Mulberry field autonomous navigation method and system based on laser radar point cloud

Abstract

The invention relates to the technical field of automatic navigation, in particular to a mulberry field autonomous navigation method and system based on laser radar point cloud. The method comprises the following steps: preprocessing a point cloud set, and performing semantic segmentation by using a semantic segmentation network model to obtain a point cloud feature comprising a point cloud label corresponding to the point cloud set; the point cloud label comprises a first label associated with the path area and a second label associated with the crop area; acquiring a crop area adjacent to the path area, and judging whether the crop area is a continuous space or a discrete space based on a second label; if the crop region is a discrete space, generating a candidate frame for a point cloud set of the crop region based on the point cloud features, and selecting a point of interest from the candidate frame; if the crop region is a continuous space, acquiring a boundary point cloud of the crop region as a point of interest; presetting a first threshold value, and obtaining a plurality of navigation points after shifting the plurality of points of interest towards the path area by the first threshold value; self-adaptive navigation line extraction of the mulberry field scene is realized, and the reliability of autonomous navigation of the mulberry field agricultural machinery is improved.

Description

Technical Field

[0001] The present invention relates to the field of automatic navigation technology, and in particular to a mulberry garden autonomous navigation method and system based on laser radar point cloud. Background Art

[0002] With the advancement of agricultural automation and intelligence, LiDAR technology is increasingly being used for autonomous navigation of agricultural machinery. However, mulberry orchards present certain unique characteristics compared to structured traffic scenarios. After summer pruning, mulberry trees are leafless, leaving only the trunks. In this state, the individual trunks are distinct. In contrast, in an unpruned orchard, the leaves are dense and lush, resulting in large point cloud data volumes and unclear boundaries between trunks. Therefore, selecting appropriate guidance line extraction methods based on the different conditions of the orchard can more accurately reflect the actual structure of the orchard and improve the reliability and efficiency of autonomous navigation.

[0003] The traditional visual extraction method of feature information such as crop rows and tractor paths has low work efficiency and is prone to human errors. Point cloud semantic segmentation models in the existing technology, such as PointNet++, mostly segment traffic scenes and lack relevant applications in agricultural mulberry garden scenes. The path navigation method disclosed in the application document with application number CN201880001094.0 determines the first path obstacle information by detecting the road surface based on the acquired depth map, combines the second path obstacle information determined by image learning based on the depth map, determines the third path obstacle information that better matches the actual road conditions, and performs navigation based on the third path obstacle information, making the navigation results more accurate.

[0004] However, the richness and accuracy of depth maps in three-dimensional spatial information are relatively low compared to lidar point cloud data. In a mulberry orchard with complex, dense mulberry trees and irregular terrain, navigation based solely on depth maps has difficulty accurately perceiving and understanding the orchard environment. Therefore, this application provides a method and system for autonomous navigation in a mulberry orchard based on lidar point clouds, which improves the reliability and efficiency of autonomous navigation of agricultural machinery in mulberry orchards, and enhances the efficiency and safety of mulberry orchard operations. Summary of the Invention

[0005] The purpose of the present invention is to address the technical problem in the existing technology that it is difficult to accurately perceive and understand the mulberry garden environment during the navigation process. Through the embodiments of this application, a mulberry garden autonomous navigation method and system based on lidar point cloud is provided, which realizes the recognition of mulberry garden scenes with unique growth cycle changes and obvious environmental characteristics. Through the adaptive navigation line extraction method, the reliability and efficiency of autonomous navigation of mulberry garden agricultural machinery are improved, damage to mulberry trees is avoided, and the efficiency and safety of mulberry garden operations are improved.

[0006] In order to achieve the above-mentioned object of the invention, the present invention provides the following technical solutions: A method for autonomous navigation in a mulberry orchard based on a laser radar point cloud, comprising: Collecting a point cloud set, and performing semantic segmentation on the preprocessed point cloud set using a semantic segmentation network model to obtain point cloud features corresponding to the point cloud set, including point cloud labels; the point cloud labels include a first label associated with a path area and a second label associated with a crop area; Acquire the crop area adjacent to the path area, and determine whether the corresponding crop area is a continuous space or a discrete space based on the second label; If the crop area is a discrete space, a candidate frame is generated for the point cloud set of the crop area based on the point cloud features, and points of interest are selected from the candidate frame; if the crop area is a continuous space, a boundary point cloud of the crop area is obtained as the point of interest; a first threshold is preset, and multiple points of interest are offset toward the path area by the first threshold to obtain multiple navigation points.

[0007] As a preferred technical solution of the present application, generating a candidate frame for the point cloud set of the crop area based on the point cloud features and selecting a point of interest from the candidate frame includes: Obtaining the point of interest in the candidate frame for representing the center position; deleting the point of interest whose distance to the mobile carrier is greater than or equal to the first threshold; Acquire multiple points of interest on both sides of the mobile carrier along the moving direction, and set indexes for the points of interest according to the moving direction; calculate the distance between the points of interest on both sides, and record the points of interest on both sides with the smallest distance as adjacent points; acquire adjacent points with index differences, and interpolate the adjacent points along the moving direction; and connect the points of interest on both sides and the interpolation points to obtain a reference line; The foot of the perpendicular between the interest point on one side and the reference line on the other side is obtained, the midpoint between the interest point and the foot of the perpendicular is calculated, and a navigation line is obtained by fitting a plurality of the midpoints.

[0008] As a preferred technical solution of the present application, obtaining the boundary point cloud of the crop area as the point of interest includes: Acquire a point cloud whose distance to the mobile carrier is greater than or equal to the first threshold and whose distance is the smallest as the point of interest; sequentially connect adjacent points of interest along the moving direction to obtain a reference line; The reference lines on both sides of the mobile carrier are obtained along the moving direction, multiple interval points are set on the reference lines according to a preset first interval, the midpoints of the lines connecting the adjacent interval points on both sides are calculated, and the navigation line is obtained after fitting the multiple midpoints.

[0009] As the preferred technical solution of the present application, the preprocessing includes: using the statistical filtering algorithm of the PCL point cloud library to remove point cloud outliers from the point cloud set, and using the straight-through filter algorithm of the PCL point cloud library to retain the point cloud within a preset area.

