Farm high-precision map making method and system

Abstract

The invention provides a farm high-precision map making method and system. The method comprises the steps of obtaining target farm three-dimensional point cloud data corresponding to a farm area to be mapped; performing point cloud semantic segmentation on the three-dimensional point cloud data of the target farm to obtain a first point cloud semantic segmentation result and a second point cloud semantic segmentation result; based on the farmland point cloud curvature, boundary point extraction processing is carried out on the first point cloud semantic segmentation result, and a target farmland boundary point set is obtained; determining an entrance and exit candidate point set according to the node degrees of the corresponding nodes of each pixel point in the road center line pixel set in the undirected graph, and determining an entrance and exit candidate point set according to the distances between the farmland boundary points in the target farmland boundary point set and the entrance and exit candidate points in the entrance and exit candidate point set; generating a ternary topology connection association relationship among the farmland points, the entrances and exits and the farmland road sections; and according to the ternary topology connection association relationship, constructing a high-precision map file of the to-be-drawn farm area. According to the invention, the precision of the farm high-precision map is effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of high-precision mapping technology, and in particular to a high-precision farmland mapping method and system. BACKGROUND

[0002] With the development of precision agriculture and agricultural automatic driving, there is an urgent need for high-precision maps in agricultural scenarios. High-precision maps in urban scenarios are usually based on multi-layer information and use standardized formats to accurately express roads and lanes. However, existing methods mainly focus on urban lane networks and are difficult to unify the vectorization and topological expression of farmland plots, entrances and exits, and field paths.

[0003] Although the existing manual dotting mapping method has high precision, it is costly, time-consuming, and has a low update frequency, making it difficult to meet the needs of agricultural monitoring. Satellite remote sensing mapping can cover a wide range, but the spatial resolution is low and affected by cloud cover, making it difficult to capture the micro boundaries of farmland plots. Ground vehicle automated collection methods are also limited by factors such as operational path layout and crop canopy shading, making it difficult to obtain complete farmland surface data, and the system integration and maintenance costs are high.

[0004] Therefore, there is an urgent need for a high-precision farmland mapping method and system to solve the above problems. SUMMARY

[0005] To solve the problems of the prior art, the present application provides a high-precision farmland mapping method and system.

[0006] The present application provides a high-precision farmland mapping method, comprising: obtaining target farmland three-dimensional point cloud data corresponding to a farmland area to be mapped; performing point cloud semantic segmentation on the target farmland three-dimensional point cloud data to obtain a first point cloud semantic segmentation result and a second point cloud semantic segmentation result, wherein the first point cloud semantic segmentation result is point cloud data with farmland point cloud semantic labels, and the second point cloud semantic segmentation result is point cloud data with road point cloud semantic labels; performing boundary point extraction processing on the first point cloud semantic segmentation result based on farmland point cloud curvature to obtain a target farmland boundary point set; determining an entrance candidate point set according to the node degrees of each pixel point in the road center line pixel set in the undirected graph, and generating a ternary topological connection association relationship between farmland points, entrances, and farmland road segments according to the distances between farmland boundary points in the target farmland boundary point set and entrance candidate points in the entrance candidate point set, wherein the road center line pixel set is constructed based on the second point cloud semantic segmentation result; According to the ternary topological connection association relationship, a high-precision map file corresponding to the farmland area to be mapped is constructed.

[0007] The farm high-precision mapping method provided by the application comprises the following steps: searching a neighborhood within a preset radius range of the KD-Tree data structure to calculate the farmland point cloud curvature corresponding to each point in the first point cloud semantic segmentation result; performing corresponding weighted mean filtering processing on each point in the first point cloud semantic segmentation result according to the farmland point cloud curvature and a preset curvature threshold to obtain filtered point cloud data; performing farmland plot segmentation on the filtered point cloud data based on a region growing algorithm of a normal vector angle and a curvature to obtain multiple farmland plot point cloud data; projecting points in each farmland plot point cloud data to a horizontal plane to obtain two-dimensional projection coordinate information corresponding to the points in each farmland plot point cloud data; dividing regions according to the two-dimensional projection coordinate information and a preset grid resolution, and constructing an initial farmland boundary point set according to a first projection coordinate point and a second projection coordinate point in each region, wherein the first projection coordinate point is a point corresponding to the largest two-dimensional projection coordinate information in each region, and the second projection coordinate point is a point corresponding to the smallest two-dimensional projection coordinate information in each region; performing filtering processing on boundary points in the initial farmland boundary point set based on an alpha-Shape algorithm to obtain a pending boundary point set; sorting the boundary points in the pending boundary point set based on any one boundary point in the pending boundary point set and a nearest neighbor greedy strategy to obtain a target farmland boundary point set.

[0008] The farm high-precision mapping method provided by the application comprises the following steps: projecting points in the second point cloud semantic segmentation result to a horizontal plane to obtain a two-dimensional projection coordinate set corresponding to the second point cloud semantic segmentation result and a projection plane; dividing the projection plane into multiple unit cells based on a preset grid size; calculating a grid point density corresponding to each unit cell according to the two-dimensional projection coordinate set, and constructing a target density map according to the grid point density; binarize the target density map to obtain a binarized target density map, and perform a closing operation and a filling operation on the binarized target density map to obtain a target binary image; Based on the skeletonization algorithm, the target binary image is subjected to road center line extraction processing to obtain the road center line pixel set; According to the road center line pixel set, a road center point undirected graph is constructed, and the node degrees corresponding to each node in the road center point undirected graph are calculated; According to the node degrees, the farmland entrance and exit points and the farmland road intersection points in the road center point undirected graph are determined, and the farmland entrance and exit points and the farmland road intersection points are taken as the entrance and exit candidate points to obtain the entrance and exit candidate point set.

[0009] According to the distance between the farmland boundary points in the target farmland boundary point set and the entrance and exit candidate points in the entrance and exit candidate point set, a ternary topological connection association relationship between the farmland points, the entrances and exits, and the farmland road segments is generated. A first minimum Euclidean distance is obtained, wherein the first minimum Euclidean distance is the minimum Euclidean distance between the farmland boundary points in the target farmland boundary point set and the entrance and exit candidate points in the entrance and exit candidate point set; The entrance and exit candidate points in the entrance and exit candidate point set whose first minimum Euclidean distance is less than or equal to a first preset Euclidean distance are determined as potential entrance and exit nodes; All the potential entrance and exit nodes are subjected to clustering processing, and the target entrance and exit position information is determined according to the centroids corresponding to each cluster in the obtained clustering processing result, wherein the target entrance and exit position information includes farmland entrance and exit position information and road entrance and exit position information; A second minimum Euclidean distance is obtained, and if it is determined that the second minimum Euclidean distance is less than a second preset Euclidean distance, a first association relationship is established; wherein the second minimum Euclidean distance is the minimum Euclidean distance between the farmland boundary points in the target farmland boundary point set and the farmland entrance and exit position information; and the first association relationship is the association relationship between the target farmland boundary point set and the farmland entrance and exit position information; Based on the road center line pixel set, each road segment node information corresponding to the farmland area to be mapped is determined; A third minimum Euclidean distance is obtained, and if it is determined that the third minimum Euclidean distance is less than a third preset Euclidean distance, a second association relationship is established; wherein the third minimum Euclidean distance is the minimum Euclidean distance between the farmland entrance and exit position information and the road segment node information; and the second association relationship is the association relationship between the farmland entrance and exit position information and the road segment node information. generate the ternary topological connection association relationship between the farmland points, the entrances and exits and the farmland road segments based on the intersection between the first association relationship and the second association relationship.

[0010] According to the ternary topological connection association relationship, a high-precision map file corresponding to the farmland region to be mapped is constructed, Based on the ternary topological connection association relationship, farmland ID information, entrance and exit coordinate information and road segment ID information corresponding to the farmland region to be mapped are generated; According to the farmland ID information, the entrance and exit coordinate information and the road segment ID information, a KML file corresponding to the farmland region to be mapped is constructed, so as to generate the high-precision map file according to the KML file.

