method relying on manual annotation is too inefficient.
[0007] Therefore, it is of great practical significance to develop a lightweight map construction method for stair environment with small data volume, high map data utilization, clear boundaries, distinct layers and terrain labels. Summary of the invention
[0008] In order to solve the technical problems of large amount of map data required for navigation of ground unmanned platforms in buildings, low data utilization rate and difficulty in direct classification, the present invention provides a lightweight map construction method for stair environment for unmanned platforms. By extracting the drivable area, the redundant information in the map data is significantly reduced, the path search efficiency and the accuracy of navigation of the ground unmanned platform in this area are improved.
[0009] A method for constructing a lightweight map of a stair environment for an unmanned platform comprises the following steps:
[0010] S1: Construct a point cloud map based on the internal point cloud data of the building obtained by LiDAR and inertial navigation;
[0011] S2: Extract the drivable area of the ground unmanned platform from the point cloud map and remove the discrete point cloud in the drivable area; establish an initial multi-layer grid map based on the point cloud map after removing the discrete point cloud, then perform map edge smoothing on the initial multi-layer grid map, and remove the data representing the ceiling part in the smoothed multi-layer grid map to obtain the final multi-layer grid map;
[0012] S3: Input the features representing the stair environment terrain in the multi-layer grid map into the Bayesian classifier for terrain classification, and obtain a multi-layer grid map with terrain labels, wherein the terrain labels include steps, stair turning platforms, and floor plans.
[0013] Furthermore, the method for extracting the drivable area of the ground unmanned platform from the point cloud map in step S2 is:
[0014] Each point in the point cloud map is used as the current point A to perform a drivable judgment operation in turn, and all points belonging to the drivable area are screened out, wherein the drivable judgment operation is as follows:
[0015] Extract the N neighboring points closest to the current point A to build a neighborhood set N;
[0016] Fit all point clouds in the neighborhood set N to a plane e, and determine whether the normal vector m of the plane e is perpendicular to the horizontal plane. If yes, the current point cloud A belongs to the point cloud of the drivable area. If no, the current point cloud A does not belong to the point cloud of the drivable area.
[0017] Furthermore, the method for removing the discrete point cloud in the drivable area in step S2 is as follows:
[0018] Different values are taken for the proportional coefficient ε in (0,1) according to the set step size, and a ratio P acquisition operation is performed for each value ε, wherein the ratio P acquisition operation is as follows:
[0019] Calculate the threshold N according to the proportionality coefficient ε at the current value neighborhood :
[0020]
[0021] where d surface is the point cloud surface density in the drivable area, R is the radius of the spherical area constructed with each point cloud in the drivable area as the center of the sphere, and R < min{D, H}, D is the step length of the step, and H is the step height;
[0022] Construct spherical areas with a radius of R with each point cloud in the drivable area as the center of the sphere, and respectively judge the number of point clouds N number contained in each spherical area. Whether it is less than N neighborhood . The point cloud with a judgment result of yes is the discrete point cloud, and at the same time, the discrete point cloud is removed. The point cloud with a judgment result of no is retained;
[0023] Take the total number N number of the point clouds belonging to the stairs in the point clouds not less than N neighborhood and the total number N StairNum of the point clouds not less than N number The ratio between them is used as P; neighborhood The total number N TotalNum ;
[0024] Take the value of the proportionality coefficient ε corresponding to the maximum value of the ratio P without residual discrete point clouds as the optimal value, and remove the discrete point clouds in the drivable area under the optimal value to obtain a point cloud map after removing the discrete point clouds.
[0025] Furthermore, the method for establishing the initial multi-layer grid map according to the point cloud map after removing the discrete point clouds in step S2 is as follows:
[0026] Divide the point clouds in the point cloud map into different levels according to height, and count the number of point clouds in different levels. Select all the height values corresponding to the levels with the number of point clouds greater than the set threshold δ1 as the grid heights corresponding to each layer of the initial multi-layer grid map;
[0027] Integrate the point clouds at each layer grid height into a point cloud horizontal plane set according to the set height resolution S z . Among them, the value of the height resolution S z satisfies: making each layer grid height correspond to a point cloud horizontal plane set and any point cloud only falls into one of the point cloud horizontal plane sets;
[0028] According to the set x-axis resolution S x of the horizontal plane and the y-axis resolution S yDivide each horizontal plane point cloud set into multiple grids, and count the number of point clouds in each grid on each horizontal plane point cloud set;
[0029] Determine whether the number of point clouds in all grids is greater than the set threshold δ2. The grid area where the judgment result is yes is the drivable plane. All drivable planes at each grid height constitute the grid map at each grid height. The grid maps at all grid heights are combined to obtain the initial multi-layer grid map.
[0030] Furthermore, the method for performing map edge smoothing processing on the initial multi-layer grid map in step S2 is:
[0031] Use the 3×3 convolution kernel K to perform convolution processing on each layer of the multi-layer grid map to obtain a smoothed multi-layer grid map, where the convolution kernel K is:
[0032]
[0033] 6. The method for constructing a lightweight map of a staircase environment for an unmanned platform as claimed in claim 1, wherein the method for removing the data representing the ceiling part in the smoothed multi-layer grid map in step S2 is:
[0034] Use each layer of the multi-layer grid map as the current layer grid map and perform the following operations:
[0035] Determine whether there are other grids with a grid value of 1 within the set distance range [δ3,δ4] below each grid on the current grid map, and delete the grids with a judgment result of yes; where δ3 and δ4 are the lower and upper limits of the set distance range respectively; a grid value of 1 indicates that there is a solid surface at the corresponding position of the grid inside the building;
[0036] It is determined whether the number of grids deleted on the current grid map is greater than a threshold value δ5. If so, the current grid map is data corresponding to the ceiling part, which is deleted.
