A navigation path generation method, system, electronic device and readable storage medium

Abstract

The application discloses a navigation path generation method, system, electronic equipment and readable storage medium, and relates to the navigation field, and solves the problems of reduced positioning accuracy and significant influence of light conditions in the prior art navigation path generation technology in a crop canopy shielding scene. The application discloses a navigation path generation method, which comprises the following steps: in a data acquisition stage, three-dimensional point cloud data is obtained by using a laser radar sensor; in a plane projection stage, vertical projection is performed on the three-dimensional point cloud data to obtain two-dimensional projection point cloud; in an inner crop row extraction stage, the two-dimensional projection point cloud is segmented to obtain left crop fitting parameters and right crop fitting parameters; and in a navigation path generation stage, three-level smoothing processing is performed based on the left crop fitting parameters and the right crop fitting parameters, smooth navigation path is obtained, and the generation of the navigation path is completed. The application is suitable for the fields of agricultural robots, autonomous navigation, laser radars, point cloud processing and precision agriculture.

Description

Technical Field

[0001] This invention relates to the field of navigation, and also to the field of agricultural technology, specifically to the field of navigation path generation. Background Technology

[0002] With the development of agricultural modernization, agricultural robots are being used more and more widely in farmland operations. During crop planting, agricultural robots need to navigate precisely along the rows of crops to complete tasks such as sowing, fertilizing, and spraying pesticides. Generating navigation paths under the crop canopy is a key technology for ensuring operational accuracy and efficiency. Currently, existing navigation methods are mainly divided into three categories, as follows: GNSS-based positioning and navigation methods rely on the Global Positioning System for path planning to enable agricultural robots to navigate. However, in scenarios where crop canopies obstruct the view, GPS signal quality deteriorates significantly, leading to reduced positioning accuracy and failing to meet the requirements for precise navigation. Furthermore, this method cannot perceive the actual position of crop rows in real time, has poor adaptability to curved or irregular crop planting conditions, and lacks obstacle avoidance capabilities, posing safety hazards during operation.

[0003] Traditional LiDAR-based navigation methods collect 3D point cloud data of crops using LiDAR, and then process the point cloud to extract crop row information to generate a navigation path. However, this method has several technical shortcomings: First, directly processing 3D point cloud data results in a massive amount of data and high computational complexity, leading to poor real-time navigation performance. Second, the point cloud distribution along the crop plant height is uneven and contains many noise points, severely affecting the accuracy of crop row fitting. Third, in multi-row crop scenarios, it is difficult to accurately identify the innermost crop row, easily leading to false detections of outer rows. Fourth, the generated navigation path exhibits severe jitter and poor stability, hindering stable tracking by agricultural robots. Fifth, the lack of an effective spatiotemporal smoothing mechanism results in insufficient continuity of the navigation path, further affecting operational accuracy.

[0004] The core of machine vision-based navigation methods is to acquire crop images through cameras and then use image processing technology to identify the center lines of crop rows, thereby achieving navigation. However, this method is significantly affected by lighting conditions. In complex lighting environments such as strong light, shadow, backlight, and nighttime, its navigation performance will drop significantly. At the same time, crop canopy occlusion will limit the camera's field of view, leading to a decrease in the accuracy of crop row recognition. Furthermore, the image processing process is computationally intensive and has poor real-time performance, making it unsuitable for the high-speed operation requirements of agricultural robots.

[0005] In the prior art, Chinese patent document CN115014358A discloses a "method, device, system, control equipment, and storage medium for crop row navigation." This technical solution is applicable to large agricultural machinery using image processing technology to observe from top to bottom. When the current driving path of the agricultural machinery interferes with the crop planting area based on the crop contour information, a corrected path for the interference area is obtained through the crop contour information, allowing the agricultural machinery to follow the corrected path for row navigation in the interference area. However, this method does not consider the scenario where the "row navigation reference" (usually the ground or the line connecting the bottom of the stem) required for crop row navigation is completely obscured by the canopy. For example, in a mature cornfield, detection is performed in the middle under the corn row canopy, resulting in an uncertain and dynamically changing offset between the "canopy surface" visible to the sensor and the true reference, making accurate navigation impossible.