[0011] According to the ternary topological connection association relationship, a high-precision map file corresponding to the farmland region to be mapped is constructed, By training the point cloud semantic segmentation model, the target farmland three-dimensional point cloud data is subjected to point cloud semantic segmentation, and the first point cloud semantic segmentation result and the second point cloud semantic segmentation result are obtained; The trained semantic segmentation model is trained by the following steps: According to the aerial image sample data of the farmland region, sample farmland three-dimensional point cloud data is obtained; The sample farmland three-dimensional point cloud data is marked with a corresponding semantic label, and a high-precision point cloud data set of the farmland is obtained, wherein the semantic label at least includes the farmland point cloud semantic label and the road point cloud semantic label; According to the high-precision point cloud data set of the farmland, the pre-trained semantic segmentation model is trained to obtain the trained semantic segmentation model.

[0012] The application also provides a high-precision map mapping system for a farmland, comprising: A three-dimensional point cloud acquisition module is configured to acquire target farmland three-dimensional point cloud data corresponding to a farmland region to be mapped; A semantic segmentation module is configured to perform point cloud semantic segmentation on the target farmland three-dimensional point cloud data to obtain a first point cloud semantic segmentation result and a second point cloud semantic segmentation result, wherein the first point cloud semantic segmentation result is point cloud data with a farmland point cloud semantic label, and the second point cloud semantic segmentation result is point cloud data with a road point cloud semantic label; A first processing module is configured to perform boundary point extraction processing on the first point cloud semantic segmentation result based on farmland point cloud curvature to obtain a target farmland boundary point set; The second processing module is configured to determine an entrance and exit candidate point set according to the node degree of each pixel point in the road center line pixel set in the undirected graph, and generate a ternary topological connection association relationship between a farmland point, an entrance and exit, and a farmland road segment according to the distance between a farmland boundary point in the target farmland boundary point set and an entrance and exit candidate point in the entrance and exit candidate point set, wherein the road center line pixel set is constructed based on the second point cloud semantic segmentation result. The mapping module is configured to construct a high-precision map file corresponding to the farmland region to be mapped according to the ternary topological connection association relationship.

[0013] The present application also provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the farmland high-precision mapping method according to any one of the above methods when executing the program.

[0014] The present application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program is executable by a processor to implement the farmland high-precision mapping method according to any one of the above methods.

[0015] The present application also provides a computer program product comprising a computer program, wherein the computer program is executable by a processor to implement the farmland high-precision mapping method according to any one of the above methods.

[0016] The present application provides a farmland high-precision mapping method and system, which performs semantic segmentation on three-dimensional point cloud data of a farmland region to be mapped to obtain two results with farmland and road point cloud semantic labels; then, a boundary point set is extracted from the farmland label point cloud according to the curvature of the farmland point cloud; at the same time, a road center line pixel set is constructed based on the road label point cloud, and an entrance and exit candidate point set is determined by analyzing the node degree of the pixel point in the undirected graph; finally, the distance between the farmland boundary point and the entrance and exit candidate point is calculated to generate a ternary topological connection association relationship, and a high-precision map file of the farmland is constructed accordingly, effectively improving the precision of the farmland high-precision map. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the present application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without any creative effort.

[0018] Figure 1 The flowchart of the farmland high-precision mapping method provided by the present application is shown in the figure. Figure 2 A flowchart of a farmland boundary extraction process provided by the present application is shown in FIG. 1. Figure 3 A schematic diagram of a construction process of a ternary topology connection association relationship provided by the present application is shown in FIG. 2. Figure 4 A schematic diagram of a high-precision farmland map provided by the present application is shown in FIG. 3. Figure 5 A schematic diagram of a structure of a high-precision farmland mapping system provided by the present application is shown in FIG. 4. Figure 6 A schematic diagram of a structure of an electronic device provided by the present application is shown in FIG. 5. DETAILED DESCRIPTION

[0019] To make the objectives, technical solutions, and advantages of the present application clearer, the technical solutions in the present application will be described below in conjunction with the accompanying drawings in the present application. Obviously, the described embodiments are some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0020] With the rapid development of precision agriculture and agricultural vehicle automatic driving, the demand for high-precision maps supporting centimeter-level positioning and path planning in agricultural scenarios is increasingly urgent. In urban scenarios, high-precision maps are usually based on multi-layer information such as geometric elements, semantic annotations, and topological relationships, and realize accurate expression of roads and lanes through standardized formats such as Lanelet2 and OpenDRIVE. However, existing high-precision mapping methods mostly focus on urban lane networks, and it is difficult to realize unified vectorization and topological expression of farmland plots, entrances and exits, and field paths.

[0021] Satellite remote sensing has long been the main means of obtaining large-scale farmland geographic information, but its spatial resolution is usually above meters, making it difficult to provide centimeter-level details of ground objects. Although ground laser radar can obtain high-precision point clouds, it is difficult to achieve large-scale continuous coverage due to limited viewing angle and high cost. In recent years, unmanned aerial vehicles equipped with real-time dynamic differential positioning (RTK) modules combined with structure from motion (SfM) technology provide a feasible way for three-dimensional reconstruction of farmland environments. SfM technology can reconstruct high-overlap images into dense and high-precision point clouds, with positioning errors controlled within centimeters and high spatial resolution, thereby significantly improving the geometric depiction accuracy of heterogeneous farmland features.

[0022] However, existing approaches relying solely on high-precision semantic segmentation of farm point clouds are unable to meet the requirements for expressing the "field-entrance-road" ternary topology required for agricultural machinery navigation. To achieve this goal, existing research typically follows a three-stage process: identifying plot boundaries, extracting road structures, and constructing topological relationships. However, these existing methods do not yet integrate point cloud semantic segmentation, plot boundary, and road network centerline node extraction into an automated pipeline, and they are unable to maintain topological consistency across large and complex farm environments.

[0023] In addition, although the existing mapping method through manual marking can achieve centimeter-level positioning accuracy, it is difficult to meet the rapidly changing needs of agricultural monitoring due to the high cost of manual collection, long field operation time and low update frequency. Although satellite remote sensing mapping has the advantage of seamless coverage over a large area, it is limited by the spatial resolution above 10dm and the problem of image acquisition being affected by cloud obstruction, making it difficult to capture the microscopic boundaries of plots. High-density point clouds and panoramic images are automatically collected by ground vehicles driving along the operating channel, but their coverage area is limited by the layout of the operating channel and obstruction by the crop canopy, making it difficult to obtain complete farmland surface data. In addition, ground undulations and rutting affect driving stability. In addition, the dust and water resistance and real-time data processing requirements of the sensor increase system integration and maintenance costs.

[0024] To address the problems existing in the above-mentioned existing technologies, the present invention adopts low-altitude aerial photography by drones combined with SfM technology to provide a feasible approach for the three-dimensional reconstruction of farmland environments. Among them, SfM technology can reconstruct highly overlapping images into dense, high-precision point clouds. Its positioning error can be controlled at the centimeter level, and it has high spatial resolution, greatly improving the geometric depiction accuracy of heterogeneous farmland objects.

[0025] Figure 1 The flowchart of the method for making a high-precision farm map provided by the present invention is as follows: Figure 1 As shown, the present invention provides a method for making a high-precision farm map, comprising: Step 101, obtaining the three-dimensional point cloud data of the target farm corresponding to the farm area to be mapped; In this method, a multi-rotor drone is used to capture aerial images of the farm to be mapped along a set route. These images are then imported into 3D reconstruction software to generate a 3D point cloud of the farm area to be mapped, representing the target farm's 3D point cloud data. Point cloud data is a collection of numerous 3D points, each containing spatial coordinate information, accurately reflecting the farm area's topography and other characteristics.