[0037] Further, the features representing the stair environment terrain in the multi-layer grid map in step S3 include: the length of the plane rectangle formed by the grid with a value of 1 in each layer of grid map, the width of the plane rectangle formed by the grid with a value of 1 in each layer of grid map, the x-axis projection coordinate of the center point of the plane rectangle formed by the grid with a value of 1 in each layer of grid map, the y-axis projection coordinate difference of the center point of the plane rectangle formed by the grid with a value of 1 in two adjacent layers of grid map, the area of the solid surface corresponding to each layer of grid map in the interior of the building, and the aspect ratio of the shape of the solid surface corresponding to each layer of grid map in the interior of the building; wherein, a grid value of 1 indicates that a solid surface exists at the corresponding position of the grid in the building;
[0038] If any layer of grid maps satisfies the following first joint decision condition, the terrain label corresponding to the grid map of this layer is a step. The first joint decision condition is: the length of the plane rectangle formed by the grid with a value of 1 in the grid map of this layer is the length of the plane rectangle of the step and the width of the plane rectangle formed is the plane width of the step, and the difference between the solid surface area corresponding to the grid map of this layer in the building and the area of the single-layer step is less than a first set threshold;
[0039] If a continuous multi-layer grid map satisfies the following second joint decision condition, then the continuous multi-layer grid map is a staircase, and the joint decision condition is: the length of the plane rectangle formed by the grid with a value of 1 in each layer of the grid map is the length of the plane rectangle of the step and the width of the plane rectangle formed is the width of the step, the x-axis projection coordinates of the center points of the plane rectangles formed by the grid with a value of 1 in each layer of the grid map are the same, and the y-axis projection coordinate differences of the center points of the plane rectangles formed by the grids with a value of 1 in any two adjacent layers of the grid map are the same;
[0040] If the difference between the physical surface area corresponding to any layer of the grid map in the building and the area of the stair turning platform is less than the second set threshold and does not meet the first joint judgment condition, then the terrain label corresponding to the grid map of this layer is the stair turning platform;
[0041] If the difference between the physical surface area corresponding to any layer of the grid map inside the building and the area of the stair turning platform is not less than the second set threshold and does not meet the first joint judgment condition, the terrain label corresponding to the grid map of this layer is the floor plan.
[0042] Furthermore, the method of inputting the features representing the stair environment terrain in the multi-layer grid map into the Bayesian classifier for terrain classification is specifically as follows:
[0043] The rules for obtaining the stair environment terrain are as follows:
[0044] Rule (1): The length and width of the rectangles on each step plane are the same;
[0045] Rule (2): For each flight of stairs, the x-axis projection coordinates of the center point of the plane rectangle of each step are the same;
[0046] Rule (3): For each section of stairs, the difference between the y-axis projection coordinates of the center points of the plane rectangles of any two adjacent steps is the same, and the y-axis projection coordinates of the center point of the plane rectangle of each step are within the range of the y-axis projection coordinates of the section of stairs;
[0047] Rule (4): For each flight of stairs, except for the first and last two steps, the upper and lower adjacent steps of each step are also steps, and they all satisfy rules (1), (2), and (3);
[0048] Obtain the feature vectors of the features of the stair environment terrain of each layer of the multi-layer grid map. If any feature vector meets the rule (1), rule (2), and rule (4) at the same time, the feature value corresponding to the feature vector is α1. If the feature vector does not meet the rule (1), rule (2), and rule (4) at the same time, the feature value corresponding to the feature vector is α2.
[0049] If any eigenvector conforms to rule (1), rule (3), and rule (4) at the same time, the eigenvalue 2 corresponding to the eigenvector is β1. If the eigenvector does not conform to rule (1), rule (3), and rule (4) at the same time, the eigenvalue 2 corresponding to the eigenvector is β2.
[0050] If the difference between the solid surface area corresponding to any eigenvector in the building and the single-layer step area is less than the first set threshold, the eigenvalue three corresponding to the eigenvector is η1; if the difference between the solid surface area corresponding to the eigenvector in the building and the staircase turning platform area is less than the second set threshold, the eigenvalue three corresponding to the eigenvector is η2; otherwise, the eigenvalue three corresponding to the eigenvector is η3;
[0051] If the difference between the aspect ratio of the solid surface shape corresponding to any eigenvector in the building and the aspect ratio of the single-layer step shape is less than the third set threshold, the eigenvalue four corresponding to the eigenvector is μ1. If the difference between the aspect ratio of the solid surface shape corresponding to the eigenvector in the building and the aspect ratio of the stair turning platform shape is less than the fourth set threshold, the eigenvalue four corresponding to the eigenvector is μ2. Otherwise, the eigenvalue four corresponding to the eigenvector is μ x ;
[0052] According to the above rules, the features of the training sample set X composed of all layers of raster maps are extracted to obtain the training sample data set T = {(x1, y1), (x2, y2), … (x N ,y N )}, N is the total number of grid map layers in the training sample, where represents the jth eigenvalue of the i-th sample formed by the i-th layer grid map, j = 1, 2, 3, 4, a jl is the lth possible value of the jth feature, l = 1, 2, ... S j , S j is the number of possible values of the j-th eigenvalue;
[0053] in, Y represents the terrain label set of the training set samples, and Y = {y1, y2, ..., y i,...,y N},y i Represents the terrain label value of the i-th sample formed by the i-th layer of the raster map;
[0054] According to the training sample data set T, calculate the prior probability P(Y=c k ):
[0055]
[0056] Where k is the category number of the terrain label, and k = 1 means that the terrain label is a step, k = 2 means that the terrain label is a staircase turning platform, k = 3 means that the terrain label is a floor plane, λ is the Laplace smoothing coefficient, c k Indicates that the terrain label is the kth category, Represents the terrain label value y in the statistical training sample set i Classified as c k The total number of grid map samples, K represents the number of classifications;
[0057] According to the training sample data set T, calculate the conditional probability P λ (X (j) =a jl |Y=c k ):
[0058]
[0059] Among them, the conditional probability P λ (X (j) =a jl |Y=c k ) means that when the label of the sample in the training sample set X is c k When the jth eigenvalue of the sample is a jl probability; Indicates that the terrain label value in all training sample sets X is c k And the selection of the jth eigenvalue is the number of samples of the lth value;
[0060] The following classification operations are performed on each layer of the constructed multi-layer grid map to be classified:
[0061] The prior probability P(Y=c k ) and conditional probability P λ (X (j) =a jl |Y=c k ), and at the same time extract the characteristic value x of each sample x to be classified formed by each layer of the grid map to be classified = (x (1) ,x (2) ,x (3) ,x(4) ) T , and then calculate the posterior probability value P (Y = c k |X=x):
[0062]
[0063] For each sample to be classified formed by each layer of grid maps to be classified, the class with the largest posterior probability is taken as the classification y of each sample to be classified x, that is, the terrain classification of the grid map of this layer:
[0064]
[0065] Among them, max is the maximum value.