[0010] As a preferred technical solution of the present application, the semantic segmentation network model includes four SA modules for extracting the point cloud features, four upsampling layers for restoring the number of the point cloud features to the size of the point cloud set, and a classification fully connected layer for classifying the point cloud labels; The SA module includes a sampling layer, a grouping layer and a feature layer for randomly sampling point clouds; the grouping layer presets multiple different grouping radii to group the point cloud set; the feature layer is composed of 3 groups of dynamic edge convolutions to extract the point cloud features.

[0011] As a preferred technical solution of the present application, the semantic segmentation network model extracts features from the point cloud sets corresponding to different grouping radii through the SA module to obtain grouping features; and splices multiple grouping features to obtain local features; The local features are used through the dynamic edge convolution to extract edge features of the adjacent point cloud set, and the local features and the edge features are fused to obtain point cloud features; the point cloud features are upsampled to restore the number of point clouds in the point cloud set to the original number; and the point cloud set is classified through a fully connected layer.

[0012] As a preferred technical solution of the present application, the extracting edge features of adjacent point cloud sets by applying the dynamic edge convolution to the local features includes: Deleting the feature corresponding to the center point from the grouping feature to obtain a new grouping feature; splicing multiple new grouping features to obtain the local feature; A weight score matrix is ​​generated according to the coordinates of the point cloud set; the local features and the preset fully connected layer are multiplied to obtain a weight parameter matrix, and the local features and the weight parameter matrix are multiplied to obtain a feature matrix; the score matrix and the feature matrix are multiplied to associate the coordinates of the point cloud set with the weight parameter matrix.

[0013] As a preferred technical solution of the present application, the step of restoring the number of the point cloud features to the size of the point cloud set includes: Obtaining the point cloud features of the previous upsampling layer as the points to be interpolated, and obtaining the point cloud features of the next upsampling layer as the reference points; Traversing the point to be interpolated, selecting multiple reference points with the smallest distance to the point to be interpolated, and calculating the sum of the reciprocals of the multiple distances; using the ratio of the reciprocal of the distance to the sum of the reciprocals as the weight of the reference point; calculating the product of the reference point and the weight, and using the average of the products corresponding to the multiple reference points as the new point cloud feature of the point to be interpolated; splicing the features before and after interpolation to obtain the point cloud feature with increased dimension.

[0014] As the preferred technical solution of this application, the semantic segmentation network model adopts the method of transfer learning, firstly performs multiple rounds of training on a public data set, and verifies it using a public data set, calculates the average intersection-over-union ratio of multiple categories of the public data set, and screens out the semantic segmentation network model corresponding to the largest average intersection-over-union ratio; Establish a first data set of mulberry garden scenes, use the training set to perform multiple rounds of training on the semantic segmentation network model, calculate the average intersection-over-union ratio of multiple categories of the first data set, and screen out the semantic segmentation network model corresponding to the largest average intersection-over-union ratio as the final semantic segmentation network model.

[0015] The present invention provides a mulberry garden autonomous navigation system based on laser radar point cloud, comprising: a semantic segmentation module for collecting a point cloud set and performing semantic segmentation on the preprocessed point cloud set using a semantic segmentation network model to obtain point cloud features corresponding to the point cloud set, including point cloud labels; the point cloud labels include a first label associated with the path area and a second label associated with the crop area; an area recognition module, configured to obtain the crop area adjacent to the path area, and determine, based on the second label, whether the corresponding crop area is a continuous space or a discrete space; A path generation module is configured to generate a candidate frame for the point cloud set of the crop area based on the point cloud features and select points of interest from the candidate frame if the crop area is a discrete space; and to obtain a boundary point cloud of the crop area as the points of interest if the crop area is a continuous space; and to preset a first threshold, offset a plurality of the points of interest toward the path area by the first threshold to obtain a plurality of navigation points, so that the mobile carrier moves along the plurality of navigation points.

[0016] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention projects the point cloud into a polar coordinate system to form a bird's-eye view, and uses the semantic segmentation model for the mulberry garden scene to improve the semantic segmentation model. After performing semantic segmentation based on the bird's-eye view under polar coordinates, the semantic segmentation result is mapped back to the Cartesian coordinate system, effectively solving the problem of uneven distribution of point clouds in the mulberry garden scene leading to uneven distribution of point cloud labels, and improving the precision and real-time performance of semantic segmentation. In the special scene of mulberry garden, the point cloud representation under the polar coordinate system can better adapt to the obvious difference between dense and sparse areas of mulberry trees, so that the semantic segmentation model provided by the present invention can more accurately extract the features of semantic categories such as mulberry trees and ground, providing a more reliable data basis for subsequent autonomous navigation perception, and significantly improving the reliability and efficiency of autonomous navigation.

[0017] 2. The BEV_Mnet backbone network uses 4 consecutive SA modules to extract point cloud features. It uses a multi-scale grouping method. After random farthest point sampling, it sets multiple grouping radii. Then, it extracts features for the point cloud sets in each group. Finally, it splices the point cloud features extracted under different grouping radii to obtain the point cloud features output by each SA module. Each SA module contains a fully connected layer MLP consisting of 3 groups of DEdgeConv convolutions. The present invention extracts point cloud features through dynamic convolution DEdgeConv. The dynamic process is determined by the position of the point cloud, and a weight score matrix for regulating the convolution kernel parameters is generated. The output result is obtained through weighted operations, which is conducive to more efficient processing of irregularly distributed and disordered point cloud data.

[0018] 3. The autonomous navigation method provided by the present invention can select a suitable navigation line extraction method according to the different states of the mulberry garden, including summer pruning and unpruned. After summer pruning, the semantic segmentation network model BEV_Mnet can be used to identify the mulberry tree trunks on the left and right sides of the mechanized farming path, generate candidate frames for each trunk, and then extract navigation lines; when unpruned, navigation lines can be extracted based on the boundaries of the left and right tree rows. This way of adaptively selecting the navigation line extraction method according to different states can more accurately reflect the actual structure of the mulberry garden and improve the adaptability and reliability of autonomous navigation. The method can perceive the mulberry garden environmental information in real time and accurately, and update the navigation line in real time according to environmental changes, so that the autonomous navigation system can better adapt to the complex environmental changes in the mulberry garden scene, effectively reduce mulberry tree damage, improve the efficiency and safety of mulberry garden operations, and promote the realization of automated operations in the mulberry garden. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 Schematic diagram of the structure of the mulberry garden autonomous navigation system provided for this application; Figure 2 A flow chart of the autonomous navigation method for a mulberry garden provided in this application; Figure 3Schematic diagram of the structure and semantic segmentation of the BEV_Mnet model; Figure 4 It is a structural diagram of the backbone network in the BEV_Mnet model; Figure 5 Schematic diagram of the process of performing dynamic edge convolution for the BEV_Mnet model; Figure 6 Schematic diagram of the process of generating dynamic convolution kernel; Figure 7 Schematic diagram of the process of generating candidate boxes for crop regions; Figure 8 Schematic diagram of the process of BEV_Mnet model training and application; Figure 9 This is a schematic diagram of autonomous navigation in a mulberry orchard after summer pruning. Figure 10 A schematic diagram of autonomous navigation in an unpruned mulberry garden. DETAILED DESCRIPTION

[0020] To make the purpose, technical solutions and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them.