[0006] In summary, existing navigation path generation technologies suffer from several drawbacks: reduced positioning accuracy in scenarios with crop canopy occlusion; significant impact from lighting conditions; poor adaptability to curved or irregular crop planting patterns and lack of obstacle avoidance capabilities; massive data volume and high computational complexity, resulting in poor real-time navigation performance; uneven point cloud distribution along crop height with numerous noise points; difficulty in accurately identifying the innermost crop row in multi-row crop scenarios; and severe jitter and poor stability in the generated navigation path. Summary of the Invention

[0007] This invention alleviates the problems of existing navigation path generation technologies, such as reduced positioning accuracy in scenarios with crop canopy occlusion; significant influence from lighting conditions; poor adaptability to curved or irregular crop planting conditions and lack of obstacle avoidance capabilities; large data volume, high computational complexity, and poor real-time navigation performance; uneven distribution of point clouds along the crop height direction with many noise points; difficulty in accurately identifying the innermost crop row in multi-row crop scenarios; and severe jitter and poor stability of the generated navigation path.

[0008] This invention provides the following solution: Option 1: A navigation path generation method, comprising the following stages: Data acquisition stage: The raw 3D point cloud data of the current area is collected using a lidar sensor; the raw 3D point cloud data is preprocessed to obtain 3D point cloud data. Planar projection stage: The three-dimensional point cloud data is vertically projected to obtain a two-dimensional projected point cloud; Inner crop row extraction stage: The two-dimensional projection point cloud is segmented to obtain the left crop projection point cloud and the right crop projection point cloud; the fitting parameters are extracted from the left crop projection point cloud and the right crop projection point cloud respectively to obtain the corresponding left crop fitting parameters and right crop fitting parameters; the fitting parameter extraction includes inner crop row extraction and line fitting. Navigation path generation stage: Based on the crop fitting parameters on the left and right sides, several path points are obtained; the path points are then subjected to three-level smoothing to obtain a smoothed navigation path, thus completing the generation of the navigation path; The three-level smoothing process includes outlier removal, spatial smoothing, and temporal smoothing, respectively.

[0009] Furthermore, in one embodiment of the present invention, the preprocessing method in the data acquisition stage is as follows: The original 3D point cloud data is filtered along the X-axis, Y-axis and Z-axis directions using a PassThrough filter to obtain the corresponding point cloud data along the X-axis, Y-axis and Z-axis directions. VoxelGrid filters are used to voxelize and downsample the unfiltered raw 3D point cloud data to obtain point cloud data in other directions; The point cloud data in the X-axis direction, Y-axis direction, Z-axis direction, and other directions are combined into three-dimensional point cloud data to complete the preprocessing.

[0010] Furthermore, in one embodiment of the present invention, the method for extracting the inner crop rows in the inner crop row extraction stage is as follows: Taking the left-side crop projection point cloud as an example, one-dimensional clustering is performed on the point cloud in the left-side crop projection point cloud according to the clustering threshold to obtain several point clusters; The point clusters are judged, and the point clusters that meet the inner crop conditions are taken as the inner crop rows, thus completing the extraction of the inner crop rows.

[0011] Furthermore, in one embodiment of the present invention, the clustering threshold is obtained by the following method: If adaptive mode is enabled, the clustering threshold is determined by...

[0012] Obtain, among which, The median of the interval. It is the gap multiplier. The minimum threshold; Otherwise, the minimum threshold will be... As a clustering threshold.

[0013] Furthermore, in one embodiment of the present invention, the method for outlier removal during the navigation path generation stage is as follows: Based on the aforementioned path points, the Euclidean distance between each path point and its adjacent path points is obtained and combined into a set of Euclidean distances. Based on the aforementioned set of Euclidean distances, the mean of the Euclidean distances is obtained. and standard deviation ; For each Euclidean distance in the set of Euclidean distances Make a judgment; if it satisfies...

[0014] Then retain the Euclidean distance. The corresponding path point and its adjacent path points; otherwise, the Euclidean distance. The adjacent path points of the corresponding path point are considered outliers and are removed. This is the outlier threshold coefficient.

[0015] Furthermore, in one embodiment of the present invention, the spatial smoothing method in the navigation path generation stage is as follows: A spatially weighted average is applied to several path points after outlier removal to obtain spatially smoothed path points. The spatial weighted average method is as follows:

[0016] in, These are the path points after spatial smoothing. For spatial smoothing weights, For the current path point, The previous neighboring path point of the current path point. It is the next adjacent path point of the current path point.

[0017] Furthermore, in one embodiment of the present invention, the time smoothing method in the navigation path generation stage is as follows: A time-weighted average is performed on several path points after spatial smoothing and the corresponding path points in the historical path queue to obtain several path points after time smoothing. These path points are then combined to form a navigation path.