[0066] Beneficial effects:
[0067] 1. The present invention provides a method for constructing a lightweight map of a stair environment for an unmanned platform. First, point cloud data established by a laser radar and an inertial navigation system are input; then, the area where the ground unmanned platform can navigate is extracted from the input point cloud map, and discrete points are removed. A multi-layer grid map is established based on the processed point cloud map, and map edge smoothing is performed and data representing the ceiling part in the multi-layer grid map is removed; then, features representing the terrain of the stair environment are extracted from the generated multi-layer grid map, and the multi-layer grid map is classified using a naive Bayes classification method with Laplace smoothing; finally, a multi-layer grid map with terrain labels is output; it can be seen that the present invention significantly reduces redundant information in map data by extracting a drivable area, realizes lightweight map construction, and improves the utilization rate of map data when the ground unmanned platform navigates and searches for paths, avoids searching in non-drivable areas such as walls, and improves path search efficiency.
[0068] 2. The present invention provides a lightweight map construction method for stair environments for unmanned platforms. According to the characteristics of the stairs, rules are formulated to remove residual discrete points, so that the point cloud representing the stair area is retained to the greatest extent. The constructed stair area map is more accurate, which improves the accuracy of navigation of ground unmanned platforms in this area.
[0069] 3. The present invention provides a lightweight map construction method for staircase environments for unmanned platforms. The map edge smoothing operation is adopted to improve the situation that the constructed map has unclear boundaries, irregular shapes, and small internal defects caused by the randomness of laser radar sampling, so that the constructed map is closer to the physical surface it represents, and the map has a regular shape and clear boundaries.
[0070] 4. The present invention provides a lightweight map construction method for stair environments for unmanned platforms. By extracting map features and using the naive Bayesian classification method, different terrains are distinguished to provide more practical and intelligent environmental information for ground unmanned platform navigation. BRIEF DESCRIPTION OF THE DRAWINGS
[0071] Figure 1 This is the overall flow chart of the lightweight map construction method for stair environments for unmanned platforms;
[0072] Figure 2 Schematic diagram for removing discrete point size calculation;
[0073] Figure 3 Schematic diagram of edge smoothing convolution;
[0074] Figure 4 This is a schematic diagram of a multi-layer grid map;
[0075] Figure 5 A schematic diagram of a grid map showing the edge of the ground;
[0076] Figure 6 This is a schematic diagram of the statistical results of the ground edge grid map in the x direction;
[0077] Figure 7 This is a schematic diagram of the statistical results of the ground edge grid map in the y direction;
[0078] Figure 8 A point cloud diagram showing the staircase environment in a building;
[0079] Fig. 9 This is a schematic diagram of the point cloud after extracting the executable area;
[0080] Fig.10 Schematic diagram of the relationship curve between ε and P. DETAILED DESCRIPTION
[0081] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.
[0082] like Figure 1 As shown, a lightweight map construction method for stair environment for unmanned platform includes the following steps:
[0083] S1: The point cloud map formed by the internal point cloud data of the building obtained by LiDAR and inertial navigation is used as the input data of the lightweight map construction method of the stair environment for unmanned platforms;
[0084] For example, the specific implementation of step S1 is as follows:
[0085] Input a dense point cloud PointCloud of a two-story floor and stairwell environment original :
[0086] PointCloud original = {P1,P2,P3,…P Q}
[0087] Q is the number of points in the set, such as Figure 8 As shown, here Q is 4229468.
[0088] S2: Extract the drivable area of the ground unmanned platform from the point cloud map and remove the discrete point cloud in the drivable area; establish an initial multi-layer grid map based on the point cloud map after removing the discrete point cloud, then perform map edge smoothing on the initial multi-layer grid map, and remove the data representing the ceiling part in the smoothed multi-layer grid map to obtain the final multi-layer grid map;
[0089] S3: Input the features representing the stair environment terrain in the multi-layer grid map into the Bayesian classifier for terrain classification, and obtain a multi-layer grid map with terrain labels, wherein the terrain labels include steps, stair turning platforms, and floor plans.
[0090] The operation process of step S2 is described in detail below.
[0091] In the first aspect, the method for extracting the drivable area of the ground unmanned platform from the point cloud map in step S2 is:
[0092] Each point in the point cloud map is used as the current point A to perform a drivable judgment operation in turn, and all points belonging to the drivable area are screened out, wherein the drivable judgment operation is as follows:
[0093] Extract the N neighboring points closest to the current point A to build a neighborhood set N:
[0094] N={p s (x s ,y s ,z s )|s=1,2,…,n}
[0095] For example, Figure 8 The point cloud map PointCloud before the drivability judgment operation is shown original , for PointCloud original For all points A in the tree, search for the coordinates of the s nearest points to form the neighborhood set N of the point, where s is 629230. It should be noted that the KD-tree method can be used to search for the nearest point set of a certain point, or other similar methods can be selected, such as Ball Tree, VP-tree, etc.
[0096] Fit all the point clouds in the neighborhood set N to a plane e, and determine whether the normal vector m of the plane e is perpendicular to the horizontal plane. If so, the current point cloud A belongs to the point cloud of the drivable area. If not, the current point cloud A does not belong to the point cloud of the drivable area.