[0021] Therefore, the following detailed description of the embodiments of the present invention is not intended to limit the scope of the claimed invention, but merely represents some embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort shall fall within the scope of protection of the present invention.

[0022] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features and technical solutions therein may be combined with each other.

[0023] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings.

[0024] In the description of the present invention, it should be noted that the terms "upper" and "lower" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, or the orientations or positional relationships in which the inventive product is typically placed when in use, or the orientations or positional relationships commonly understood by those skilled in the art. Such terms are intended solely to facilitate the description of the present invention and simplify the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limitations on the present invention. Furthermore, the terms "first" and "second" and the like are used solely for distinction and should not be construed as indicating or implying relative importance.

[0025] For example Figure 1 The present embodiment shows a mulberry orchard autonomous navigation system based on a laser radar point cloud, including: a semantic segmentation module for collecting a point cloud set and performing semantic segmentation on the preprocessed point cloud set using a semantic segmentation network model to obtain point cloud features corresponding to the point cloud set, including point cloud labels; the point cloud labels include a first label associated with the path area and a second label associated with the crop area; an area recognition module, configured to obtain the crop area adjacent to the path area, and determine, based on the second label, whether the corresponding crop area is a continuous space or a discrete space; A path generation module is configured to generate a candidate frame for the point cloud set of the crop area based on the point cloud features and select points of interest from the candidate frame if the crop area is a discrete space; and to obtain a boundary point cloud of the crop area as the points of interest if the crop area is a continuous space; and to preset a first threshold, offset a plurality of the points of interest toward the path area by the first threshold to obtain a plurality of navigation points, so that the mobile carrier moves along the plurality of navigation points.

[0026] The mulberry garden autonomous navigation system is used to execute a mulberry garden autonomous navigation method based on laser radar point cloud provided in this application, for example Figure 2 Shown, including: A mobile vehicle-mounted 3D LiDAR system was used to scan the mulberry orchard environment, collecting 3D LiDAR point clouds within the scene to form a point cloud. Mulberry orchards are often located in hilly and mountainous areas, with small, densely packed plots and varying elevations. The collected point cloud includes information such as mulberry trees, crop rows, ground, and soil slopes, resulting in a relatively dense point cloud. Because the point cloud is widely distributed throughout the measurement range, some points may be out of alignment, necessitating preprocessing of the point cloud.

[0027] Since the data package of the lidar includes the horizontal rotation angle , laser beam vertical angle and distance d. To present a 3D point cloud, the angles and distances in stereoscopic space are converted into x, y, and z coordinates in a Cartesian coordinate system, centered on the LiDAR. A Cartesian coordinate system is established with the center of the LiDAR as the origin, and the direction of movement of the LiDAR's moving carrier as the positive x-axis. A rectangular coordinate system is established using the right-hand rule. The LiDAR's rotation direction is 0 degrees, and it rotates 360 degrees clockwise. The conversion formula is as follows: ; Where d is the distance, is the vertical angle of the laser beam, is the horizontal rotation angle of the laser beam, and R is the plane radius of the laser point from the origin. The laser radar data packet also provides reflection intensity information to characterize the reflection intensity of the measured object, with a range of [1-255]. In this embodiment, the point cloud set obtained is P={x i ,y i , z i , r i}, where (x i ,y i , z i ) is the Cartesian coordinate of the point cloud data, r i is the reflection intensity.

[0028] For the 3D LiDAR point cloud dataset from the Mulberry Garden environment, the PCL point cloud library's statistical filtering algorithm was used to remove outliers from the dataset, while the PCL point cloud library's pass-through filter algorithm was used to retain point cloud data within a preset area. The statistical filtering algorithm involves selecting multiple point clouds closest to a reference point cloud in multiple directions around it, calculating the mean, standard deviation, and variance of the first distances between these point clouds and the reference point cloud. The difference between the first distance and the mean is then calculated, and outliers whose difference is greater than the product of the standard deviation and variance are removed.

[0029] For agricultural machinery operations in mulberry fields, the area of ​​greatest interest is typically in front of the machine. However, the original point cloud contains a 360° range of points, resulting in a large amount of data. The straight-through filtering algorithm sets thresholds for the coordinate axes (x, y, and z), defining minimum and maximum values ​​for each axis to determine a predefined region. The algorithm then iterates through each point in the point cloud, retaining those points whose coordinates fall within the predefined region.

[0030] The statistical filtering algorithm of the PCL point cloud library is used to remove outlier point clouds, and the pass-through filter algorithm of the PCL point cloud library is used to retain point clouds within the preset area, which helps to improve the quality of point cloud data before semantic segmentation of the point cloud set.

[0031] Each point cloud in the Cartesian coordinate system is projected onto a plane where the x-axis and y-axis are located from a bird's-eye view perspective to form a two-dimensional bird's-eye view. The two-dimensional bird's-eye view can use a two-dimensional convolutional network to extract point cloud features from the point cloud set, which can balance computational cost and accuracy. Then, with the origin of the laser radar coordinate system in the plane where the x-axis and y-axis are located as the center point, a polar coordinate system is established for the two-dimensional bird's-eye view, and the polar coordinate system is set according to the preset offset angle. As the step size, the polar coordinate system is divided into two equal parts At the same time, the radial direction with the coordinate origin as the center is adjusted according to the preset offset radius , dividing the point cloud into multiple concentric circles. Different angles and different concentric circles divide the point cloud in the polar coordinate system into multiple fan-shaped grids. Polar coordinates can balance the distribution of point clouds in different ranges, making it more likely for the semantic segmentation network model used for semantic segmentation of point cloud sets to learn and identify features at locations close to the lidar, thereby minimizing the information loss caused by feature extraction. i ,y i , z i , r i}Convert to the angle in polar coordinate system and distance , the calculation formula is as follows: ; Represent the point cloud set as , where K is the number of point cloud coordinate elements, that is, the dimension of the input feature, which is set to 4; N is the number of point clouds in the point cloud set. The corresponding different fan-shaped grids group the point cloud sets, and the point cloud sets in each fan-shaped grid are represented as , H and W are the dimensions of the two-dimensional bird's-eye view, is the number of point clouds in each sector grid.