[0018] Option 2: A navigation path generation system, comprising the following modules: Module 1 is used to collect raw 3D point cloud data of the current area using a lidar sensor; and to preprocess the raw 3D point cloud data to obtain 3D point cloud data. Module 2 is used to perform vertical projection on the three-dimensional point cloud data to obtain a two-dimensional projected point cloud; Module 3 is used to segment the two-dimensional projection point cloud to obtain the left crop projection point cloud and the right crop projection point cloud; and to extract fitting parameters from the left crop projection point cloud and the right crop projection point cloud respectively to obtain the corresponding left crop fitting parameters and right crop fitting parameters; the fitting parameter extraction includes inner crop row extraction and line fitting. Module 4 is used to obtain several path points based on the crop fitting parameters on the left and right sides; and to perform three-level smoothing on the several path points to obtain a smoothed navigation path, thus completing the generation of the navigation path. The three-level smoothing process includes outlier removal, spatial smoothing, and temporal smoothing, respectively.

[0019] Option 3: An electronic device includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When a processor executes a program stored in memory, it implements the method described in Scheme 1.

[0020] Option 4: A computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the method described in Option 1.

[0021] The navigation path generation method described in this invention is based on 3D LiDAR, effectively alleviating the problems of existing navigation path generation technologies, such as large data volume, high computational complexity, poor real-time navigation; reduced positioning accuracy in crop canopy occlusion scenarios; significant influence from lighting conditions; uneven point cloud distribution along crop height with numerous noise points; difficulty in accurately identifying the innermost crop row in multi-row crop scenarios; poor adaptability to curved or irregular crop planting conditions and lack of obstacle avoidance capabilities; and severe jitter and poor stability of the generated navigation path. Specific beneficial effects include: 1. The navigation path generation method of the present invention combines data acquisition, planar projection, inner crop row extraction and navigation path generation, and simultaneously meets the practical operational requirements of high computational efficiency, high detection accuracy and good path quality.

[0022] 2. The navigation path generation method described in this invention, through data acquisition and planar projection, uses the concept of dimensionality reduction to transform the complex three-dimensional point cloud processing problem into a simple two-dimensional plane problem, and significantly reduces the computational complexity through XY plane projection.

[0023] 3. The navigation path generation method described in this invention improves detection accuracy by extracting inner crop rows and adopting an adaptive clustering strategy, thus solving problems such as reduced positioning accuracy in crop canopy occlusion scenarios; significant influence from lighting conditions; uneven point cloud distribution along crop height with many noise points; and difficulty in accurately identifying the innermost crop row in multi-row crop scenarios.

[0024] 4. The navigation path generation method of the present invention improves path quality from different dimensions by using a three-level smoothing mechanism through navigation path generation.

[0025] This invention is applicable to fields such as agricultural robots, autonomous navigation, lidar, point cloud processing, and precision agriculture. Attached Figure Description

[0026] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of the navigation path generation method described in Implementation Method 1. Detailed Implementation

[0027] Various embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings. The embodiments described with reference to the drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0028] Implementation Method 1: A navigation path generation method described in this implementation method, such as... Figure 1 As shown, the navigation path generation method includes the following stages: Data acquisition stage: The raw 3D point cloud data of the current area is collected using a lidar sensor; the raw 3D point cloud data is preprocessed to obtain 3D point cloud data. Planar projection stage: The three-dimensional point cloud data is vertically projected to obtain a two-dimensional projected point cloud; Inner crop row extraction stage: The two-dimensional projection point cloud is segmented to obtain the left crop projection point cloud and the right crop projection point cloud; the fitting parameters are extracted from the left crop projection point cloud and the right crop projection point cloud respectively to obtain the corresponding left crop fitting parameters and right crop fitting parameters; the fitting parameter extraction includes inner crop row extraction and line fitting. Navigation path generation stage: Based on the crop fitting parameters on the left and right sides, several path points are obtained; the path points are then subjected to three-level smoothing to obtain a smoothed navigation path, thus completing the generation of the navigation path; The three-level smoothing process includes outlier removal, spatial smoothing, and temporal smoothing, respectively.

[0029] In this embodiment, during the data acquisition stage, a lidar sensor installed on an agricultural robot performs three-dimensional data acquisition. The raw three-dimensional point cloud data includes the three-dimensional coordinate information and reflection intensity information of the crop plants.

[0030] In this embodiment, during the planar projection stage, the vertical projection involves projecting the preprocessed 3D point cloud data vertically onto the XY horizontal plane, thereby reducing the data dimensionality. The method for vertical projection is as follows: Traverse each 3D point Extract its and coordinate; Generate the corresponding two-dimensional projection points ; To maintain the visualization effect, the z-coordinates of all projection points are set to fixed small values; This creates a two-dimensional projection point cloud, simplifying subsequent processing complexity.