[0097] That is to say, the present invention fits all the points in the neighborhood N to an optimal plane e, and uses the normal vector m of the plane to represent the inclination of the infinitesimal plane where point A is located in space. For PointCloud original Solve the normal vector for all the points in it. According to the motion ability limit of the ground unmanned platform, by restricting the direction of the normal vector m of each point, remove the points in non-drivable areas such as vertical and inclined areas to obtain PointCloud extracted . Figure 8 Is the point cloud map before extracting the drivable area of the ground unmanned platform, Fig. 9 Is the point cloud map after extraction.
[0098] In the second aspect, the method for removing discrete point clouds in the drivable area in step S2 is as follows:
[0099] Take different values for the proportionality coefficient ε in (0, 1) according to the set step size, and perform the ratio P obtaining operation for each value of ε. Among them, the ratio P obtaining operation is as follows:
[0100] Calculate the threshold N according to the proportionality coefficient ε under the current value neighborhood :
[0101]
[0102] Among them, d surface Is the point cloud surface density in the drivable area, R is the radius of the spherical area constructed with each point cloud in the drivable area as the center of the sphere, and R < min{D, H}, D is the step tread length, and H is the step height; for example, as Figure 2 Shown, take the radius R of the spherical judgment domain as 0.1m and 0.15m respectively, the step tread length D of the stairs is 0.25m, and the step height H is 0.16m. For each point in PointCloud extracted , count the number N of the surrounding points within the spherical judgment domain of this point number .
[0103] Construct spherical areas with a radius of R with each point cloud in the drivable area as the center of the sphere, and respectively judge whether the number of point clouds N number Contained in each spherical area is less than N neighborhood . The point cloud with the judgment result being yes is the discrete point cloud, and at the same time, the discrete point cloud is removed, and the point cloud with the judgment result being no is retained;
[0104] N number Not less than N neighborhood The total number of point clouds belonging to stairs in the point cloud N StairNum With N number Not less than N neighborhood The total number of point clouds N TotalNum The ratio between
[0105] The value of the proportional coefficient ε corresponding to the maximum value of the ratio P when there is no residual discrete point cloud is taken as the optimal value, and the discrete point cloud in the drivable area under the optimal value is eliminated to obtain the point cloud map after removing the discrete point cloud.
[0106] For example, the present invention takes R as 0.1m and 0.15m respectively to conduct experiments and obtain the relationship between ε and P, as shown in FIG. Fig.10 As shown, according to the experimental results, ε is taken as 0.2.
[0107] In a third aspect, the method for establishing an initial multi-layer grid map according to the point cloud map after removing the discrete point cloud in step S2 is:
[0108] Divide each point cloud in the point cloud map into different levels according to the height, count the number of point clouds in different levels, and select the height values corresponding to all levels where the number of point clouds is greater than the set threshold δ1 as the grid heights corresponding to each grid layer of the initial multi-layer grid map;
[0109] According to the set height resolution S z The point clouds at each layer of grid height are integrated into a point cloud horizontal plane set, where the height resolution S z The value of satisfies: each layer of grid height corresponds to a point cloud horizontal plane set and any point cloud only falls into one of the point cloud horizontal plane sets;
[0110] According to the set horizontal x-axis resolution S x , y-axis resolution S y Divide each horizontal plane point cloud set into multiple grids, and count the number of point clouds in each grid on each horizontal plane point cloud set;
[0111] Determine whether the number of point clouds in all grids is greater than the set threshold δ2. The grid area where the judgment result is yes is the drivable plane. All drivable planes at each grid height constitute the grid map at each grid height. The grid maps at all grid heights are combined to obtain the initial multi-layer grid map.
[0112] That is to say, the present invention divides the point cloud data into different levels according to the height distribution, counts the number of point clouds at each height, selects the height with more clustered points (that is, the height value corresponding to the number of clustered points greater than the threshold δ1) as the height of each layer of the grid map, and then calculates the height according to the height resolution S z Integrate point clouds into different point cloud level sets N c Optional, take δ1 as 0.01m, S z It is 0.03m.
[0113] Next, according to the horizontal resolution S x , S y The point cloud horizontal plane set N of each plane c Split into subsets G stored in the raster at this resolution c , and record the number of points in each grid. Optionally, take S x 0.05m, S y It is 0.05m.
[0114] If the number of points in the grid is greater than the threshold δ2, the area represented by the grid is considered to be a drivable plane, and a grid map M representing each plane is established. c (here δ2 is 50), the multi-layer grid map MGM is M c Optionally, δ2 is 50.
[0115] In a fourth aspect, the method for performing map edge smoothing on the initial multi-layer grid map in step S2 is:
[0116] Use the 3×3 convolution kernel K to perform convolution processing on each layer of the multi-layer grid map to obtain a smoothed multi-layer grid map, such as Figure 3 Medium F i As shown, the convolution kernel K is:
[0117]
[0118] Fifthly, it should be noted that according to the building specifications, the horizontal coordinates of the ceiling plane of each floor coincide with the horizontal coordinates of the ground plane and the height difference is fixed. The map area representing the ceiling is removed by setting the following conditions. Specifically, the method for removing the data representing the ceiling part in the smoothed multi-layer grid map in step S2 is:
[0119] Use each layer of the multi-layer grid map as the current layer grid map and perform the following operations:
[0120] Determine whether there are other grids with a grid value of 1 within the set distance range [δ3, δ4] below each grid on the current grid map, and delete the grids with a judgment result of yes; wherein δ3 and δ4 are the lower limit and upper limit of the set distance range respectively, and the values of δ3 and δ4 refer to the distance between the ceiling and the floor in general building specifications; a grid value of 1 indicates that there is a solid surface at the corresponding position of the grid inside the building; optionally, δ3 is 2m and δ4 is 3.3m.
[0121] Determine whether the number of grids deleted on the current grid map is greater than the threshold δ5, where δ5 is 50. If yes, the current grid map is the data corresponding to the ceiling part, which is removed to complete the map construction algorithm and obtain a multi-layer grid map, such as Figure 4 shown.
[0122] The operation process of step S3 is described in detail below.
[0123] It should be noted that the present invention converts the image edge extraction Laplace operator into a 3×3 convolution template K laplace , and convolve the operator template with the multi-layer grid map to obtain a grid map representing the edge of the ground, such as Figure 5 shown.