[0032] The grouped point cloud set is input into a semantic segmentation network model for semantic segmentation, obtaining point cloud features corresponding to the point cloud set; the point cloud features include the point cloud labels and polar coordinates of the point cloud set. In this embodiment, the semantic segmentation network model is constructed as a BEV_Mnet model, comprising a backbone network composed of four SA (set abstraction) modules for feature extraction and four upsampling layers for restoring the number of features.

[0033] According to the point cloud features after semantic segmentation by the semantic segmentation network model, the corresponding point cloud is extracted from the point cloud set based on the preset point cloud labels and point cloud coordinates. The point cloud polar coordinates are mapped back to the Cartesian coordinate system; for each point cloud , get the corresponding polar coordinates The point cloud tag of the sector grid is assigned to the point cloud, and the Cartesian coordinates of the point cloud and the point cloud tag are combined into a new point cloud feature, which is used in the bin candidate box generation method to generate a candidate box for the crop area based on the second tag and the corresponding coordinates. In this embodiment, the preset point cloud tags include a first tag for representing the path, a second tag for representing the crop, and a third tag (tag value 0) for representing background information; wherein the point cloud set corresponding to the first tag is the path area, including the ground (tag value 40), and the point cloud set corresponding to the second tag is the crop area, including the left tree row (tag value 48) and the right tree row (tag value 51). The path area and the crop area are adjacent.

[0034] Select appropriate navigation line extraction methods according to different states of the mulberry garden. In order to promote the early maturity of mulberry leaves and increase the yield of spring leaves, modern mulberry gardens reserve 10%-15% of summer-cut mulberry trees during winter management. The invention provided by this application is based on the fact that the agricultural machinery operations in the mulberry garden are mainly in "management" and "harvesting". At this time, the state of the mulberry garden is mainly divided into two states: summer-cut pruning and unpruned. After summer-cut pruning, the mulberry trees only retain trunks no higher than 25 cm, with no mulberry leaves at all. The corresponding extracted features are discrete multiple point cloud sets, forming multiple discrete crop areas; when not pruned, the mulberry leaves of the mulberry trees can grow to 50~150 cm high, the mulberry leaves are lush and dense, and the trunks are almost invisible. The corresponding extracted features are point cloud sets with continuous distribution of point clouds, forming a single continuous crop area. The present invention determines whether the point cloud features extracted from the crop area are continuous or discrete based on whether summer-cut pruning is performed, and selects different navigation line extraction strategies. A point of interest is selected from the candidate box; a first threshold is preset, and multiple navigation points are obtained by offsetting multiple points of interest toward the path area by the first threshold. The mobile carrier moves along the multiple navigation points or the navigation line formed by connecting the multiple navigation points, and the calculated navigation line is fed back to the autonomous navigation system in real time. Combined with the vehicle's dynamic model and sensor data, autonomous navigation in the mulberry garden scene is realized.

[0035] In one embodiment, for example Figure 9As shown, when the mulberry orchard is in the summer felling pruning state, the mulberry trees have no leaves, only the trunks serve as the crop area, and the point cloud density is low. The semantic segmentation network model is used to identify the mulberry trunks on the left and right sides of the mechanized farming path, generate candidate boxes for the trunks, and extract the navigation line based on the midpoint of the line connecting the center points of the candidate boxes on both sides. It should be noted that the first threshold in the summer felling pruning state is used to retain the center position of the mulberry trunks on both sides of the current mechanized farming path. The first threshold can be the coordinate axis distance between the laser radar origin and the point cloud corresponding to the left boundary of the left tree row, or the point cloud corresponding to the right boundary of the right tree row; it can also be the distance on the coordinate axis corresponding to 1 / 2 of the mechanized farming path width and the sum of the tree row widths. In this embodiment, the first threshold is set to the coordinate axis unit length of 1, that is, through the coordinate unit length transformation, the coordinate axis distance represented by the first threshold is converted to unit length to reduce the amount of calculation. Specifically, the geometric center point of the generated candidate box is obtained as the point of interest, the center position of each mulberry tree trunk is extracted, and the center point of the mulberry trunk that deviates from the current mechanized farming path is removed. The method for removing deviated center points includes: using the mobile carrier as the origin, for the center points corresponding to multiple tree trunks on the left, deleting points with y-axis coordinates greater than 1; for the center points corresponding to the right tree trunk, deleting points with y-axis coordinates less than -1; and assigning index values ​​to the center points of the left and right trunks sequentially starting from 0 according to the moving direction of the mobile carrier. Calculating the distance between the left and right center points, and recording the center points with the smallest distance between the left and right sides as adjacent points. Determine whether the index values ​​of adjacent points are the same. If there is a difference in the index, add compensation center points at intervals of 0.1m along the moving direction for the left and right center points to make the line connecting the left and right center points along the moving direction smoother. The center points and the compensation center points are sequentially connected to obtain a reference line.

[0036] Count the number of left and right indexes. Since the more navigation points there are, the more accurate and smooth the navigation line of the mobile carrier is, therefore, draw a perpendicular line from the center point of the reference line on the side with more numbers to the center point of the reference line on the side with fewer numbers to obtain the foot point of the perpendicular; the midpoint between the center point and the foot point of the reference line on the side with more numbers is used as the navigation point. For example, if the number of right indexes is greater than that on the left, for each center point on the right, obtain its foot point on the left reference line, and the x-axis coordinate of the foot point is within the x-coordinate range of the current point cloud set; calculate the midpoint between the center point and the foot point on the right. Use multiple midpoints as navigation points, store the navigation points, interpolate between the navigation points at intervals of 0.05m, add compensation navigation points, and connect the navigation points and compensation navigation points in sequence to fit the navigation line.