[0031] In this embodiment, during the inner crop row extraction stage, the segmentation involves dividing the projected point cloud into a left crop row and a right crop row based on the Y-coordinate, which are then used as the left crop projected point cloud and the right crop projected point cloud, respectively. The segmentation method is as follows: Left crop row: All projection points; Right-hand crop row: All projection points.

[0032] In this embodiment, the method for straight line fitting during the inner crop row extraction stage is as follows: Fit the linear equation using the least squares method. ; Calculate the slope and intercept ; Abnormal slopes are subjected to amplitude limiting to ensure the reasonableness of the fitting results.

[0033] In this embodiment, the method for obtaining several path points in the navigation path generation stage is as follows: pass To obtain the average slope; pass To obtain the average intercept; In the odom global coordinate system, path points are generated with a fixed step size from the robot's current position x_start to the front x_end. The coordinates of each path point are: ; Calculate the path orientation angle based on the slope. .

[0034] This implementation also includes a coordinate system transformation and publishing stage: The navigation path in the odom coordinate system is transformed to the base_link robot body coordinate system using TF transformation for visualization. Maintain the path in the odom coordinate system for robot control; The processed point cloud data and navigation path are then published to the downstream control system.

[0035] This implementation also includes a Fallback phase, which promptly reverts to a conservative strategy or publishes an empty path when an anomaly is detected, ensuring system security.

[0036] The navigation path generation method described in this embodiment uses data acquisition and planar projection to transform the complex three-dimensional point cloud processing problem into a simple two-dimensional planar problem by applying the concept of dimensionality reduction. This method is not a conventional dimensionality reduction that simply discards feature information, but rather a comprehensive technical solution that combines precise purification of the preceding point cloud, preservation of core features, precise fitting of the subsequent point cloud, and multi-level smooth correction. This fundamentally overcomes the technical defect that conventional dimensionality reduction can easily lead to a decrease in detection accuracy.

[0037] This implementation uses a three-dimensional LiDAR to collect data during the data acquisition stage, completely replacing the visual camera. It does not rely on image brightness and color, and is not affected by strong light, shadows, backlight, or nighttime. It only collects three-dimensional coordinates and reflection intensity, maintaining stable perception under full illumination conditions and eliminating light interference from the hardware source.

[0038] In this implementation, during the data acquisition and planar projection stages, the lidar can penetrate the gaps in the crop canopy to obtain the effective point cloud of the lower plants; the effective height range of the crop canopy is locked by Z-axis height filtering, eliminating interference from the ground and invalid shading; and then the XY projection retains the positional features between rows, unaffected by the field of view cutoff caused by canopy shading, ensuring the complete and accurate extraction of crop rows.

[0039] This implementation method uses PassThrough filtering to remove outliers in the height direction during the planar projection stage and the navigation path generation stage. Then, it uses VoxelGrid voxel downsampling to equalize the point cloud density and remove noise. The 3D point cloud is vertically projected onto the XY plane to eliminate the fitting bias caused by uneven distribution in the height direction. Finally, through three-level processing of outlier removal, spatial weighted smoothing, and temporal history frame smoothing, the fitting stability and accuracy are significantly improved.

[0040] The three-level smoothing process in the navigation path generation stage of this embodiment is an improvement based on existing technology. However, directly applying the existing three-level smoothing process to navigation scenarios under crop canopies would present significant technical challenges and fail to solve the practical problems of agricultural robot navigation, as detailed below: Farmland point clouds are noisy and highly random. Ordinary smoothing will filter out "actual crop row deviations" as noise, causing the path to deviate from the center of the rows. Agricultural robots require high real-time performance. General time smoothing will introduce lag, and the path will not keep up when turning or changing rows. In the case of multiple rows of crops, missing seedlings, or broken rows, single-point anomalies will contaminate the entire path. Ordinary smoothing cannot accurately remove anomalies. The navigation path must be continuous and jitter-free. Simple spatial smoothing cannot eliminate inter-frame jumps, and the robot is prone to oscillations when tracking.

[0041] To address the aforementioned issues, this implementation method first removes anomalies and then smooths the surface. It first calculates the Euclidean distance between adjacent path points, then dynamically identifies and forcibly removes outliers using the mean and standard deviation, preventing noise points from being introduced into the smoothing process. This method is specifically designed to handle sudden changes in farmland noise and seedling loss.