[0124]
[0125] It should be noted that the Laplace operator may also be replaced by other operators having image edge extraction functions, such as the Sobel operator, the Prewitt operator, the Canny operator, and the like.
[0126] Depend on Figure 5 It can be seen that the plane distribution of the staircase part has a regular pattern. Figure 5 Taking the bottom flight of stairs as an example, the following rules can be drawn:
[0127] (1) The length and width of the rectangle representing each step plane are approximately the same, such as Figure 5 As shown, the length of the plane rectangle of each step in this section is a1a3, and the width is b1b2.
[0128] (2) For each section of stairs, the coordinates of the center points of the plane rectangles of each step coincide in the X direction and are a2.
[0129] (3) For each section of stairs, the difference between the coordinate value of the center point of each step rectangle on the Y axis and the coordinate value of the center point of its adjacent step rectangle on the Y axis is a fixed value, the length of the difference is b1b2, and the center point of each step rectangle is distributed within the range of b1b3 in the Y direction.
[0130] (4) For each flight of stairs, except for the first and last two steps, the steps above and below each step are also steps, and they all have the characteristics of rules (1), (2), and (3).
[0131] Based on this, it can be concluded that the features representing the stair environment terrain in the multi-layer grid map in step S3 include: the length of the plane rectangle formed by the grid with a value of 1 in each layer of the grid map, the width of the plane rectangle formed by the grid with a value of 1 in each layer of the grid map, the x-axis projection coordinate of the center point of the plane rectangle formed by the grid with a value of 1 in each layer of the grid map, the y-axis projection coordinate difference of the center point of the plane rectangle formed by the grid with a value of 1 in two adjacent layers of the grid map, the area of the solid surface corresponding to each layer of the grid map in the interior of the building, and the aspect ratio of the shape of the solid surface corresponding to each layer of the grid map in the interior of the building; wherein a grid value of 1 indicates that there is a solid surface at the corresponding position of the grid inside the building;
[0132] If any layer of grid maps satisfies the following first joint decision condition, the terrain label corresponding to the grid map of this layer is a step. The first joint decision condition is: the length of the plane rectangle formed by the grid with a value of 1 in the grid map of this layer is the length of the plane rectangle of the step and the width of the plane rectangle formed is the plane width of the step, and the difference between the solid surface area corresponding to the grid map of this layer in the building and the area of the single-layer step is less than a first set threshold;
[0133] If a continuous multi-layer grid map satisfies the following second joint decision condition, then the continuous multi-layer grid map is a staircase, and the joint decision condition is: the length of the plane rectangle formed by the grid with a value of 1 in each layer of the grid map is the length of the plane rectangle of the step and the width of the plane rectangle formed is the width of the step, the x-axis projection coordinates of the center points of the plane rectangles formed by the grid with a value of 1 in each layer of the grid map are the same, and the y-axis projection coordinate differences of the center points of the plane rectangles formed by the grids with a value of 1 in any two adjacent layers of the grid map are the same;
[0134] If the difference between the physical surface area corresponding to any layer of the grid map in the building and the area of the stair turning platform is less than the second set threshold and does not meet the first joint judgment condition, then the terrain label corresponding to the grid map of this layer is the stair turning platform;
[0135] If the difference between the physical surface area corresponding to any layer of the grid map inside the building and the area of the stair turning platform is not less than the second set threshold and does not meet the first joint judgment condition, the terrain label corresponding to the grid map of this layer is the floor plan.
[0136] Furthermore, the method of inputting the features representing the stair environment terrain in the multi-layer grid map into the Bayesian classifier for terrain classification is specifically as follows:
[0137] The rules for obtaining the stair environment terrain are as follows:
[0138] Rule (1): The length and width of the rectangles on each step plane are the same;
[0139] Rule (2): For each flight of stairs, the x-axis projection coordinates of the center point of the plane rectangle of each step are the same;
[0140] Rule (3): For each section of stairs, the difference between the y-axis projection coordinates of the center points of the plane rectangles of any two adjacent steps is the same, and the y-axis projection coordinates of the center point of the plane rectangle of each step are within the range of the y-axis projection coordinates of the section of stairs;
[0141] Rule (4): For each flight of stairs, except for the first and last two steps, the upper and lower adjacent steps of each step are also steps, and they all satisfy rules (1), (2), and (3);
[0142] Obtain the feature vectors of the features of the stair environment terrain of each layer of the multi-layer grid map. If any feature vector meets the rule (1), rule (2), and rule (4) at the same time, the feature value corresponding to the feature vector is α1. If the feature vector does not meet the rule (1), rule (2), and rule (4) at the same time, the feature value corresponding to the feature vector is α2.
[0143] If any eigenvector conforms to rule (1), rule (3), and rule (4) at the same time, the eigenvalue 2 corresponding to the eigenvector is β1. If the eigenvector does not conform to rule (1), rule (3), and rule (4) at the same time, the eigenvalue 2 corresponding to the eigenvector is β2.
[0144] If the difference between the solid surface area corresponding to any eigenvector in the building and the single-layer step area is less than the first set threshold, the eigenvalue three corresponding to the eigenvector is η1; if the difference between the solid surface area corresponding to the eigenvector in the building and the staircase turning platform area is less than the second set threshold, the eigenvalue three corresponding to the eigenvector is η2; otherwise, the eigenvalue three corresponding to the eigenvector is η3;
[0145] If the difference between the aspect ratio of the solid surface shape corresponding to any eigenvector in the interior of the building and the aspect ratio of the shape of the single-layer step is less than the third set threshold, the eigenvalue four corresponding to the eigenvector is μ1; if the difference between the aspect ratio of the solid surface shape corresponding to the eigenvector in the interior of the building and the aspect ratio of the shape of the stair turning platform is less than the fourth set threshold, the eigenvalue four corresponding to the eigenvector is μ2; otherwise, the eigenvalue four corresponding to the eigenvector is μ3;
[0146] For example, the present invention counts the number of grids in each layer of grid map in Figure 5The distribution in the X and Y directions is shown in the following figure. Figure 6 , Figure 7 shown.