[0026] In step 102, point cloud semantic segmentation is performed on the target farm three-dimensional point cloud data to obtain a first point cloud semantic segmentation result and a second point cloud semantic segmentation result, wherein the first point cloud semantic segmentation result is point cloud data with a farmland point cloud semantic label, and the second point cloud semantic segmentation result is point cloud data with a road point cloud semantic label.

[0027] In the present application, each point in the target farm three-dimensional point cloud data is assigned a specific semantic label by point cloud semantic segmentation technology to distinguish different categories of objects. The present application divides the point cloud data into two parts: one part is point cloud data with a farmland point cloud semantic label (i.e., the first point cloud semantic segmentation result), which represents the farmland area; the other part is point cloud data with a road point cloud semantic label (i.e., the second point cloud semantic segmentation result), which represents the road area. Through semantic segmentation, the location and range of farmland and roads in the farm can be clearly identified.

[0028] In step 103, boundary point extraction processing is performed on the first point cloud semantic segmentation result based on the farmland point cloud curvature to obtain a target farmland boundary point set.

[0029] In the present application, point cloud curvature reflects the degree of bending of point cloud data in a local area. The point cloud curvature at the boundary of the farmland usually changes significantly, and by using this characteristic, boundary points can be extracted from the farmland point cloud data. These boundary points constitute the target farmland boundary point set, which can accurately depict the outline of the farmland and provide important information for subsequent map construction.

[0030] In step 104, an entrance candidate point set is determined according to the node degree of each pixel point in the road centerline pixel set in the undirected graph, and a ternary topological connection relationship between farmland points, entrances, and farmland road segments is generated according to the distance between the farmland boundary points in the target farmland boundary point set and the entrance candidate points in the entrance candidate point set, wherein the road centerline pixel set is constructed based on the second point cloud semantic segmentation result.

[0031] In the present application, the road centerline pixel set is constructed based on the second point cloud semantic segmentation result (road point cloud data). Then, the entrance candidate point set is determined according to the node degree of each pixel point in the road centerline pixel set in the undirected graph.

[0032] In the present application, the node degree reflects the connection of the node with other nodes, and in the road network, the node at the entrance and exit position usually has a specific degree feature. Then, the distance between the farmland boundary points in the target farmland boundary point set and the entrance and exit candidate points in the entrance and exit candidate point set is calculated, and a ternary topological connection association relationship between the farmland points, the entrance and exit and the farmland road segments is generated according to the distance information. This association relationship can accurately describe the connection between the farmland, the entrance and exit and the road, and provide a basis for the topological structure of the map.

[0033] In step 105, a high-precision map file corresponding to the farmland area to be mapped is constructed according to the ternary topological connection association relationship.

[0034] In the present application, after obtaining the topological connection association relationship between the farmland, the entrance and exit and the road, the point cloud data and the semantic segmentation result obtained before are integrated by using the map construction algorithm and technology, and a high-precision map file of the farmland area to be mapped is generated. The map file can accurately reflect the topography, farmland distribution, road direction and topological relationship of the farmland area, and provide reliable map support for precision agriculture and automatic driving of agricultural machinery.

[0035] The farmland high-precision mapping method provided by the present application can obtain two parts of results with farmland and road point cloud semantic labels by performing semantic segmentation on the three-dimensional point cloud data of the farmland area to be mapped. Then, the boundary point set is extracted from the farmland labeled point cloud according to the farmland point cloud curvature. Meanwhile, the road center line pixel set is constructed based on the road labeled point cloud, and the entrance and exit candidate point set is determined by analyzing the node degree of the pixel points in the undirected graph. Finally, the distance between the farmland boundary points and the entrance and exit candidate points is calculated, the ternary topological connection association relationship is generated, and the high-precision map file of the farmland is constructed accordingly, which effectively improves the precision of the high-precision map of the farmland.

[0036] On the basis of the above embodiment, the boundary point extraction processing is performed on the first point cloud semantic segmentation result based on the farmland point cloud curvature to obtain the target farmland boundary point set, including: The neighborhood within the preset radius range of the KD-Tree data structure is searched, and the farmland point cloud curvature corresponding to each point in the first point cloud semantic segmentation result is calculated; According to the farmland point cloud curvature and the preset curvature threshold, the corresponding weighted mean filtering processing is performed on each point in the first point cloud semantic segmentation result to obtain the filtered point cloud data; Based on the region growing algorithm of the normal vector angle and the curvature, the farmland land segmentation is performed on the filtered point cloud data to obtain a plurality of farmland land point cloud data. Projecting the points in each of the farmland plot point cloud data to a horizontal plane to obtain two-dimensional projection coordinate information corresponding to the points in each of the farmland plot point cloud data; According to the two-dimensional projection coordinate information and a preset grid resolution, the points in each of the farmland plot point cloud data projected to the horizontal plane are regionally divided, and an initial farmland boundary point set is constructed according to first projection coordinate points and second projection coordinate points in each region, wherein the first projection coordinate points are points corresponding to the largest two-dimensional projection coordinate information in each region; and the second projection coordinate points are points corresponding to the smallest two-dimensional projection coordinate information in each region. Based on an alpha-Shape algorithm, the boundary points in the initial farmland boundary point set are filtered to obtain a pending boundary point set. Based on any one boundary point in the pending boundary point set and a nearest neighbor greedy strategy, the boundary points in the pending boundary point set are sorted to obtain a target farmland boundary point set.

[0037] The KD-Tree is a data structure for efficiently searching points in a multi-dimensional space. In the first point cloud semantic segmentation result (i.e., farmland point cloud data), for each point, the KD-Tree is used to search its neighborhood points within a preset radius range in the present application. This preset radius range can be set according to actual requirements and the density of the point cloud data, etc., which determines the size of the neighborhood range considered when calculating the curvature.

[0038] In the present application, the farmland point cloud curvature corresponding to the point is calculated according to the searched neighborhood points. The curvature reflects the bending degree of the point cloud data in the local region, and the point cloud curvature at the farmland boundary usually changes significantly, so the curvature information can be used for subsequent farmland plot segmentation and boundary extraction.

[0039] Further, a curvature threshold, i.e., a preset curvature threshold, is preset. This preset curvature threshold is used to determine whether the points in the point cloud data need to be filtered. In the present application, the selection of the preset curvature threshold needs to consider factors such as the noise level of the point cloud data and the complexity of the farmland terrain. Then, according to the comparison result of the calculated farmland point cloud curvature and the preset curvature threshold, corresponding weighted mean filtering processing is performed on each point in the first point cloud semantic segmentation result. Through weighted mean filtering, the point cloud data can be smoothed, noise interference can be reduced, and the key features of the farmland terrain can be preserved at the same time.

[0040] In the present application, the point cloud is segmented based on the region growing algorithm of the normal vector angle and curvature. Specifically, the algorithm first selects a seed point, and then gradually merges the neighborhood points with similar normal vector angle and curvature information to the seed point into the same region until a certain stopping condition is met. In this way, the filtered point cloud data is divided into multiple farmland plot point cloud data, and each plot point cloud data represents an independent farmland area.

[0041] Further, the points in each farmland plot point cloud data are projected onto the horizontal plane. The projection process can be achieved by ignoring the vertical coordinates (height information) of the points and only retaining the horizontal coordinates (two-dimensional projection coordinates) of the points. After projection, the two-dimensional projection coordinate information corresponding to the points in each farmland plot point cloud data is obtained, which will be used for subsequent region division and boundary extraction. Then, according to the two-dimensional projection coordinate information and the preset grid resolution, the points in each farmland plot point cloud data projected onto the horizontal plane are regionally divided. In the present application, the preset grid resolution determines the size of each region. The higher the grid resolution, the finer the region division, and the more detailed boundary information can be captured. In each region, the point corresponding to the largest two-dimensional projection coordinate information (i.e. the first projection coordinate point) and the point corresponding to the smallest two-dimensional projection coordinate information (i.e. the second projection coordinate point) are found. These points are usually located on the boundary of the region, and they are collected to construct an initial farmland boundary point set. Although the initial farmland boundary point set contains possible boundary points, it may contain some noise points or redundant points, which need to be further processed.