[0042] Spatial smoothing employs a weighted average of the current point and a dedicated formula for constraints on preceding and following points. A fixed-weighted average is used for intermediate path points, smoothing only local fluctuations without altering the overall path direction, ensuring the path always adheres to the object's centerline without deviation.

[0043] Temporal smoothing employs a historical path queue and point-by-point weighted inter-frame fusion. It maintains the path of the most recent N frames and only weights points at the same location, which suppresses inter-frame jitter without introducing control lag, thus adapting to the robot's high-speed real-time navigation.

[0044] The third-level smoothing is strongly coupled with the preceding processes. The smoothed object is a high-quality path that has been purified, dimensionality reduced, and adaptively clustered to extract the inner rows. The input is clean, and the smoothing effect is stable and reliable. All smoothing aims to maintain the centering between rows, stable orientation, and no jitter, rather than simply removing noise. The final path can be directly fed into the chassis control.

[0045] In this implementation, through the inner crop row extraction stage and the navigation path generation stage, it does not rely on fixed row lines and GNSS preset trajectories, but extracts the current innermost crop row in real time frame by frame and fits the actual direction using the least squares method; the adaptive threshold can adapt to field morphology with missing seedlings, bends, and uneven density, and always follows the actual distribution of crops to calculate the center path, and has strong adaptability to bends and irregular crop rows.

[0046] During the fitting stage, abnormal slopes are forcibly limited to avoid sudden jitter. During the smoothing stage, outliers are first removed, then spatial weighted smoothing is used to suppress local jitter, and finally historical frame time smoothing is used to eliminate inter-frame jumps, so that the path is continuous and smooth, meeting the requirements for stable robot tracking.

[0047] In this implementation method, path points are generated uniformly at fixed step lengths during the navigation path generation stage to ensure continuous point positions; three-level smoothing ensures the trajectory is continuous and uninterrupted from three dimensions: single point, local, and temporal; and stable release combined with coordinate system transformation ensures that the path moves without jumps, interruptions, or deviations, significantly improving operational accuracy.

[0048] Implementation Method Two: This implementation method further defines the navigation path generation method described in Implementation Method One. In this implementation method, the preprocessing method in the data acquisition stage is as follows: The original 3D point cloud data is filtered along the X-axis, Y-axis and Z-axis directions using a PassThrough filter to obtain the corresponding point cloud data along the X-axis, Y-axis and Z-axis directions. VoxelGrid filters are used to voxelize and downsample the unfiltered raw 3D point cloud data to obtain point cloud data in other directions; The point cloud data in the X-axis direction, Y-axis direction, Z-axis direction, and other directions are combined into three-dimensional point cloud data to complete the preprocessing.

[0049] In this embodiment, a PassThrough filter is used to filter the original 3D point cloud data along the X-axis, retaining the forward distance range. The point cloud within the area is limited in length to the processing region.

[0050] In this embodiment, a PassThrough filter is used to filter the original 3D point cloud data along the Y-axis, preserving the lateral range. The point cloud within the area limits the width of the processing region.

[0051] In this embodiment, a PassThrough filter is used to filter the original 3D point cloud data along the Z-axis, preserving the height range. Point cloud within, of which and Based on the crop canopy height setting, interference from ground weeds and excessively tall vegetation is removed.

[0052] In this embodiment, a VoxelGrid filter is used to voxelize and downsample the unfiltered raw 3D point cloud data to obtain point cloud data in other directions, thereby reducing the amount of data while maintaining the geometric features of the point cloud.

[0053] This implementation further defines the data acquisition stage and describes the preprocessing. The method filters and downsamples the original 3D point cloud data to remove noise points and redundant data. Triaxial PassThrough filtering strictly limits the range of crop canopy height, lateral width, and forward distance, eliminating interference points such as ground weeds, excessively tall vegetation, and redundant noise. VoxelGrid voxel downsampling compresses the data volume while preserving the geometric features of the point cloud, ensuring that all point clouds entering the projection stage are valid feature points of the crop rows, eliminating the impact of invalid points on subsequent detection accuracy from the source. Furthermore, by discarding only redundant information in the height direction, the core features of the crop rows in the XY plane, such as lateral position, longitudinal orientation, and row spacing, are fully preserved. After projection, the actual distribution information of the crop rows is still accurately reflected, without the loss of key positioning information, thus avoiding any reduction in detection accuracy.