[0147] Figure 6 Each unit on the X-axis represents each single-layer grid map in the multi-layer grid map. Figure 5 The Z axis represents the situation in the middle; the Y axis represents the situation of all grids in the multi-layer grid map. Figure 5 The distribution on the X axis; the Z axis represents the current single-layer grid map c i (i∈[1,45]) in X=a n (n∈[0,86]). Corresponding to the above-mentioned rules (1), (2), (3), and (4): Figure 6 In the example, taking the small cells with similar protrusions in the Z direction, there are small cells with similar protrusions from a1 to a3 on the Y axis, and the length corresponds to Figure 5 The length of each step plane rectangle in the figure conforms to rule (1); a2 corresponds to Figure 5 The coordinate value of the center point of each step plane on the X-axis conforms to rule (2); and except for the first and last steps of each staircase, each step is closely connected to other steps, forming the convex part framed by the dotted parallelogram in the figure, which conforms to rule (4); By analogy, it can be judged that Figure 6 Each dotted parallelogram in the figure represents a staircase. Therefore, each layer of the grid map is regarded as a separate sample. If the sample meets the rules (1), (2), and (4), the feature value of the sample is set to the value α1 with preliminary staircase features, otherwise it is set to the value α2 indicating that it does not have preliminary staircase features.
[0148] Figure 7 Each unit on the X-axis represents each single-layer grid map in the multi-layer grid map. Figure 5 The Z axis represents the situation in the middle; the Y axis represents the situation of all grids in the multi-layer grid map. Figure 5 The distribution on the Y axis; the Z axis represents the current single-layer grid map c i (i∈[1,45]) in Y=b n The number of grids when (n∈[0,163]); corresponding to the above-mentioned rules (1), (2), (3), and (4): Figure 7 In the example, taking the small grid with similar protrusion degree in the Z direction as an example, there is a protrusion from b1 to b2 on the Y axis, which corresponds to Figure 5The width of each step plane rectangle in the formula complies with rule (1), and similar protrusions are evenly, increasing or decreasingly distributed in the range of b1 to b3 along the Y direction, which complies with rules (3) and (4). Each layer of the grid map is regarded as a separate sample. If the sample complies with rules (1), (3) and (4), the eigenvalue 2 of the sample is set to the value β1 with preliminary staircase features, otherwise it is set to the value β2 indicating that it does not have preliminary staircase features.
[0149] It can be seen that according to the rules shown in the staircase area in the statistical graph, such as Figure 6 , Figure 7 As shown, the feature is quantified into classification feature value 1 and feature value 2 used by the Bayesian classifier.
[0150] Furthermore, according to the above rules, the features of the training sample set X composed of all layers of raster maps are extracted to obtain the training sample data set T = {(x1, y1), (x2, y2), … (x N ,y N )}, N is the total number of grid map layers in the training sample, where represents the jth eigenvalue of the i-th sample formed by the i-th layer grid map, j = 1, 2, 3, 4, a jl is the lth possible value of the jth feature, l = 1, 2, ... S j , S j is the number of possible values of the j-th eigenvalue;
[0151] in, Y represents the terrain label set of the training set samples, and Y = {y1, y2, ..., y i ,...,y N},y i Represents the terrain label value of the i-th sample formed by the i-th layer of the raster map;
[0152] According to the training sample data set T, calculate the prior probability P(Y=c k ):
[0153]
[0154] Where k is the category number of the terrain label, and k = 1 means that the terrain label is a step, k = 2 means that the terrain label is a staircase turning platform, k = 3 means that the terrain label is a floor plane, λ is the Laplace smoothing coefficient, c k Indicates that the terrain label is the kth category, Represents the terrain label value y in the statistical training sample set i Classified as c kThe total number of grid map samples, K represents the number of classifications;
[0155] According to the training sample data set T, calculate the conditional probability P λ (X (j) =a jl |Y=c k ):
[0156]
[0157] Among them, the conditional probability P λ (X (j) =a jl |Y=c k ) means that when the label of the sample in the training sample set X is c k When the jth eigenvalue of the sample is a jl probability; Indicates that the terrain label value in all training sample sets X is c k And the selection of the jth eigenvalue is the number of samples of the lth value;
[0158] The following classification operations are performed on each layer of the constructed multi-layer grid map to be classified:
[0159] The prior probability P(Y=c k ) and conditional probability P λ (X (j) =a jl |Y=c k ), and at the same time extract the characteristic value x of each sample x to be classified formed by each layer of the grid map to be classified = (x (1) ,x (2) ,x (3) ,x (4) ) T , and then calculate the posterior probability value P (Y = c k |X=x):
[0160]
[0161] For each sample to be classified formed by each layer of grid maps to be classified, the class with the largest posterior probability is taken as the classification y of each sample to be classified x, that is, the terrain classification of the grid map of this layer:
[0162]
[0163] Among them, max is the maximum value.
[0164] In summary, the present invention has the following advantages due to the adoption of the above technical solution:
[0165] First, the method of extracting drivable areas significantly reduces the redundant information in the map data, achieving lightweight map construction. It also improves the utilization rate of map data when the ground unmanned platform navigates and searches for paths, avoids searching in non-drivable areas such as walls, and improves the efficiency of path search.
[0166] Second, according to the characteristics of the stairs, rules are formulated to remove residual discrete points, so that the point cloud representing the stair area is retained to the greatest extent. The constructed stair area map is more accurate, which improves the accuracy of navigation of the ground unmanned platform in this area.
[0167] 3. The map edge smoothing operation is used to improve the situation where the constructed map has unclear boundaries, irregular shapes, and small internal defects caused by the randomness of lidar sampling, making the constructed map closer to the physical surface it represents, and the map has a regular shape and clear boundaries.
[0168] Fourth, by extracting map features and using the naive Bayes classification method, different terrains can be distinguished to provide more practical and intelligent environmental information for ground unmanned platform navigation.