[0037] In another embodiment, for example Figure 10As shown, when the mulberry orchard is unpruned, the leaves are dense and numerous, making it difficult to segment individual trunks. The semantic segmentation network model also struggles to generate candidate boxes for each trunk. In this case, the semantic segmentation network model identifies the tree rows on the left and right sides of the tractor path, using the rows as the crop areas. The navigation line is extracted along the tree row boundaries. It is important to note that the first threshold in the unpruned state is used to reduce the selection of noisy point clouds close to the lidar as boundary point clouds. The first threshold can be the coordinate axis distance between the radar origin and the point cloud corresponding to the right boundary of the left tree row or the point cloud corresponding to the left boundary of the right tree row; it can also be the coordinate axis distance corresponding to half the width of the tractor path. In this embodiment, the first threshold is set to a coordinate axis unit length of 1. This means that the coordinate axis distance represented by the first threshold is converted to a unit length through a coordinate unit length transformation to reduce computational complexity. Specifically, the boundary point clouds of the left and right tree rows are extracted from the point cloud features. Since mulberry trees in a row may have died of disease and been dug up, point clouds within the same row may be discontinuous. Therefore, the point cloud sets on the left and right sides are segmented every 5m along the direction of movement, and the index values ​​of the segments are marked in order according to the direction of movement starting from 0. After the left point cloud is segmented, the boundary points with the smallest y-axis coordinate and a y-axis coordinate greater than 1 in each segment are extracted as points of interest, and the boundary points of each segment are connected in order of index values ​​to form the left reference line. After the right point cloud is segmented, the boundary points with the largest y-axis coordinate and a y-axis coordinate less than -1 in each segment are extracted as points of interest, and the boundary points of each segment are connected in order of index values ​​to form the right reference line. Based on the points of interest on the left and right sides, fitting is performed along the direction of movement, interpolation is performed with a step size of 1m, compensation boundary points are added, and the points of interest and compensation boundary points are connected in sequence to obtain the reference lines on the left and right sides. For the reference lines on the left and right sides, interval points are set at intervals of 1m, the midpoints of the Euclidean distances of adjacent interval points in the two boundary lines are calculated, and multiple midpoints are used as navigation points. After sequential connection, the navigation line is fitted. By segmenting the tree rows and setting index values ​​for each segment, the algorithm for calculating navigation points is executed on long-distance mulberry tree rows, and then the navigation points are connected in index order to reduce navigation errors caused by discontinuous mulberry tree rows.

[0038] The semantic segmentation network model performs semantic segmentation on the point cloud, extracts point cloud features, and predicts corresponding point cloud labels, such as "road," "building," "left tree row," and "right tree row," expressed as numerical values. The final semantic segmentation result is obtained by converting the polar coordinate system back to the Cartesian coordinate system and reshaping the 2D semantic segmentation prediction back to 3D voxels to separate labels at different heights along the z-axis, where Z represents the number of voxel layers along the z-axis.

[0039] In one embodiment, the semantic segmentation network model is a deep learning-based semantic segmentation model BEV_Mnet model, including: a point cloud feature extraction network and a candidate box generation network, wherein the point cloud feature extraction network includes a backbone network and a segmentation network, for example Figure 3 Specifically, for example Figure 4 As shown in Figure 1, the backbone network includes four consecutive SA modules for feature encoding, which are used to extract multi-scale and high-dimensional point cloud features from point cloud sets. The SA (Set Abstraction) module includes a sampling layer for sampling point clouds, a grouping layer for grouping point cloud sets according to multiple preset grouping radii in different SA modules, and a feature layer for extracting point cloud features. Specifically, the SA module uses the farthest point sampling to randomly sample npoint point clouds in the original point cloud. The input point cloud dimension is , where C in is the number of input channels, that is, the number of point cloud coordinate elements, which is set to 4 in this embodiment; with the sampling point as the center, select S points for sampling in a sphere with a specified radius for grouping, and the point cloud dimension after grouping is . Since only a single grouping radius scale and a fixed number of sampling points are used, the grouped point cloud will be affected by the point cloud density; if the point cloud is too sparse, it will be difficult to group with a small radius; the subsequent SA layers are all based on the previous layer. If the first layer of SA does not extract features, the subsequent layer will not be able to extract the features of the corresponding sparse point cloud. Therefore, the present invention uses a multi-scale grouping (MSG) method to extract point cloud features. After random farthest point sampling, multiple grouping radii are set respectively, and then feature extraction is performed on the point cloud sets in each group respectively. Finally, the point cloud features extracted under different grouping radii are spliced ​​to obtain the point cloud features output by each SA module.

[0040] To better handle irregular and disordered point cloud data, e.g. Figure 5 As shown in the figure, the BEV_Mnet model of the present invention uses dynamic edge convolution DEdgeConv (Dynamic Edge Convolution) to extract point cloud features, which not only considers the features of the point itself, but also integrates the relative features of the surrounding point clouds, further focusing on the local features of the point cloud, making the point cloud features richer and more effective.

[0041] Specifically, the feature extraction part of each SA module in the BEV_Mnet backbone network contains a convolutional fully connected layer MLP composed of three groups of DEdgeConv convolutions. First, after multi-scale grouping, the edge convolution method is used to subtract the group center point (i.e., sampling point) feature from the point cloud feature within the group to obtain the point cloud relative center feature with a dimension of ,in, Represents the input point cloud dimension, S represents the number of grouped sampling points; then feature splicing is performed to splice the original point cloud features input to the SA module with the center relative features to obtain Dimensional point cloud features.