[0054] Implementation Method 3: This implementation method further defines the navigation path generation method described in Implementation Method 1. In this implementation method, the method for extracting the inner crop rows in the inner crop row extraction stage is as follows: Taking the left-side crop projection point cloud as an example, one-dimensional clustering is performed on the point cloud in the left-side crop projection point cloud according to the clustering threshold to obtain several point clusters; The point clusters are judged, and the point clusters that meet the inner crop conditions are taken as the inner crop rows, thus completing the extraction of the inner crop rows.

[0055] In this embodiment, the inner crop condition is that the number of point clouds in the point cluster is at least 10, and the average y-coordinate of the point clouds in the point cluster is the smallest.

[0056] With a minimum of 10 points in both the left and right crop projection point clouds, this approach ensures sufficient points for reliable straight-line fitting while avoiding path jitter caused by noise or a few outliers, thus balancing accuracy with real-time performance requirements.

[0057] The point cloud with the smallest average y-coordinate in the point cluster is used to extract the inner crop row that is closest to the robot's longitudinal centerline.

[0058] This implementation further defines the inner crop row extraction stage, focusing on the extraction of inner crop rows. The method is described as follows: the point cloud is divided into left and right crop row regions according to the Y coordinate; the point cloud on each side is sorted according to the Y direction and the adjacent spacing is calculated. One-dimensional clustering with adaptive threshold is adopted, and the cluster boundary is automatically determined based on the median of the spacing, the gap multiplier, and the minimum threshold; the point cluster closest to the center of the row and satisfying the minimum number of points is selected as the innermost crop row, so as to avoid false detection of the outer row from the algorithm mechanism.

[0059] Implementation Method Four: This implementation method further defines the navigation path generation method described in Implementation Method Three. In this implementation method, the clustering threshold is obtained through the following method: If adaptive mode is enabled, the clustering threshold is determined by...

[0060] Obtain, among which, The median of the interval. It is the gap multiplier. The minimum threshold; Otherwise, the minimum threshold will be... As a clustering threshold.

[0061] In this embodiment, the lower limit of the minimum threshold range is greater than or equal to 3 to 5 times the sensor noise, and the upper limit of the minimum threshold range is less than half of the minimum expected line spacing.

[0062] The lower limit of the minimum threshold range is used to avoid the clustering threshold from being too sensitive; the upper limit of the minimum threshold range is used to distinguish different crop rows.

[0063] In this embodiment, the minimum threshold is preferably in the range of 0.05m to 0.3m; when the navigation path generation method is applied to cornfields, the minimum threshold is in the range of 0.1m to 0.2m; when the navigation path generation method is applied to fruit tree environments, the minimum threshold is in the range of 0.15m to 0.3m.

[0064] In this implementation, whether to enable the adaptive mode depends on factors such as the magnitude of changes in the farmland environment and the flexibility of the operational tasks. When the farmland environment changes significantly, such as in the same field where crops are lush in some areas and sparse in others, or when different varieties of crops are mixed, enabling the adaptive mode automatically adjusts the threshold based on the local point cloud distribution to avoid over-segmentation in dense areas and under-segmentation in sparse areas.

[0065] In farmland environments with minimal changes, within stable and known standardized farmland, row and plant spacing are strictly standardized, and the adaptive mode is not enabled. The minimum threshold is used as a fixed threshold; after one parameter tuning optimization, this fixed threshold can be used long-term, offering greater stability and reliability. When high flexibility is required for operational tasks, a single robot needs to navigate in various farmland conditions. Using adaptive mode reduces the frequency of manual parameter tuning and improves system adaptability. In single-field, single-season, single-crop scenarios, using a fixed threshold avoids unnecessary computational overhead and simplifies system behavior.

[0066] In this embodiment, the median of the spacing Obtained through the following methods: Taking the left crop projection point cloud as an example, the point clouds in the left crop projection point cloud are sorted from smallest to largest according to the Y-axis coordinate to obtain the sorted point cloud sequence; Based on the point cloud sequence, the Y-axis spacing of each group of adjacent point clouds is obtained and combined into a Y-axis spacing sequence; the median of the Y-axis spacing sequence is used as the median spacing. .

[0067] This embodiment further defines the extraction of inner crop rows and explains the clustering threshold. This method automatically adjusts the clustering parameters according to the actual distribution of the point cloud, rather than using a fixed threshold, thereby improving adaptability to different planting densities and seedling shortages.

[0068] While existing adaptive clustering is a conventional technical concept, it cannot be directly applied to navigation scenarios under crop canopies. Direct application would lead to technical problems such as false detection of crop rows, poor adaptability to missing seedlings and uneven density, and failure of inner row positioning.