[0042] The α-Shape algorithm is an algorithm for extracting boundaries from a point set. By introducing a parameter α, the shape and complexity of the boundary can be controlled. In the present application, for the boundary points in the initial farmland boundary point set, the α-Shape algorithm is used for filtering to remove some noise points and redundant points, and to obtain a tentative boundary point set. The tentative boundary point set more accurately reflects the shape of the farmland boundary.

[0043] Further, in the present application, based on any one boundary point in the tentative boundary point set, the nearest neighbor greedy strategy is used to sort the boundary points in the tentative boundary point set. The basic idea of the nearest neighbor greedy strategy is to select the boundary point closest to the current point as the next point each time, and connect all the boundary points in turn until returning to the starting point. In this way, the boundary points in the tentative boundary point set are arranged in a certain order to obtain a target farmland boundary point set. The target farmland boundary point set can accurately describe the boundary of the farmland, and provides an important basis for subsequent map construction and farmland management.

[0044] Figure 2 The flowchart of the farmland boundary extraction process provided by the present application can be referred toFigure 2 As shown in the present application, the farmland plot boundary is extracted from the point set belonging to the "farmland" category in the semantic marked point cloud, and the specific process is as follows: First, input the farmland points In order to smooth the noise while preserving the edge features of the field, a weighted mean filtering based on point cloud curvature is introduced. The present application constructs a KD-Tree, and searches for the neighborhood within a radius r Range, calculates the farmland point cloud curvature i of each point in the first point cloud semantic segmentation result: Among them, , , The characteristic value of the neighborhood point, and .

[0045] Further, according to the preset curvature threshold and the enhancement factor , the weighted mean filtering processing or the original boundary point is retained: Among them, the weight , is the distance between the first i point and the first j point; The set of neighborhood points of , the first point adjacent to the first i point, j The first point after weighted mean filtering processing or the original boundary point is retained. This filtering can not only remove small-scale noise, but also retain the farmland edge at high curvature.

[0046] Further, in order to ensure the independence of adjacent farmland clusters, the present application uses a region growing algorithm based on the angle between the normal vector and the curvature threshold to segment the farmland plot. Then, the multiple farmland plot point cloud data obtained by segmentation is projected onto the horizontal plane to obtain two-dimensional projection coordinate information, and the projection area is divided according to the preset grid resolution R. The minimum and maximum projection coordinate points are selected in each row and column grid to form an initial farmland boundary point set.

[0047] Next, the initial farmland boundary point set is executed α-Shape algorithm, and the edges with a distance greater than the threshold value from the adjacent points are removed, and the true farmland boundary points (i.e. the pending boundary points) are retained. Then, starting from any boundary point in these pending boundary points, the nearest neighbor greedy strategy is used to generate a boundary ordered point sequence, so as to obtain the final target farmland boundary point set.

[0048] ​​The application is based on a weighted mean filtering algorithm of point cloud curvature, which preserves the fine feature edge of objects while denoising, and realizes high-precision farmland boundary extraction by combining the region growing and alpha-shape optimization of normal vector and curvature.

[0049] On the basis of the above-mentioned embodiments, the entrance candidate point set is determined according to the node degree of each pixel point in the road center line pixel set in the undirected graph. The points in the second point cloud semantic segmentation result are projected onto a horizontal plane to obtain a two-dimensional projection coordinate set corresponding to the second point cloud semantic segmentation result and a projection plane; Based on a preset grid size, the projection plane is divided into cells to obtain a plurality of cells; According to the two-dimensional projection coordinate set, the grid point density corresponding to each cell is calculated, and a target density map is constructed according to the grid point density; The target density map is binarized to obtain a binarized target density map, and the binarized target density map is closed and filled to obtain a target binary image; Based on a skeleton algorithm, the target binary image is subjected to road center line extraction processing to obtain the road center line pixel set; According to the road center line pixel set, a road center point undirected graph is constructed, and the node degree corresponding to each node in the road center point undirected graph is calculated; According to the node degree, the farmland entrance and exit points and the farmland road intersection points in the road center point undirected graph are determined, and the farmland entrance and exit points and the farmland road intersection points are taken as the entrance and exit candidate points to obtain the entrance and exit candidate point set.

[0050] In the application, the second point cloud semantic segmentation result is point cloud data with road point cloud semantic labels, the points are projected onto a horizontal plane, the vertical coordinates (height information) of the points are ignored, and only the horizontal coordinates (two-dimensional projection coordinates) are retained. Through this projection operation, a two-dimensional projection coordinate set corresponding to the second point cloud semantic segmentation result is obtained, and these coordinate points represent the position distribution of the road on the horizontal plane. At the same time, the plane where all the projected points are located is the projection plane, which provides a basic space for subsequent cell division and density calculation.

[0051] Further, according to the actual demand and the fine degree of road distribution, a grid size, i.e., a preset grid size, is set in advance, which determines the size of each cell. The smaller the grid size, the more refined the division, and the more detailed road information can be captured. Based on the preset grid size, the projection plane is divided into cells, so that the projection plane is divided into a plurality of cells of the same size. Each cell can be regarded as a small area, and the density of the point will be calculated in the area subsequently.

[0052] Further, according to the two-dimensional projection coordinate set, the number of projection points contained in each cell is counted, which is the grid point density corresponding to the cell. The grid point density reflects the density of the road point cloud in the cell. The road area usually has a higher grid point density. Then, the grid point density information of all cells is represented in the form of an image, and the target density map is constructed.

[0053] In the present application, the target density map is binarized, and a threshold is set. The pixel points corresponding to the cells with a density higher than the threshold are set to white (or a high brightness value), and the pixel points corresponding to the cells with a density lower than the threshold are set to black (or a low brightness value). Through the binarization processing, the target density map is converted into a black and white image, and the road area (white part) and the non-road area (black part) are highlighted.

[0054] Further, the image after the binarization processing may have some noise or holes, which affects the integrity of the road area. The present application can fill small holes in the image through closing operation, connect adjacent white areas, and make the road area more continuous. The filling processing further fills the holes in the image after the closing operation, ensures that there is no blank area in the road area, and obtains the target binary image.

[0055] Further, based on the skeletonization algorithm, the target binary image is subjected to road center line extraction processing. The purpose of the skeletonization algorithm is to reduce the road area (white part) to a curve with a width of one pixel, which is the center line of the road. Through the skeletonization processing, the center position of the road can be accurately extracted, and the road center line pixel set is obtained, which contains the coordinate information of each pixel point on the road center line. In the present application, according to the road center line pixel set, a road center point undirected graph is constructed. Each pixel point on the road center line is regarded as a node in the graph. If two nodes are adjacent on the road center line, an edge is added between them, so that an undirected graph is formed. The road center point undirected graph can intuitively represent the topological structure of the road.

[0056] Further, in the road center point undirected graph, the node degree corresponding to each node is calculated. The node degree reflects the connection of the node in the road network, for example, the degree of the intersection node is usually high because it connects multiple roads. According to the node degree, the farmland entrance and exit point and the farmland road intersection point in the road center point undirected graph are determined, specifically, the node with a degree of 1 can be a farmland entrance and exit point because it is connected to only one road; the node with a degree greater than 2 can be a farmland road intersection point because it connects multiple roads. Then, the determined farmland entrance and exit point and the farmland road intersection point are taken as entrance and exit candidate points, and are collected to obtain an entrance and exit candidate point set. The entrance and exit candidate point set provides important reference information for subsequent map construction, path planning and the like, for example, in the automatic driving of agricultural machinery, the agricultural machinery needs to determine the path for entering and exiting the farmland according to the entrance and exit candidate points.