[0042] Furthermore, the dynamic convolution process of DEdgeConv to extract point cloud features includes: generating a weight score matrix for regulating the convolution kernel parameters according to the position of the point cloud, and obtaining the output result through weighted operation. Specifically, for example Figure 6 As shown, define a weight library , where each B q Is a weight parameter matrix. In the ScoreNet network, the convolution kernel weight score matrix is ​​generated according to the point cloud coordinate features, and its output dimension is , Q represents the number of convolution kernel weight parameters. The input of the ScoreNet network is the point cloud center point p i and its neighboring point p j The positional relationship between them is transformed nonlinearly by a multi-layer perceptron to obtain an intermediate feature representation. In order to ensure the reasonable distribution of scores, the output of the ScoreNet network is normalized by the Softmax function to obtain a normalized weight score matrix ,in Represents the qth weight parameter matrix B q The score of row i and column j when constructing the convolution kernel. Weight score matrix M ij Each element represents the weight parameter matrix B q In building the convolution kernel The weight score matrix M generated based on the ScoreNet network ij , the convolution kernel is dynamically assembled by the following formula: ; From the weight parameter matrix B q Weighted selection of the required convolution kernel parameters, weight parameter matrix B q The dimension is . Combine the spliced ​​point cloud features with the weight parameter matrix B q Multiplying together gives the dimension The feature matrix of , the weight score matrix is ​​multiplied by the feature matrix to get the weighted result, whose dimension is The weighting operation is performed on the Q dimension, which is equivalent to combining all parameters in the Q dimension into a single parameter of the convolution kernel through weighting. By multiplying the weight score matrix with the feature matrix and performing a weighted summation on the Q dimension, the convolution kernel parameters are associated with the point cloud coordinate positions. Different feature extraction methods are obtained for point cloud sets corresponding to different grouping radii.

[0043] Since the maximum pooling is only related to the maximum value and has nothing to do with the order, it conforms to the disordered characteristics of the point cloud. The maximum pooling operation is performed after each DEdgeConv dynamic convolution. The point cloud features of dimensions are output after maximum pooling The backbone network achieves deep feature extraction by connecting four consecutive SA modules in series. The output includes a point cloud feature list sa_features, a list of sampled point cloud coordinates, and a list of sampling point indices.

[0044] Specifically, the segmentation network is tasked with classifying each point cloud. Since the number of point clouds decreases after successive convolution operations of the SA module, the segmentation network needs to classify the original point cloud. Therefore, the segmentation network needs to upsample the point cloud features extracted by the SA module to restore the number of points in the point cloud feature to the original number, perform loss calculations, and iterate model training. The segmentation network of the BEV_Mnet model employs four layers of feature upsampling. Each layer uses interpolation to restore the number of points, and the interpolation depends on the features of the previous and next layers. The interpolation process involves using the first layer (M1 points) as the interpolated points and the second layer (M2 points) as the reference points. The distances between these M1 interpolated points and the M2 reference points are calculated. The first layer has M1 points and a feature dimension of D1, while the second layer has M2 points and a feature dimension of D2. Then, the k=3 closest reference points are selected from each of the M1 interpolated points. The weighted average of the features of these k points is used as the new feature of the interpolated point, with a dimension of D2. Each point to be interpolated will get a new feature, which comes from the weighted average of the features of the next layer of points. The weighting coefficient is equal to the inverse of the distance of each point divided by the sum of the inverses of the distances of k points. The closer the distance, the larger the weighting coefficient. Using interpolation can make the point cloud of the previous layer obtain features similar to those of the next layer. The features before and after interpolation are spliced ​​together, and an additional layer of input and output channels is added. The convolution of , keeps the feature dimension at D2. The point cloud features output by the BEV_Mnet model feature extraction network ultimately include two parts: upsampled point cloud coordinates and upsampled point cloud features.

[0045] Specifically, the point cloud feature extraction network extracts point cloud features through the backbone network, performs semantic segmentation on the point cloud features through the segmentation network, obtains point cloud features including point cloud label classification, and generates candidate boxes for the point cloud set after label classification of the original point cloud set based on the point cloud labels. In order to generate candidate boxes more efficiently and obtain a small number of candidate boxes with high recall rates, this paper takes the target point as the center and directly extracts the point cloud labels of each point cloud. When a point belongs to the target point, its corresponding candidate box is more likely to become a real candidate box. The bin candidate box generation method is used to re-plan the original space size according to the bin size as the unit length to generate the candidate box. The candidate box generation network process is as follows: Figure 7shown.

[0046] In one embodiment, the BEV_Mnet model uses transfer learning, e.g. Figure 8 As shown, multiple rounds of training are first performed using multiple public datasets, and the public labels with the highest intersection-over-union ratio are selected. The spatial attributes of the object categories corresponding to the public labels and the object categories to be segmented determined in the mulberry garden scene are compared, such as the length and / or width and / or height of the object categories, to obtain the corresponding BEV_Mnet model with the highest spatial attribute similarity; the BEV_Mnet model is then trained using the dataset collected from the mulberry garden scene to improve the classification accuracy of the semantic labels in the mulberry garden scene, and then the model corresponding to the largest average intersection-over-union ratio is selected as the semantic segmentation network model for the final application, and finally applied to the actual autonomous navigation scene in the mulberry garden.

[0047] In one specific embodiment, the public dataset uses the SemanticKITTI dataset. This invention uses the SemanticKITTI dataset's lidar point cloud data format. After manual annotation, each point cloud has a point cloud label. First, a SemKITTI_label_name dictionary is created, containing the names of all categories in the SemanticKITTI dataset and their corresponding label values. Then, all label values ​​are retrieved and sorted, removing label values ​​corresponding to background or invalid categories. In PyTorch, many operations (such as the cross-entropy loss function and label processing in classification tasks) require category labels to start at 0. Therefore, the label value index is converted to start at 0 to meet PyTorch's requirements.

[0048] The training process involves creating a packaged BEV_Mnet model, selecting the polar coordinate model as the model type, setting the grid size of the bird's-eye view BEV to [480, 360, 32], the training batch size to 2, the feature dimension to 9, the maximum number of points per SA module to 256, the output point cloud feature dimension to 512, the convolution kernel size to 1, random sampling of the point cloud, and feature compression to 32. At the beginning of each training epoch, the data is randomly shuffled, and four threads are used to accelerate data loading. An empty list, loss_list, is initialized to store the loss values ​​for each batch in the current epoch. A progress bar object, pbar, is created using the tqdm library to display training progress. The total length of the progress bar is set to the length of the training data loader, which is the total number of training data batches. Each batch of data in the training data accelerator is iterated over, using the try method to catch any exceptions. The point cloud of the training data is converted to the required format for training, and the model is forward propagated on the training data to obtain predictions. Calculate the prediction loss, including the Lovász-Softmax loss and cross-entropy loss. Set the ignore index to 255 to exclude losses for invalid classes or background. Backpropagate the loss values ​​and calculate the gradients. Update the model parameters using the Adam optimizer and add the current batch's loss values ​​to the loss_list. Clear the optimizer's gradients to prepare for the next iteration and update the progress bar to indicate the completion of training for one batch. If any exceptions occur during training, capture the exception information.