[0069] This method adapts the model to the environment under the canopy. Based on XY plane projection, the adaptive clustering is limited to one-dimensional spacing in the Y direction, and only the horizontal distribution between rows is calculated. This fits the arrangement pattern of the plant rows and avoids multidimensional interference.

[0070] Based on the triple constraints of median spacing, gap multiplier, and minimum threshold, it automatically adapts to different crop densities, missing seedlings, and gap mutations, rather than a universal threshold.

[0071] After clustering, the cluster of points closest to the center of the row is selected as the inner crop row. At the same time, it is coupled with the pre-processing point cloud filtering and post-processing fitting smoothing depth, which significantly improves the accuracy and robustness of crop row extraction.

[0072] Implementation Method 5: This implementation method further defines the navigation path generation method described in Implementation Method 1. In this implementation method, the outlier removal method in the navigation path generation stage is as follows: Based on the aforementioned path points, the Euclidean distance between each path point and its adjacent path points is obtained and combined into a set of Euclidean distances. Based on the aforementioned set of Euclidean distances, the mean of the Euclidean distances is obtained. and standard deviation ; For each Euclidean distance in the set of Euclidean distances Make a judgment; if it satisfies...

[0073] Then retain the Euclidean distance. The corresponding path point and its adjacent path points; otherwise, the Euclidean distance. The adjacent path points of the corresponding path point are considered outliers and are removed. This is the outlier threshold coefficient.

[0074] In this embodiment, the outlier threshold coefficient is shown in Table 1.

[0075] Table 1

[0076] In this embodiment, the outlier removal can employ the Z-Score method, with the threshold coefficient selected according to the 3σ principle. If the threshold coefficient is 2.0 and the confidence level is 95%, then 95% ± 5% of the data is retained. This method effectively removes obvious outliers without excessively removing normal data, making it suitable for sensor noise suppression in agricultural environments.

[0077] Implementation Method Six: This implementation method further defines the navigation path generation method described in Implementation Method One. In this implementation method, the spatial smoothing method in the navigation path generation stage is as follows: A spatially weighted average is applied to several path points after outlier removal to obtain spatially smoothed path points. The spatial weighted average method is as follows:

[0078] in, These are the path points after spatial smoothing. For spatial smoothing weights, For the current path point, The previous neighboring path point of the current path point. It is the next adjacent path point of the current path point.

[0079] In this embodiment, the previous adjacent path point and the next adjacent path point are adjacent path points in index order. Taking the direction of the navigation path as the positive x-axis as an example, if the x value of the current path point i is 5, then the x value of the previous adjacent path point i-1 is less than 5, and the x value of the next adjacent path point i+1 is greater than 5.

[0080] Implementation Method Seven: This implementation method further defines the navigation path generation method described in Implementation Method One. In this implementation method, the time smoothing method in the navigation path generation stage is as follows: A time-weighted average is performed on several path points after spatial smoothing and several path points within the corresponding window in the historical path queue to obtain several path points after time smoothing. These path points are then combined to form a navigation path.

[0081] In this embodiment, the time-weighted average method is as follows:

[0082] in, These are the path points after time smoothing. For the first in the window Path points, This represents the total number of path points within the window.

Claims

1. A navigation path generation method, characterized in that, The navigation path generation method includes the following stages: Data acquisition stage: The raw 3D point cloud data of the current area is collected using a lidar sensor; the raw 3D point cloud data is preprocessed to obtain 3D point cloud data. Planar projection stage: The three-dimensional point cloud data is vertically projected to obtain a two-dimensional projected point cloud; Inner crop row extraction stage: The two-dimensional projection point cloud is segmented to obtain the left crop projection point cloud and the right crop projection point cloud; Fitting parameters are extracted from the left crop projection point cloud and the right crop projection point cloud respectively to obtain the corresponding left crop fitting parameters and right crop fitting parameters; the fitting parameter extraction includes inner crop row extraction and straight line fitting. Navigation path generation stage: Based on the crop fitting parameters on the left and right sides, several path points are obtained; the path points are then subjected to three-level smoothing to obtain a smoothed navigation path, thus completing the generation of the navigation path; The three-level smoothing process includes outlier removal, spatial smoothing, and temporal smoothing, respectively. In the inner crop row extraction stage, the method for extracting the inner crop rows is as follows: Taking the left-side crop projection point cloud as an example, one-dimensional clustering is performed on the point cloud in the left-side crop projection point cloud according to the clustering threshold to obtain several point clusters; The point clusters are judged, and the point clusters that meet the inner crop conditions are taken as the inner crop rows, thus completing the extraction of the inner crop rows; The inner crop condition is that the number of point clouds in the point cluster is at least 10, and the average y-coordinate of the point clouds in the point cluster is the smallest; The clustering threshold is obtained by the following method: If adaptive mode is enabled, the clustering threshold is determined by... Obtain, among which, The median of the interval. It is the gap multiplier. The minimum threshold; Otherwise, the minimum threshold will be... As a clustering threshold; The lower limit of the minimum threshold range is greater than or equal to 3 to 5 times the sensor noise, and the upper limit of the minimum threshold range is less than half of the minimum expected line spacing; The spatial smoothing method mentioned in the navigation path generation stage is as follows: A spatially weighted average is applied to several path points after outlier removal to obtain spatially smoothed path points. The spatial weighted average method is as follows: in, These are the path points after spatial smoothing. For spatial smoothing weights, For the current path point, The previous neighboring path point of the current path point. It is the next adjacent path point of the current path point.