[0057] In an embodiment, the construction process of the entrance and exit candidate point set is specifically described, and the steps are as follows: First, the road class point cloud (i.e., the second point cloud semantic segmentation result) is projected to the horizontal plane to obtain a two-dimensional projection coordinate set, and the projection coordinate set is divided into a plurality of cells according to a preset grid size g The projection plane is divided into a plurality of cells, and the cell density is calculated to form a target density map.

[0058] Then, a threshold value τ The target density map is binarized, and the binarized density map is subjected to a closing operation and a small hole filling process to obtain a target binary image, so as to maintain the road connectivity and remove isolated noise.

[0059] Next, a skeletonization algorithm is performed on the morphologically processed binary image (i.e., the target binary image) to obtain a road center line pixel set (skeleton pixel set), and the road center line pixel set is converted into a road center point undirected graph G = ( V , E ), the node degree is calculated as the basis for subsequent entrance and exit discrimination, and the calculation formula of the node degree is as follows: ; Wherein, ( u , v ) ∈ E represents that in the road center point undirected graph G =( V , E ), there is an undirected edge between node u and node v .

[0060] Further, from the road center point undirected graph GThe candidate set of the node degree of 1 (farmland entrance and exit) or the node degree greater than 2 (road intersection) is extracted, and the formula is: ; Wherein, The entrance candidate point set is represented.

[0061] On the basis of the above-mentioned embodiment, the ternary topological connection association relationship between the farmland point, the entrance and exit and the farmland road segment is generated according to the distance between the farmland boundary point in the target farmland boundary point set and the entrance and exit candidate point in the entrance and exit candidate point set, and the ternary topological connection association relationship includes: The first minimum Euclidean distance is obtained, wherein the first minimum Euclidean distance is the minimum Euclidean distance between the farmland boundary point in the target farmland boundary point set and the entrance and exit candidate point in the entrance and exit candidate point set; The entrance and exit candidate point in the entrance and exit candidate point set with the first minimum Euclidean distance less than or equal to the first preset Euclidean distance is determined as a potential entrance and exit node; The clustering processing is performed on all the potential entrance and exit nodes, and the target entrance and exit position information is determined according to the centroid corresponding to each cluster in the obtained clustering processing result, wherein the target entrance and exit position information includes farmland entrance and exit position information and road entrance and exit position information; The second minimum Euclidean distance is obtained, and if it is determined that the second minimum Euclidean distance is less than the second preset Euclidean distance, the first association relationship is established; wherein the second minimum Euclidean distance is the minimum Euclidean distance between the farmland boundary point in the target farmland boundary point set and the farmland entrance and exit position information; and the first association relationship is the association relationship between the target farmland boundary point set and the farmland entrance and exit position information; Based on the road center line pixel set, the node information of each road segment corresponding to the farmland area to be mapped is determined; The third minimum Euclidean distance is obtained, and if it is determined that the third minimum Euclidean distance is less than the third preset Euclidean distance, the second association relationship is established; wherein the third minimum Euclidean distance is the minimum Euclidean distance between the farmland entrance and exit position information and the node information of the road segment; and the second association relationship is the association relationship between the farmland entrance and exit position information and the node information of the road segment; Based on the intersection between the first association relationship and the second association relationship, the ternary topological connection association relationship between the farmland point, the entrance and exit and the farmland road segment is generated.

[0062] In the present application, for each farmland boundary point in the target farmland boundary point set, the Euclidean distance between it and each entrance and exit candidate point in the entrance and exit candidate point set is calculated respectively, and then the minimum value in all the Euclidean distances corresponding to each farmland boundary point is found, and the minimum value in these minimum values is the first minimum Euclidean distance, which reflects the closest distance in space between the target farmland boundary point set and the entrance and exit candidate point set.

[0063] In the present application, a first preset Euclidean distance is set in advance, and the entrance and exit candidate points in the entrance and exit candidate point set whose first minimum Euclidean distance with the target farmland boundary point set is less than or equal to the first preset Euclidean distance are determined as potential entrance and exit nodes. These potential entrance and exit nodes are the entrance and exit candidate points that are relatively close to the farmland boundary in space and are likely to be actual farmland entrances and exits.

[0064] Further, clustering processing is performed on all the potential entrance and exit nodes, and through the clustering processing, the potential entrance and exit nodes that are close in distance can be classified into a class. According to the clustering processing result, the centroid corresponding to each cluster is calculated, and the centroid can be regarded as the average position of all the nodes in the cluster and represents the central position of the cluster. In the present application, the centroid position information of each cluster is determined as the target entrance and exit position information, and these information includes farmland entrance and exit position information and road entrance and exit position information, which correspond to the connection points between the farmland and the outside world and the connection points between the road and the outside world respectively.

[0065] In the present application, for each farmland boundary point in the target farmland boundary point set, the Euclidean distance between it and the farmland entrance and exit position information is calculated, and then the minimum value in all the Euclidean distances is found, which is the second minimum Euclidean distance, measuring the spatial closeness between the target farmland boundary point and the farmland entrance and exit position. In the present application, a second preset Euclidean distance is set in advance, and if the second minimum Euclidean distance is less than the second preset Euclidean distance, it indicates that the target farmland boundary point and the farmland entrance and exit position are close enough in space, and it can be considered that there is a correlation between them, thereby establishing a first correlation relationship. The correlation relationship reflects the corresponding relationship between the target farmland boundary point set and the farmland entrance and exit position information.

[0066] In the present application, the road centerline pixel set contains the coordinate information of each pixel point on the road centerline. Through analysis and processing of the road centerline pixel set, the node information of each road segment corresponding to the farmland to be mapped can be determined. These node information can represent the key positions of the road, such as the starting point, the ending point and the turning point of the road, etc., which are used to describe the topological structure of the road.

[0067] For each position point in the farmland entrance and exit position information, the present application calculates the Euclidean distance between each position point and each node in the road segment node information, and then finds the minimum value, which is the third minimum Euclidean distance, to measure the spatial proximity between the farmland entrance and exit position and the road segment node. In the present application, a third preset Euclidean distance is preset. If the third minimum Euclidean distance is less than the third preset Euclidean distance, it indicates that the farmland entrance and exit position and the road segment node are close enough in space, and it is considered that there is a correlation between them, so as to establish a second correlation relationship, which reflects the corresponding relationship between the farmland entrance and exit position information and the road segment node information.

[0068] In the present application, the first correlation relationship describes the association between the target farmland boundary point set and the farmland entrance and exit position information, and the second correlation relationship describes the association between the farmland entrance and exit position information and the road segment node information. By finding the intersection of the two correlation relationships, a ternary topological connection correlation relationship between the farmland point (target farmland boundary point set), the entrance and exit (farmland entrance and exit position information) and the farmland road segment (road segment node information) can be obtained. This correlation relationship completely describes the spatial connection relationship between the farmland, the entrance and exit and the road, and provides accurate data for subsequent map construction, path planning and farmland management.

[0069] Figure 3 The construction process of the ternary topological connection correlation relationship provided by the present application is shown in the schematic diagram Figure 3 As shown, after two-dimensional projection of the input road point cloud, the corresponding density map is generated and binarization processing is performed, and the binarization image is subjected to closing operation and other processing to remove noise; then, the target binarization image is subjected to skeleton processing to construct a road center point undirected graph.

[0070] Further, let the target farmland boundary point set be F k For each point in the entrance and exit candidate point set v ∈ V cand Calculate the minimum Euclidean distance, which is the first minimum Euclidean distance , and the formula is: ; Wherein, V cand represents the entrance and exit candidate point set, represents the point in the target farmland boundary point set.

[0071] When , the point v in the entrance and exit candidate point set is identified as a potential entrance and exit node, and the set of potential entrance and exit nodes is , wherein, represents a first preset Euclidean distance.

[0072] Further, the set of potential access nodes application DBSCAN clustering, merging similar candidate points, taking the center of each cluster As the target access location information, including farmland access location information and road access location information.

[0073] Then, for the target farmland boundary point set F k The point and the target access location Calculate the second minimum Euclidean distance If , the association is established , that is, the first association relationship, wherein, represents a second preset Euclidean distance.