[0049] Set the validation interval to 150 and the batch size to 2 during validation. Check if the current global iteration count has reached the validation interval. If so, perform model validation. Set the model to evaluation mode. Do not shuffle the data during validation, preserving the data order. Initialize an empty list, hist_list, to store the confusion matrix for each batch during validation. Initialize an empty list, val_loss_list, to store the loss values ​​for each batch during validation. Gradients are not calculated during validation, which saves resources and increases validation speed. Iterate over each batch of data in the validation data loader, convert the validation data to the format required for training, and use the model to make predictions on the validation data to obtain predicted labels. Calculate the loss for the predictions, including the lovasz_softmax maximum loss and cross-entropy loss. To avoid calculating gradients during validation, use the detach method to detach the tensor from the computational graph. Convert the predicted logits to category indices by removing the maximum index at each position. Move the predictions from the GPU to the CPU and convert them to NumPy arrays. For each validation example, calculate the confusion matrix between the predicted results and the true labels. Quickly calculate the histogram between the predicted values ​​and the true labels, combine the predicted values ​​and the true labels into an integer to count the number of occurrences of each category pair, and thus construct a confusion matrix. After calculating the histogram, only the relevant parts of the desired category are retained and added to hits_list. The loss value of the current batch is added to val_loss_list. To evaluate the performance of the model, this paper uses the intersection over union (IoU) for evaluation, which is calculated as follows: ; Among them, TP i , FP i and FN i The number of true positives, false positives, and false negatives corresponding to category i, respectively, is the arithmetic mean of the IoUs, calculated as follows, where N is the total number of semantic categories: ; After verification is complete, the model is reset to training mode. The confusion matrices for all batches are summed to calculate the Intersection over Union (IoU) for each category. Each category name and corresponding IoU value is iterated over to calculate the verified mIoU. After verification, the IoU and mIoU for each category are printed. After verification, the variables used during verification are deleted to free up memory. The verified mIoU is checked to see if it is higher than the previously recorded best mIoU. If so, the best mIoU is updated to the current value and the current model weights are saved, ensuring that the best performing model is retained during training.

[0050] During the training process, the BEV_Mnet model was compared with other models to verify the effectiveness of the present invention. The inference time of the BEV_Mnet model and other control models provided by the present invention was evaluated. For example, as shown in Table 1, which reports the inference latency of a single frame of point cloud using the largest batch size that can be inferred on a single graphics card, PolarNet can infer a maximum of 5 frames of point cloud at a time on a single graphics card, while the present invention method can infer a maximum of 10 frames of point cloud at a time.

[0051] Table 1 Single-frame point cloud inference time tested on the SemanticKITTI dataset ; Table 2 shows the results of testing different network models on the SemanticKITTI dataset. Comparing the semantic segmentation visualization results, PolarNet easily confuses fences with vegetation and misclassifies terrain, roads, and parking spaces. The proposed method demonstrates higher accuracy in identifying roads and fences.

[0052] Table 2 Comparison of semantic segmentation results of SemanticKITTI dataset ; After training the optimal BEV_Mnet model based on the public SemanticKITTI dataset, the present invention established the Mulberry Planting dataset for mulberry orchards. The dataset was collected during the winter renovation, summer felling and pruning, and the normal growing seasons of spring and autumn. Multiple point cloud labels were assigned, including: pedestrian, building, mulberry tree trunk, fence, left-tree, right-tree, road, other agricultural machinery, and pole. The point cloud data was stored in a floating-point binary file format, with each data point represented by a four-digit hexadecimal number (floating point) and separated by a line feed. A point cloud data point consists of coordinates (x, y, z) and r (reflection intensity). The dataset contains 22 sequences of lidar point cloud data: sequences 00-07 and 09-10 are used for training, sequence 08 is used for validation, and sequences 11-21 are used for testing. The labeled data is stored as a txt file. The tags contain the labeled data for multiple training scenarios, and each labeled file contains multiple attributes. Table 3 shows the performance of different network models on the Mulberry Planting dataset. The proposed model achieves the highest IoU in identifying the tractor road, the left-tree row of mulberry trees, and the right-tree row of mulberry trees.

[0053] Table 3 Comparison of semantic segmentation results of MulberryPlanting dataset ; Based on the test results in Tables 2 and 3, it can be concluded that, when tested on the SemanticKITTI dataset, the semantic segmentation model BEV_Mnet based on deep learning achieved a mIoU exceeding 57.0%, and an IoU of 97.5% for road category recognition. In a real-world mulberry orchard scenario, tested on the Mulberry Planting dataset, the method was able to accurately identify the plowing road, the left-tree row of mulberry trees, and the right-tree row of mulberry trees, demonstrating its broad application value. In terms of mulberry orchard navigation, this method can accurately identify mulberry trees and road surfaces in real time, providing important data support for the refined management of mulberry orchards. In terms of autonomous navigation, this method can quickly locate the ground and plan a route in real time based on the direction of the tree rows on both sides, reducing mulberry tree damage. The method can also identify the location of mulberry trees, facilitating the accuracy of automated harvesting and reducing errors during the harvesting process.

[0054] Those skilled in the art will appreciate that all or some of the steps and systems in the method disclosed above can be implemented as software, firmware, hardware, and appropriate combinations thereof. Some physical components or all physical components can be implemented as software executed by a processor, such as a central processing unit, a digital signal processor, or a microprocessor, or implemented as hardware, or implemented as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, and the computer-readable medium can include computer storage media (or non-transitory media) and communication media (or temporary media). As known to those skilled in the art, the term computer storage media is included in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data) and is volatile and non-volatile, removable, and non-removable. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory, or other memory technology, CD-ROM, digital versatile disks (DVD), or other optical disk storage, magnetic cassettes, magnetic tapes, disk storage, or other magnetic storage devices, or any other medium that can be used to store desired information and can be accessed by a computer. Furthermore, as is well known to those skilled in the art, communication media typically embodies computer-readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media.

[0055] The above embodiments are only used to illustrate the present invention and are not intended to limit the technical solutions described in the present invention. Although this specification has described the present invention in detail with reference to the above embodiments, the present invention is not limited to the above specific implementation methods. Therefore, any modification or equivalent replacement of the present invention; and all technical solutions and improvements thereof that do not depart from the spirit and scope of the invention are included in the scope of the claims of the present invention.