[0169] Of course, the present invention may have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may certainly make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.
Claims
1. A lightweight map construction method for staircase environment for unmanned platforms, characterized in that: The following steps are involved: S1: Construct a point cloud map based on the internal point cloud data of the building obtained by LiDAR and inertial navigation; S2: Extract the drivable area of the ground unmanned platform from the point cloud map and remove the discrete point cloud in the drivable area; establish an initial multi-layer grid map based on the point cloud map after removing the discrete point cloud, then perform map edge smoothing on the initial multi-layer grid map, and remove the data representing the ceiling part in the smoothed multi-layer grid map to obtain the final multi-layer grid map; S3: Input the features representing the stair environment terrain in the multi-layer grid map into the Bayesian classifier for terrain classification, and obtain a multi-layer grid map with terrain labels, wherein the terrain labels include steps, stair turning platforms, and floor plans.
2. A method for constructing a lightweight map of a staircase environment for an unmanned platform as claimed in claim 1, characterized in that: The method for extracting the drivable area of the ground unmanned platform from the point cloud map in step S2 is: Each point in the point cloud map is used as the current point A to perform a drivable judgment operation in turn, and all points belonging to the drivable area are screened out, wherein the drivable judgment operation is as follows: Extract the N neighboring points closest to the current point A to build a neighborhood set N; Fit all point clouds in the neighborhood set N to a plane e, and determine whether the normal vector m of the plane e is perpendicular to the horizontal plane. If yes, the current point cloud A belongs to the point cloud of the drivable area. If no, the current point cloud A does not belong to the point cloud of the drivable area.
3. A method for constructing a lightweight map of a stair environment for an unmanned platform as claimed in claim 1, characterized in that: The method for removing the discrete point cloud in the drivable area in step S2 is as follows: Different values are taken for the proportional coefficient ε in (0,1) according to the set step size, and a ratio P acquisition operation is performed for each value ε, wherein the ratio P acquisition operation is as follows: Calculate the threshold N based on the current proportional coefficient ε neighborhood : where d surface is the point cloud surface density in the drivable area, R is the radius of the spherical area constructed with each point cloud in the drivable area as the center of the sphere, and R < min{D, H}, D is the step length, and H is the step height; Construct a spherical area with a radius R for each point cloud with each point cloud in the drivable area as the sphere center, and determine the number of point clouds N contained in each spherical area. number Is it less than N? neighborhood , the point cloud with a judgment result of yes is a discrete point cloud, and the discrete point cloud is removed, while the point cloud with a judgment result of no is retained; N number Not less than N neighborhood The total number of point clouds belonging to stairs in the point cloud N StairNum With N number Not less than N neighborhood The total number of point clouds N TotalNum The ratio between them is taken as P; The value of the proportional coefficient ε corresponding to the maximum value of the ratio P when there is no residual discrete point cloud is taken as the optimal value, and the discrete point cloud in the drivable area under the optimal value is eliminated to obtain the point cloud map after removing the discrete point cloud.
4. The method for constructing a lightweight map of a staircase environment for an unmanned platform as claimed in claim 1, characterized in that: The method for establishing the initial multi-layer grid map according to the point cloud map after removing the discrete point cloud in step S2 is: Divide each point cloud in the point cloud map into different levels according to the height, count the number of point clouds in different levels, and select the height values corresponding to all levels where the number of point clouds is greater than the set threshold δ1 as the grid heights corresponding to each grid layer of the initial multi-layer grid map; According to the set height resolution S z The point clouds at each layer of grid height are integrated into a point cloud horizontal plane set, where the height resolution S z The value of satisfies: each layer of grid height corresponds to a point cloud horizontal plane set and any point cloud only falls into one of the point cloud horizontal plane sets; According to the set horizontal x-axis resolution S x , y-axis resolution S y Divide each horizontal plane point cloud set into multiple grids, and count the number of point clouds in each grid on each horizontal plane point cloud set; Determine whether the number of point clouds in all grids is greater than the set threshold δ2. The grid area where the judgment result is yes is the drivable plane. All drivable planes at each grid height constitute the grid map at each grid height. The grid maps at all grid heights are combined to obtain the initial multi-layer grid map.
5. The method for constructing a lightweight map of a staircase environment for an unmanned platform as claimed in claim 1, characterized in that: The method for performing map edge smoothing on the initial multi-layer grid map in step S2 is: Use the 3×3 convolution kernel K to perform convolution processing on each layer of the multi-layer grid map to obtain a smoothed multi-layer grid map, where the convolution kernel K is:
6. The method for constructing a lightweight map of a staircase environment for an unmanned platform as claimed in claim 1, characterized in that: The method for removing the data representing the ceiling part in the smoothed multi-layer grid map in step S2 is: Use each layer of the multi-layer grid map as the current layer grid map and perform the following operations: Determine whether there are other grids with a grid value of 1 within the set distance range [δ3,δ4] below each grid on the current grid map, and delete the grids with a judgment result of yes; where δ3 and δ4 are the lower and upper limits of the set distance range respectively; a grid value of 1 indicates that there is a solid surface at the corresponding position of the grid inside the building; It is determined whether the number of grids deleted on the current grid map is greater than a threshold value δ5. If so, the current grid map is data corresponding to the ceiling part, which is deleted.