2. The navigation path generation method according to claim 1, characterized in that, The preprocessing method described in the data acquisition stage is as follows: The original 3D point cloud data is filtered along the X-axis, Y-axis and Z-axis directions using a PassThrough filter to obtain the corresponding point cloud data along the X-axis, Y-axis and Z-axis directions. VoxelGrid filters are used to voxelize and downsample the unfiltered raw 3D point cloud data to obtain point cloud data in other directions; The point cloud data in the X-axis direction, Y-axis direction, Z-axis direction, and other directions are combined into three-dimensional point cloud data to complete the preprocessing.

3. The navigation path generation method according to claim 1, characterized in that, The outlier removal method during the navigation path generation stage is as follows: Based on the aforementioned path points, the Euclidean distance between each path point and its adjacent path points is obtained and combined into a set of Euclidean distances. Based on the aforementioned set of Euclidean distances, the mean of the Euclidean distances is obtained. and standard deviation ; For each Euclidean distance in the set of Euclidean distances Make a judgment if the condition is met. Then retain the Euclidean distance. The corresponding path point and its adjacent path points; otherwise, the Euclidean distance. The adjacent path points of the corresponding path point are considered outliers and are removed. This is the outlier threshold coefficient.

4. The navigation path generation method according to claim 1, characterized in that, The time smoothing method mentioned in the navigation path generation stage is as follows: A time-weighted average is performed on several path points after spatial smoothing and the corresponding path points in the historical path queue to obtain several path points after time smoothing. These path points are then combined to form a navigation path.

5. A navigation path generation system, characterized in that, Includes the following modules: Module 1 is used to collect raw 3D point cloud data of the current area using a lidar sensor; and to preprocess the raw 3D point cloud data to obtain 3D point cloud data. Module 2 is used to perform vertical projection on the three-dimensional point cloud data to obtain a two-dimensional projected point cloud; Module 3 is used to segment the two-dimensional projection point cloud to obtain the left crop projection point cloud and the right crop projection point cloud; Fitting parameters are extracted from the left crop projection point cloud and the right crop projection point cloud respectively to obtain the corresponding left crop fitting parameters and right crop fitting parameters; the fitting parameter extraction includes inner crop row extraction and straight line fitting. Module 4 is used to obtain several path points based on the crop fitting parameters on the left and right sides; and to perform three-level smoothing on the several path points to obtain a smoothed navigation path, thus completing the generation of the navigation path. The three-level smoothing process includes outlier removal, spatial smoothing, and temporal smoothing, respectively. In Module 3, the method for extracting the inner crop rows is as follows: Taking the left-side crop projection point cloud as an example, one-dimensional clustering is performed on the point cloud in the left-side crop projection point cloud according to the clustering threshold to obtain several point clusters; The point clusters are judged, and the point clusters that meet the inner crop conditions are taken as the inner crop rows, thus completing the extraction of the inner crop rows; The clustering threshold is obtained by the following method: If adaptive mode is enabled, the clustering threshold is determined by... Obtain, among which, The median of the interval. It is the gap multiplier. The minimum threshold; Otherwise, the minimum threshold will be... As a clustering threshold; The lower limit of the minimum threshold range is greater than or equal to 3 to 5 times the sensor noise, and the upper limit of the minimum threshold range is less than half of the minimum expected line spacing; The spatial smoothing method described in Module 4 is as follows: A spatially weighted average is applied to several path points after outlier removal to obtain spatially smoothed path points. The spatial weighted average method is as follows: in, These are the path points after spatial smoothing. For spatial smoothing weights, For the current path point, The previous neighboring path point of the current path point. It is the next adjacent path point of the current path point.

6. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the method of any one of claims 1-4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1-4.

Images