Claims

1. A mulberry garden autonomous navigation method based on laser radar point cloud, characterized in that: include: Collecting a point cloud set, and performing semantic segmentation on the preprocessed point cloud set using a semantic segmentation network model to obtain point cloud features corresponding to the point cloud set, including point cloud labels; the point cloud labels include a first label associated with a path area and a second label associated with a crop area; Acquire the crop area adjacent to the path area, and determine whether the corresponding crop area is a continuous space or a discrete space based on the second label; If the crop area is a discrete space, generating a candidate box for the point cloud set of the crop area based on the point cloud features, and selecting a point of interest from the candidate box; If the crop area is a continuous space, obtaining a boundary point cloud of the crop area as the point of interest; A first threshold is preset, and a plurality of the points of interest are offset toward the path area by the first threshold to obtain a plurality of navigation points.

2. The method according to claim 1, wherein Generating a candidate frame for the point cloud set of the crop area based on the point cloud features, and selecting a point of interest from the candidate frame, comprises: Obtaining the point of interest in the candidate frame for representing the center position; deleting the point of interest whose distance to the mobile carrier is greater than or equal to the first threshold; Acquire multiple points of interest on both sides of the mobile carrier along the moving direction, and set indexes for the points of interest according to the moving direction; calculate the distance between the points of interest on both sides, and record the points of interest on both sides with the smallest distance as adjacent points; acquire adjacent points with index differences, and interpolate the adjacent points along the moving direction; and connect the points of interest on both sides and the interpolation points to obtain a reference line; The foot of the perpendicular of the reference line on the other side of the point of interest on one side is obtained, and the midpoint between the point of interest and the foot of the perpendicular is calculated to obtain the navigation point.

3. The method according to claim 1, wherein The obtaining of the boundary point cloud of the crop area as the point of interest includes: Acquire a point cloud whose distance to the mobile carrier is greater than or equal to the first threshold and whose distance is the smallest as the point of interest; sequentially connect adjacent points of interest along the moving direction to obtain a reference line; The reference lines on both sides of the mobile carrier are obtained along the moving direction, a plurality of interval points are set on the reference lines according to a preset first interval, and the midpoint of the line connecting the adjacent interval points on both sides is used as the navigation point.

4. The method according to any one of claims 1 to 3, wherein: The preprocessing includes: using the statistical filtering algorithm of the PCL point cloud library to remove point cloud outliers from the point cloud set, and using the straight-through filter algorithm of the PCL point cloud library to retain the point cloud within a preset area.

5. The method according to any one of claims 1 to 3, wherein: The semantic segmentation network model includes four SA modules for extracting the point cloud features, four upsampling layers for restoring the number of the point cloud features to the size of the point cloud set, and a classification fully connected layer for classifying the point cloud labels; The SA module includes a sampling layer, a grouping layer and a feature layer for randomly sampling point clouds; the grouping layer presets multiple different grouping radii to group the point cloud set; the feature layer is composed of 3 groups of dynamic edge convolutions to extract the point cloud features.

6. The method according to claim 5, wherein The semantic segmentation network model extracts features from the point cloud sets corresponding to different grouping radii through the SA module to obtain grouping features; and splices multiple grouping features to obtain local features; Extracting edge features of adjacent point cloud sets by applying the dynamic edge convolution to the local features, and fusing the local features with the edge features to obtain point cloud features; The point cloud features are upsampled to restore the number of point clouds in the point cloud set to the original number; and the point cloud set is classified through a fully connected layer.

7. The method according to claim 6, wherein The extracting edge features of adjacent point cloud sets by applying the local features to the dynamic edge convolution comprises: Deleting the feature corresponding to the center point from the grouping feature to obtain a new grouping feature; splicing multiple new grouping features to obtain the local feature; A weight score matrix is ​​generated according to the coordinates of the point cloud set; the local features and the preset fully connected layer are multiplied to obtain a weight parameter matrix, and the local features and the weight parameter matrix are multiplied to obtain a feature matrix; the score matrix and the feature matrix are multiplied to associate the coordinates of the point cloud set with the weight parameter matrix.

8. The method according to claim 5, wherein Restoring the number of the point cloud features to the size of the point cloud set includes: Obtaining the point cloud features of the previous upsampling layer as the points to be interpolated, and obtaining the point cloud features of the next upsampling layer as the reference points; Traversing the point to be interpolated, selecting multiple reference points with the smallest distance to the point to be interpolated, and calculating the sum of the reciprocals of the multiple distances; using the ratio of the reciprocal of the distance to the sum of the reciprocals as the weight of the reference point; calculating the product of the reference point and the weight, and using the average of the products corresponding to the multiple reference points as the new point cloud feature of the point to be interpolated; splicing the features before and after interpolation to obtain the point cloud feature with increased dimension.

9. The method according to claim 1, wherein The semantic segmentation network model adopts the transfer learning method, firstly undergoes multiple rounds of training with a public dataset, and then verifies it with a public dataset, calculates the average intersection-over-union ratio of multiple categories of the public dataset, and selects the semantic segmentation network model corresponding to the largest average intersection-over-union ratio; Establish a first data set of mulberry garden scenes, use the training set to perform multiple rounds of training on the semantic segmentation network model, calculate the average intersection-over-union ratio of multiple categories of the first data set, and screen out the semantic segmentation network model corresponding to the largest average intersection-over-union ratio as the final semantic segmentation network model.

10. A mulberry orchard autonomous navigation system based on laser radar point cloud, characterized in that: include: a semantic segmentation module for collecting a point cloud set and performing semantic segmentation on the preprocessed point cloud set using a semantic segmentation network model to obtain point cloud features corresponding to the point cloud set, including point cloud labels; the point cloud labels include a first label associated with the path area and a second label associated with the crop area; an area recognition module, configured to obtain the crop area adjacent to the path area, and determine, based on the second label, whether the corresponding crop area is a continuous space or a discrete space; a path generation module, for generating a candidate box for the point cloud set of the crop area according to the point cloud features, and selecting a point of interest from the candidate box, if the crop area is a discrete space; If the crop area is a continuous space, obtaining a boundary point cloud of the crop area as the point of interest; A first threshold is preset, and a plurality of the points of interest are offset toward the path area by the first threshold to obtain a plurality of navigation points, so that the mobile carrier moves along the plurality of navigation points.

Images