7. The method for constructing a lightweight map of a staircase environment for an unmanned platform as claimed in claim 1, characterized in that: The features representing the stair environment terrain in the multi-layer grid map in step S3 include: the length of the plane rectangle formed by the grid with a value of 1 in each layer of grid map, the width of the plane rectangle formed by the grid with a value of 1 in each layer of grid map, the x-axis projection coordinate of the center point of the plane rectangle formed by the grid with a value of 1 in each layer of grid map, the y-axis projection coordinate difference of the center point of the plane rectangle formed by the grid with a value of 1 in two adjacent layers of grid map, the area of the solid surface corresponding to each layer of grid map in the interior of the building, and the aspect ratio of the shape of the solid surface corresponding to each layer of grid map in the interior of the building; wherein a grid value of 1 indicates that a solid surface exists at the corresponding position of the grid in the building; If any layer of grid maps satisfies the following first joint decision condition, the terrain label corresponding to the grid map of this layer is a step. The first joint decision condition is: the length of the plane rectangle formed by the grid with a value of 1 in the grid map of this layer is the length of the plane rectangle of the step and the width of the plane rectangle formed is the plane width of the step, and the difference between the solid surface area corresponding to the grid map of this layer in the building and the area of the single-layer step is less than a first set threshold; If a continuous multi-layer grid map satisfies the following second joint decision condition, then the continuous multi-layer grid map is a staircase, and the joint decision condition is: the length of the plane rectangle formed by the grid with a value of 1 in each layer of the grid map is the length of the plane rectangle of the step and the width of the plane rectangle formed is the width of the step, the x-axis projection coordinates of the center points of the plane rectangles formed by the grid with a value of 1 in each layer of the grid map are the same, and the y-axis projection coordinate differences of the center points of the plane rectangles formed by the grids with a value of 1 in any two adjacent layers of the grid map are the same; If the difference between the physical surface area corresponding to any layer of the grid map in the building and the area of the stair turning platform is less than the second set threshold and does not meet the first joint judgment condition, then the terrain label corresponding to the grid map of this layer is the stair turning platform; If the difference between the physical surface area corresponding to any layer of the grid map inside the building and the area of the stair turning platform is not less than the second set threshold and does not meet the first joint judgment condition, the terrain label corresponding to the grid map of this layer is the floor plan.
8. A method for constructing a lightweight map of a staircase environment for an unmanned platform as claimed in claim 7, characterized in that: The method of inputting the features representing the staircase environment terrain in the multi-layer grid map into the Bayesian classifier for terrain classification is as follows: The rules for obtaining the stair environment terrain are as follows: Rule (1): The length and width of the rectangles on each step plane are the same; Rule (2): For each flight of stairs, the x-axis projection coordinates of the center point of the plane rectangle of each step are the same; Rule (3): For each section of stairs, the difference between the y-axis projection coordinates of the center points of the plane rectangles of any two adjacent steps is the same, and the y-axis projection coordinates of the center point of the plane rectangle of each step are within the range of the y-axis projection coordinates of the section of stairs; Rule (4): For each flight of stairs, except for the first and last two steps, the upper and lower adjacent steps of each step are also steps, and they all satisfy rules (1), (2), and (3); Obtain the feature vectors of the features of the stair environment terrain of each layer of the multi-layer grid map. If any feature vector meets the rule (1), rule (2), and rule (4) at the same time, the feature value corresponding to the feature vector is α1. If the feature vector does not meet the rule (1), rule (2), and rule (4) at the same time, the feature value corresponding to the feature vector is α2. If any eigenvector conforms to rule (1), rule (3), and rule (4) at the same time, the eigenvalue 2 corresponding to the eigenvector is β1. If the eigenvector does not conform to rule (1), rule (3), and rule (4) at the same time, the eigenvalue 2 corresponding to the eigenvector is β2. If the difference between the solid surface area corresponding to any eigenvector in the building and the single-layer step area is less than the first set threshold, the eigenvalue three corresponding to the eigenvector is η1; if the difference between the solid surface area corresponding to the eigenvector in the building and the staircase turning platform area is less than the second set threshold, the eigenvalue three corresponding to the eigenvector is η2; otherwise, the eigenvalue three corresponding to the eigenvector is η3; If the difference between the aspect ratio of the solid surface shape corresponding to any eigenvector in the interior of the building and the aspect ratio of the shape of the single-layer step is less than the third set threshold, the eigenvalue four corresponding to the eigenvector is μ1; if the difference between the aspect ratio of the solid surface shape corresponding to the eigenvector in the interior of the building and the aspect ratio of the shape of the stair turning platform is less than the fourth set threshold, the eigenvalue four corresponding to the eigenvector is μ2; otherwise, the eigenvalue four corresponding to the eigenvector is μ3; According to the above rules, the features of the training sample set X composed of all layers of raster maps are extracted to obtain the training sample data set T = {(x1, y1), (x2, y2), … (x N ,y N )}, N is the total number of grid map layers in the training sample, where represents the jth eigenvalue of the i-th sample formed by the i-th layer grid map, j = 1, 2, 3, 4, a jl is the lth possible value of the jth feature, l = 1, 2, ... S j , S j is the number of possible values of the j-th eigenvalue; in, Y represents the terrain label set of the training set samples, and Y = {y1, y2, ..., y i ,...,y N },y i Represents the terrain label value of the i-th sample formed by the i-th layer of the raster map; According to the training sample data set T, calculate the prior probability P(Y=c k ): Where k is the category number of the terrain label, and k = 1 means that the terrain label is a step, k = 2 means that the terrain label is a staircase turning platform, k = 3 means that the terrain label is a floor plane, λ is the Laplace smoothing coefficient, c k Indicates that the terrain label is the kth category, Represents the terrain label value y in the statistical training sample set i Classified as c k The total number of grid map samples, K represents the number of classifications; According to the training sample data set T, calculate the conditional probability P λ (X (j) =a jl |Y=c k ): Among them, the conditional probability P λ (X (j) =a jl |Y=c k ) means that when the label of the sample in the training sample set X is c k When the jth eigenvalue of the sample is a jl probability; Indicates that the terrain label value in all training sample sets X is c k And the selection of the jth eigenvalue is the number of samples of the lth value; The following classification operations are performed on each layer of the constructed multi-layer grid map to be classified: The prior probability P(Y=c k ) and conditional probability P λ (X (j) =a jl |Y=c k ), and at the same time extract the characteristic value x of each sample x to be classified formed by each layer of the grid map to be classified = (x (1) ,x (2) ,x (3) ,x (4) ) T , and then calculate the posterior probability value P (Y = c k |X=x): For each sample to be classified formed by each layer of grid maps to be classified, the class with the largest posterior probability is taken as the classification y of each sample to be classified x, that is, the terrain classification of the grid map of this layer: Among them, max is the maximum value.