[0074] Further, the road skeleton graph corresponding to the road center line pixel set is divided into multiple road segments, and the set of node information of each road segment is obtained R ℓ}, and the minimum distance from the target access location to each road segment node information set, that is, the third minimum Euclidean distance, if the third minimum Euclidean distance ≤ third preset Euclidean distance d r , the association , that is, the second association relationship.

[0075] Preferably, according to the intersection between the first association relationship and the second association relationship, a ternary topological logical relationship set is formed, that is, the ternary topological connection association relationship between the farmland point, the access and the farmland road segment: .

[0076] The application combines road skeletonization with undirected graph analysis, can accurately identify farmland access and road intersection, and construct ternary topological logic of farmland, access and road segment, providing support for intelligent path planning and agricultural machinery decision planning.

[0077] On the basis of the above embodiment, according to the ternary topological connection association relationship, a high-precision map file corresponding to the farmland area to be mapped is constructed Based on the ternary topological connection association relationship, farmland ID information, access coordinate information and road segment ID information corresponding to the farmland area to be mapped are generated; According to the farmland ID information, the entrance and exit coordinate information and the road segment ID information, a KML file corresponding to the to-be-mapped farm area is constructed to generate the high-precision map file according to the KML file.

[0078] In a farm area, there can be multiple independent farmlands. In order to accurately distinguish and identify each farmland in the map, a unique identifier, i.e., farmland ID, needs to be assigned to each farmland. Based on the ternary topological connection association relationship obtained in the above embodiment, it can be clearly determined which point cloud data or area belongs to the same farmland, and then the corresponding farmland ID information of each farmland is generated. For example, different IDs can be assigned according to the geographical position, area size or semantic segmentation of the farmland, ensuring that each farmland has a corresponding identifier in the map.

[0079] Entrances and exits are important nodes in a farm that connect different areas, such as farmland and roads. According to the ternary topological connection association relationship, the present application can determine the position of the entrances and exits in three-dimensional space. By calculating the coordinate values of these entrance and exit points in a specific coordinate system (such as a geographic coordinate system or a local coordinate system), the entrance and exit coordinate information is obtained. These coordinate information accurately locates the position of the entrances and exits, for example, in automatic driving of agricultural machinery, the agricultural machinery needs to accurately know the position of the entrances and exits in order to smoothly enter and exit the farmland.

[0080] Further, the roads in a farm are usually composed of multiple road segments, each of which can have different attributes (such as length, width and material, etc.). In order to accurately represent and manage these road segments in the map, a unique ID needs to be assigned to each road segment. According to the ternary topological connection association relationship, different road segments can be identified and corresponding road segment ID information can be generated for them. In this way, each road segment can be individually labeled and processed in the map, such as setting different colors, line types, etc. to distinguish different types of road segments.

[0081] KML (Keyhole Markup Language) is a file format based on XML syntax standard, which is used to describe and save geographic information data. According to the generated farmland ID information, entrance and exit coordinate information and road segment ID information, the information is organized according to the syntax rules of KML file to construct KML file. In the KML file, different tags and elements can be used to represent geographic features such as farmland, entrance and exit and road segment, and their attributes (such as name, color, style, etc.) can be set. Then, the constructed KML file is imported into the corresponding map software or tool, and these software and tools will display the geographic features such as farmland, entrance and exit and road segment in a visual way according to the information in the KML file, to generate high-precision map file. The generated high-precision map file can be used in various application scenarios, such as farmland management in precision agriculture, path planning for automatic driving of agricultural machinery, visualization of farm resources, etc. Users can zoom in, pan, query and other operations on the map through the map software, to intuitively understand the geographic information of the farm area. Figure 4 The schematic diagram of the high-precision map of the farm provided by the present application is shown in Figure 4 , wherein, Figure 4 The serial numbers 1 to 17 in

[0082] On the basis of the above-mentioned embodiment, the point cloud semantic segmentation of the target farm three-dimensional point cloud data to obtain the first point cloud semantic segmentation result and the second point cloud semantic segmentation result comprises: The trained point cloud semantic segmentation model is used to perform point cloud semantic segmentation on the target farm three-dimensional point cloud data to obtain the first point cloud semantic segmentation result and the second point cloud semantic segmentation result; The trained semantic segmentation model is obtained by the following steps: According to the aerial image sample data of the farm area, sample farm three-dimensional point cloud data is obtained; The sample farm three-dimensional point cloud data is labeled with corresponding semantic labels according to the surface feature categories, to obtain a high-precision point cloud data set of the farm, wherein the semantic labels at least include the farmland point cloud semantic label and the road point cloud semantic label; According to the high-precision point cloud data set of the farm, a pre-trained semantic segmentation model is trained to obtain the trained semantic segmentation model.

[0083] In the present application, the three-dimensional point cloud data of the target farm is input into the trained point cloud semantic segmentation model. The model analyzes and processes the input point cloud data, determines the semantic category of each point according to the characteristics of the point cloud (such as the spatial position, color, and reflection intensity of the point), and thus completes the point cloud semantic segmentation operation. After model processing, the first point cloud semantic segmentation result and the second point cloud semantic segmentation result are obtained.

[0084] In the present application, aerial image sample data of the farm area is used. The aerial image sample data is obtained by photographing and scanning the farm area using aerial photography equipment such as a camera carried by a drone or a laser radar. Then, corresponding three-dimensional point cloud data is extracted from the aerial image sample data. This process involves the fusion and processing of multi-source data, such as the registration and fusion of optical images and laser radar point cloud data, to obtain more accurate and rich three-dimensional point cloud data of the farm. In an embodiment, a multi-rotor drone is used to take aerial photographs along a set route, and the obtained aerial image sample data is imported into three-dimensional reconstruction software to generate sample farm three-dimensional point cloud data.

[0085] Then, the ground object categories in the sample farm three-dimensional point cloud data are labeled, and each point is assigned a corresponding semantic label. Each point in the point cloud data can be classified by manual or semi-automatic tools to determine which ground object category it belongs to. In the present application, the semantic labels at least include farmland point cloud semantic labels and road point cloud semantic labels, i.e., it is clear which points in the point cloud data belong to farmland and which points belong to road. In addition, it can also include labels of other ground object categories such as buildings and vegetation, depending on actual needs and application scenarios. In an embodiment, based on the cloudcompare point cloud labeling software, different ground object categories (farmland, road, vegetation, building, and others) in the farm point cloud are assigned semantic labels, and the labeling results are cross-checked and verified for consistency, thereby generating a unified directory structure and metadata file, and exporting them in a standardized storage format. The labeled point cloud data is divided into training set, validation set, and test set according to the proportion, and a reusable farm high-precision point cloud data set is constructed for subsequent model training.

[0086] Further, a pre-trained semantic segmentation model is selected as the basis. The pre-trained model has been trained on a large-scale data set and has learned some general feature representations, which can accelerate the subsequent training process and improve the generalization ability of the model. The pre-trained semantic segmentation model includes, but is not limited to, RandLA-Net, SCF-Net, PointTransformer, and BAFLAC, etc.

[0087] During the training process, the pre-trained semantic segmentation model will continuously adjust its parameters based on the input point cloud data and corresponding semantic labels to minimize the error between the predicted results and the true labels. Through multiple iterations of training, the model gradually learns the feature patterns of different ground object categories in the farm point cloud data. After sufficient training, a trained semantic segmentation model is obtained, which can be used for semantic segmentation of new target farm three-dimensional point cloud data.

[0088] The farm high-precision mapping system provided by the present application is described below, and the farm high-precision mapping system described below can be mutually corresponding to the farm high-precision mapping method described above.

[0089] Figure 5 The structure diagram of the farm high-precision mapping system provided by the present application is shown in Figure 5 The present application provides a farm high-precision mapping system, which comprises a three-dimensional point cloud acquisition module 501, a semantic segmentation module 502, a first processing module 503, a second processing module 504 and a mapping module 505. The three-dimensional point cloud acquisition module 501 is used to acquire target farm three-dimensional point cloud data corresponding to a farm area to be mapped. The semantic segmentation module 502 is used to perform point cloud semantic segmentation on the target farm three-dimensional point cloud data to obtain a first point cloud semantic segmentation result and a second point cloud semantic segmentation result. The first point cloud semantic segmentation result is point cloud data with farmland point cloud semantic labels. The second point cloud semantic segmentation result is point cloud data with road point cloud semantic labels. The first processing module 503 is used to perform boundary point extraction processing on the first point cloud semantic segmentation result based on farmland point cloud curvature to obtain a target farmland boundary point set. The second processing module 504 is used to determine an entrance candidate point set according to the node degree of each pixel point in the road center line pixel set in the undirected graph, and generate a ternary topological connection association relationship between farmland points, entrances and farmland road segments according to the distance between farmland boundary points in the target farmland boundary point set and entrance candidate points in the entrance candidate point set. The road center line pixel set is constructed based on the second point cloud semantic segmentation result. The mapping module 505 is used to construct a high-precision map file corresponding to the farm area to be mapped according to the ternary topological connection association relationship.

[0090] The farm high-precision mapping system provided by the application obtains two part results with farmland and road point cloud semantic labels by performing semantic segmentation on the three-dimensional point cloud data of the farmland region to be mapped; then, boundary point sets are extracted from the farmland label point cloud according to the curvature of the farmland point cloud; at the same time, road center line pixel sets are constructed based on the road label point cloud, and entrance candidate point sets are determined by analyzing the node degree of the pixel points in the undirected graph; finally, the distance between the farmland boundary points and the entrance candidate points is calculated, a ternary topological connection correlation is generated, and a high-precision map file of the farm is constructed accordingly, thereby effectively improving the precision of the high-precision map of the farm.

[0091] The system provided by the embodiments of the application is used to execute the above-mentioned method embodiments, and the specific process and detailed content are referred to the above-mentioned embodiments, which will not be described here.

[0092] Figure 6 The structure schematic diagram of the electronic device provided by the application is shown as Figure 6 The electronic device can include a processor (Processor) 601, a communication interface (Communications Interface) 602, a memory (Memory) 603 and a communication bus 604, wherein the processor 601, the communication interface 602 and the memory 603 complete mutual communication through the communication bus 604. The processor 601 can call the logic instructions in the memory 603 to execute the farm high-precision mapping method, which includes: obtaining target farmland three-dimensional point cloud data corresponding to a farmland region to be mapped; performing point cloud semantic segmentation on the target farmland three-dimensional point cloud data to obtain a first point cloud semantic segmentation result and a second point cloud semantic segmentation result, wherein the first point cloud semantic segmentation result is point cloud data with farmland point cloud semantic labels; the second point cloud semantic segmentation result is point cloud data with road point cloud semantic labels; based on the curvature of the farmland point cloud, boundary point extraction processing is performed on the first point cloud semantic segmentation result to obtain a target farmland boundary point set; according to the node degree of the corresponding nodes of each pixel point in the undirected graph in the road center line pixel set, the entrance candidate point set is determined, and according to the distance between the farmland boundary points in the target farmland boundary point set and the entrance candidate points in the entrance candidate point set, a ternary topological connection correlation between farmland points, entrances and farmland road segments is generated, wherein the road center line pixel set is constructed based on the second point cloud semantic segmentation result; according to the ternary topological connection correlation, a high-precision map file corresponding to the farmland region to be mapped is constructed.

[0093] In addition, the logic instructions in the memory 603 described above can be implemented in the form of a software function unit and sold or used as an independent product, which can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the part that contributes to the prior art or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, including a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in various embodiments of the present application. The foregoing storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0094] In another aspect, the present application also provides a computer program product, which comprises a computer program stored on a non-transitory computer readable storage medium, and the computer program comprises program instructions, when the program instructions are executed by a computer, the computer can execute the farm high-precision mapping method provided by the above-mentioned method, which comprises: obtaining target farm three-dimensional point cloud data corresponding to a to-be-mapped farm area; performing point cloud semantic segmentation on the target farm three-dimensional point cloud data to obtain a first point cloud semantic segmentation result and a second point cloud semantic segmentation result, wherein the first point cloud semantic segmentation result is point cloud data with farmland point cloud semantic labels; the second point cloud semantic segmentation result is point cloud data with road point cloud semantic labels; based on the farmland point cloud curvature, performing boundary point extraction processing on the first point cloud semantic segmentation result to obtain a target farmland boundary point set; determining an entrance candidate point set according to the node degree of each pixel point in the road center line pixel set in the undirected graph, and generating a ternary topological connection association relationship between farmland points, entrances and farmland road segments according to the distance between farmland boundary points in the target farmland boundary point set and entrance candidate points in the entrance candidate point set, wherein the road center line pixel set is constructed based on the second point cloud semantic segmentation result; and constructing a high-precision map file corresponding to the to-be-mapped farm area according to the ternary topological connection association relationship.

[0095] In yet another aspect, the present application also provides a non-transitory computer-readable storage medium having stored thereon a computer program, which, when executed by a processor, implements the farm high-precision mapping method provided by the above embodiments, and the method comprises: obtaining target farm three-dimensional point cloud data corresponding to a to-be-mapped farm region; performing point cloud semantic segmentation on the target farm three-dimensional point cloud data to obtain a first point cloud semantic segmentation result and a second point cloud semantic segmentation result, wherein the first point cloud semantic segmentation result is point cloud data with farmland point cloud semantic labels; the second point cloud semantic segmentation result is point cloud data with road point cloud semantic labels; performing boundary point extraction processing on the first point cloud semantic segmentation result based on farmland point cloud curvature to obtain a target farmland boundary point set; determining an entrance candidate point set according to the node degree of each pixel point in the road center line pixel set in the undirected graph, and generating a ternary topological connection association relationship between farmland points, entrances and farmland road segments according to the distance between farmland boundary points in the target farmland boundary point set and entrance candidate points in the entrance candidate point set, wherein the road center line pixel set is constructed based on the second point cloud semantic segmentation result; and constructing a high-precision map file corresponding to the to-be-mapped farm region according to the ternary topological connection association relationship.

[0096] The device embodiments described above are only schematic, wherein the units illustrated as separate components can or can not be physically separate, and the components illustrated as units can or can not be physical units, i.e., can be located in one place or distributed on a plurality of network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the present embodiment. Those skilled in the art can understand and implement without creative labor.

[0097] From the above description of the embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software plus necessary universal hardware platforms, and of course, can also be realized by hardware. Based on such understanding, the above technical solutions, essentially or in other words, the part that contributes to the prior art can be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in the various embodiments or some parts of the embodiments.

[0098] It should be pointed out finally that the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit the same; and although the present application has been described in detail with reference to the foregoing embodiments, it should be appreciated by those skilled in the art that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features thereof can be replaced equivalently; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for making high-precision farm maps, characterized in that: include: Obtain the three-dimensional point cloud data of the target farm corresponding to the farm area to be mapped; Performing point cloud semantic segmentation on the three-dimensional point cloud data of the target farm to obtain a first point cloud semantic segmentation result and a second point cloud semantic segmentation result, wherein the first point cloud semantic segmentation result is point cloud data with a farmland point cloud semantic label; and the second point cloud semantic segmentation result is point cloud data with a road point cloud semantic label; Based on the curvature of the farmland point cloud, performing boundary point extraction processing on the semantic segmentation result of the first point cloud to obtain a target farmland boundary point set; Determining an entrance and exit candidate point set based on the node degree of each pixel point in the road centerline pixel set corresponding to a node in an undirected graph, and generating a ternary topological connection association relationship between the farmland point, the entrance and exit, and the farmland road segment based on the distance between the farmland boundary point in the target farmland boundary point set and the entrance and exit candidate point in the entrance and exit candidate point set, wherein the road centerline pixel set is constructed based on the semantic segmentation result of the second point cloud; According to the ternary topological connection association relationship, a high-precision map file corresponding to the farm area to be mapped is constructed.

2. The farm high-precision map making method according to claim 1, characterized in that: The step of performing boundary point extraction processing on the first point cloud semantic segmentation result based on the farmland point cloud curvature to obtain a target farmland boundary point set includes: Searching a neighborhood within a preset radius of the KD-Tree data structure, and calculating the curvature of the farmland point cloud corresponding to each point in the first point cloud semantic segmentation result; According to the curvature of the farmland point cloud and a preset curvature threshold, performing corresponding weighted mean filtering on each point in the first point cloud semantic segmentation result to obtain filtered point cloud data; Based on the region growing algorithm of the normal vector angle and the curvature, the filtered point cloud data is segmented into farmland plots to obtain a plurality of farmland plot point cloud data; Projecting each point in the point cloud data of the farmland plot onto a horizontal plane to obtain two-dimensional projection coordinate information corresponding to each point in the point cloud data of the farmland plot; Based on the two-dimensional projection coordinate information and the preset grid resolution, the points projected onto the horizontal plane in the point cloud data of each farmland plot are divided into regions, and an initial farmland boundary point set is constructed based on the first projection coordinate point and the second projection coordinate point in each region, wherein the first projection coordinate point is the point corresponding to the maximum two-dimensional projection coordinate information in each region; and the second projection coordinate point is the point corresponding to the minimum two-dimensional projection coordinate information in each region; Based on the α-Shape algorithm, filtering is performed on the boundary points in the initial farmland boundary point set to obtain a pending boundary point set; Based on any boundary point in the pending boundary point set and a nearest neighbor greedy strategy, the boundary points in the pending boundary point set are sorted to obtain a target farmland boundary point set.

3. The farm high-precision map making method according to claim 1, characterized in that: The step of determining the entrance and exit candidate point set according to the node degree of the corresponding node of each pixel point in the road centerline pixel set in the undirected graph includes: Projecting the points in the second point cloud semantic segmentation result onto a horizontal plane to obtain a two-dimensional projection coordinate set and a projection plane corresponding to the second point cloud semantic segmentation result; Based on a preset grid size, dividing the projection plane into cells to obtain a plurality of cells; Calculating the grid point density corresponding to each of the cells according to the two-dimensional projection coordinate set, and constructing a target density map according to the grid point density; performing a binarization process on the target density map to obtain a binarized target density map, and performing a closing operation and a filling process on the binarized target density map to obtain a target binary image; Based on the skeletonization algorithm, the road centerline extraction process is performed on the target binary image to obtain the road centerline pixel set; Constructing an undirected graph of road center points based on the road center line pixel set, and calculating the node degree corresponding to each node in the undirected graph of road center points; According to the node degree, the farmland entrance and exit points and the farmland road intersection points in the road center point undirected graph are determined, and the farmland entrance and exit points and the farmland road intersection points are used as the entrance and exit candidate points to obtain the entrance and exit candidate point set.

4. The farm high-precision map making method according to claim 1 or 3, characterized in that: The generating of a ternary topological connection association relationship between a farmland point, an entrance and an exit, and a farmland road segment according to the distance between the farmland boundary point in the target farmland boundary point set and the entrance and exit candidate points in the entrance and exit candidate point set includes: Obtaining a first minimum Euclidean distance, wherein the first minimum Euclidean distance is a minimum Euclidean distance between a farmland boundary point in the target farmland boundary point set and an entrance / exit candidate point in the entrance / exit candidate point set; Determine the entrance and exit candidate points in the entrance and exit candidate point set whose first minimum Euclidean distance is less than or equal to the first preset Euclidean distance as potential entrance and exit nodes; Performing clustering processing on all the potential entrance and exit nodes, and determining target entrance and exit location information based on the centroid corresponding to each cluster in the obtained clustering processing results, wherein the target entrance and exit location information includes farmland entrance and exit location information and road entrance and exit location information; Obtaining a second minimum Euclidean distance, and if it is determined that the second minimum Euclidean distance is less than a second preset Euclidean distance, establishing a first association relationship; wherein the second minimum Euclidean distance is the minimum Euclidean distance between a farmland boundary point in the target farmland boundary point set and the farmland entrance and exit location information; and the first association relationship is the association relationship between the target farmland boundary point set and the farmland entrance and exit location information; Determining the node information of each road segment corresponding to the farm area to be mapped based on the road centerline pixel set; Obtaining a third minimum Euclidean distance, and if it is determined that the third minimum Euclidean distance is less than a third preset Euclidean distance, establishing a second association relationship; wherein the third minimum Euclidean distance is the minimum Euclidean distance between the farmland entrance and exit location information and the road segment node information; and the second association relationship is the association relationship between the farmland entrance and exit location information and the road segment node information; Based on the intersection between the first association relationship and the second association relationship, the ternary topological connection association relationship between the farmland point, the entrance and exit, and the farmland road segment is generated.

5. The farm high-precision map making method according to claim 1, characterized in that: The high-precision map file corresponding to the farm area to be mapped is constructed based on the ternary topological connection association relationship. Based on the ternary topological connection association relationship, generating farmland ID information, entrance and exit coordinate information, and road segment ID information corresponding to the farm area to be mapped; According to the farmland ID information, the entrance and exit coordinate information and the road segment ID information, a KML file corresponding to the farm area to be mapped is constructed, so as to generate the high-precision map file according to the KML file.

6. The farm high-precision map making method according to claim 1, characterized in that: The performing point cloud semantic segmentation on the target farm three-dimensional point cloud data to obtain a first point cloud semantic segmentation result and a second point cloud semantic segmentation result includes: Performing point cloud semantic segmentation on the three-dimensional point cloud data of the target farm using the trained point cloud semantic segmentation model to obtain the first point cloud semantic segmentation result and the second point cloud semantic segmentation result; The trained semantic segmentation model is obtained by training through the following steps: Based on the aerial image sample data of the farm area, obtain the three-dimensional point cloud data of the sample farm; Semantic labels corresponding to the feature category labels in the three-dimensional point cloud data of the sample farm are assigned to obtain a high-precision point cloud dataset of the farm, wherein the semantic labels at least include the semantic labels of the farmland point cloud and the semantic labels of the road point cloud; The pre-trained semantic segmentation model is trained according to the farm high-precision point cloud dataset to obtain the trained semantic segmentation model.

7. A high-precision farm map making system, characterized in that: include: A 3D point cloud acquisition module is used to obtain the 3D point cloud data of the target farm corresponding to the farm area to be mapped; a semantic segmentation module, configured to perform point cloud semantic segmentation on the three-dimensional point cloud data of the target farm to obtain a first point cloud semantic segmentation result and a second point cloud semantic segmentation result, wherein the first point cloud semantic segmentation result is point cloud data with a farmland point cloud semantic label; and the second point cloud semantic segmentation result is point cloud data with a road point cloud semantic label; A first processing module is configured to perform boundary point extraction processing on the semantic segmentation result of the first point cloud based on the curvature of the farmland point cloud to obtain a target farmland boundary point set; a second processing module, configured to determine an entrance and exit candidate point set based on the node degree of a corresponding node in an undirected graph for each pixel point in the road centerline pixel set, and generate a ternary topological connection association relationship between the farmland point, the entrance and exit, and the farmland road segment based on the distance between the farmland boundary point in the target farmland boundary point set and the entrance and exit candidate point in the entrance and exit candidate point set, wherein the road centerline pixel set is constructed based on the semantic segmentation result of the second point cloud; A mapping module is used to construct a high-precision map file corresponding to the farm area to be mapped based on the ternary topological connection association relationship.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, it implements the farm high-precision map making method as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the high-precision farm map making method as described in any one of claims 1 to 6 is implemented.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the high-precision farm map making method as described in any one of claims 1 to 6 is implemented.

Images