Lidar-based farmland furrow three-dimensional point cloud data correction method and system
By processing point cloud data under dynamic agricultural machinery operations using posture compensation and adaptive region growing algorithms, the distortion problem was solved, and accurate reconstruction and real-time detection of the three-dimensional morphology of furrows were achieved, thus improving the accuracy and reliability of plowing operations.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- JILIN AGRICULTURAL UNIV
- Filing Date
- 2026-04-02
- Publication Date
- 2026-06-12
Smart Images
Figure CN121962564B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of three-dimensional image processing, and more specifically to three-dimensional image geometric correction technology. Background Technology
[0002] Plowing is a crucial step in agricultural production, and its quality directly determines the effectiveness of subsequent agronomic measures such as sowing and irrigation, ultimately affecting crop yield. In precision agriculture systems, sowing, fertilization, and plant protection can achieve precise closed-loop management of "perception-decision-execution" using prescription maps and variable operation techniques. However, the quality detection and evaluation of plowing, as a preliminary and fundamental step, is generally still in an open-loop state: agronomists can set target furrow parameters, but there is a lack of effective means to verify in real time and comprehensively whether these parameters are executed with high fidelity. This creates a control blind spot in the source stage of precision agriculture, hindering further improvements in overall resource utilization efficiency and operational consistency. Therefore, developing an online, automatic, and precise furrow quality detection method is of key value for achieving closed-loop precision management of plowing and improving the precision agriculture technology system.
[0003] To achieve the above objectives, various technologies have been tried in this field, but a series of unique technical problems remain unresolved: (1) Directly acquired images or point cloud data suffer from geometric distortion, making it difficult to accurately reflect the furrow shape: During the continuous movement of agricultural machinery, the undulation of the farmland surface, engine vibration, and high-frequency impact of the plow and soil cause the lidar or camera sensor to undergo pitch, roll, and other attitude changes with the agricultural machinery, resulting in geometric distortion in the images or point cloud data directly acquired by the lidar or camera sensor; (2) Dependence on manual labor and lack of real-time performance: Existing assessments almost entirely rely on post-event manual sampling, making it impossible to obtain quality feedback while the agricultural machinery is operating, leading to a lag in operation adjustments, which is the primary obstacle to achieving closed-loop control; (3) Insufficient three-dimensional geometric perception capability: Although lidar provides three-dimensional point cloud data, existing agricultural applications mostly focus on discrete object recognition (such as fruit trees and crop rows). Or large-scale terrain modeling, lack of effective algorithms for robust instance segmentation of continuous, strip-shaped furrow structures, complex unstructured field micro-topography and point cloud distortion caused by agricultural machinery vibration further exacerbate data distortion, making simple clustering or thresholding methods easily fail; (4) Algorithm generalization and model mismatch: Deep learning-based solutions are limited by the strong variability of field scenes (such as soil moisture, crop residue, and light changes), resulting in serious domain adaptation problems and high model maintenance costs; while traditional rule-based geometric methods often oversimplify furrow profiles into regular geometric models (such as parabolas), which are seriously inconsistent with the irregular and asymmetrical shapes formed by soil plasticity and agricultural machinery disturbance in real farming, and even if complex models are used, the real shape cannot be recovered from the distorted data when there is already geometric distortion at the data source, resulting in fundamental errors in the extraction of key parameters such as depth and cross-sectional area.
[0004] In summary, the core problem with existing technologies is that the raw point cloud data directly acquired under dynamic agricultural machinery operation conditions has significant geometric distortion and cannot truly reflect the shape of the furrows. At the same time, existing automation technology solutions are not adaptable enough to unstructured field environments and cannot achieve robust, real-time and high-precision perception of the true three-dimensional shape of the furrows, thus hindering the establishment of a closed-loop control system for precision plowing operations. Summary of the Invention
[0005] This invention solves the problems of significant geometric distortion in the raw point cloud data directly acquired under dynamic agricultural machinery operation conditions in the prior art, which cannot truly reflect the shape of the furrows; at the same time, the automation method is not adaptable to complex field terrain and cannot achieve robust, real-time and high-precision perception of the true three-dimensional shape of the furrows.
[0006] A method for correcting 3D point cloud data of farmland furrows based on lidar, wherein the farmland furrow segmentation method includes the following steps:
[0007] S1 collects raw point cloud data of farmland and agricultural machinery attitude information in real time;
[0008] S2, Based on the agricultural machinery attitude information, the original point cloud data is projected onto a standard horizontal reference plane by constructing a rotation and translation matrix to obtain the attitude-compensated point cloud data;
[0009] S3, Extracting furrow boundary features from the pose-compensated point cloud data, including the following steps:
[0010] S31, a local nonmaximum suppression method based on spatial windows, relies solely on the local morphological distribution of points to achieve adaptive recognition and filter furrow boundary feature points;
[0011] S32, use the DBSCAN clustering algorithm to cluster the feature points of the furrow boundary to obtain the clusters of each furrow boundary feature point, and fit the boundary lines of the clusters using the linear least squares method to obtain the left and right boundary lines of each furrow.
[0012] S4. Based on the left and right boundary lines of each furrow, an adaptive region growing algorithm that integrates global linear prior and local geometric consistency is used to perform point cloud clustering and region division on the pose-compensated point cloud data to obtain a three-dimensional point cloud set of a single furrow and reconstruct the three-dimensional point cloud data of farmland furrows.
[0013] In a further preferred embodiment, the method further includes determining the maximum allowable time threshold for single-frame point cloud processing based on the agricultural machinery operating speed, the effective look-ahead distance of the lidar, and the scanning frequency.
[0014] In a further optimized scheme, in S2, projecting the original point cloud data onto a standard horizontal reference plane by constructing a rotation and translation matrix refers to... Through continuous transformations of rotation and translation matrices, the sensor coordinate system is transformed. To local coordinate system Mapping:
[0015] ,
[0016] in, This indicates that the original point cloud data has been transformed into three-dimensional coordinate vector data in the local coordinate system of the farmland. To collect point cloud observation vector data in the sensor coordinate system of the lidar sensor where the raw point cloud data is located, It is a dynamic rotation matrix that changes in real time with the movement of the agricultural machinery. This dynamic rotation matrix is driven by the attitude angles obtained by the inertial measurement unit that collects the attitude information of the agricultural machinery.
[0017] In a further optimized scheme, S4, the adaptive region growing algorithm that integrates global linear prior and local geometric consistency specifically includes:
[0018] S41, Feature Space Construction: Construct a spatial prior field based on principal component analysis, and calculate the local normal vector of each point in the pose-compensated point cloud data;
[0019] S42, Seed initialization: Based on the feature space, extract the global minimum point of the region of interest in the pose-compensated point cloud data as the seed point, and initialize a parallel priority queue for managing the growth process;
[0020] S43, Hybrid cost calculation: Starting from the seed point and the parallel priority queue, the geometric consistency is evaluated through the local normal vector, the spatial weighting weight is obtained through the spatial prior field, and the hybrid clustering cost from the current seed point to its neighboring points is calculated by combining the geometric consistency and the spatial weighting weight.
[0021] S44, Competition Determination: The cost of the hybrid clustering is used as the determination criterion, and the minimum cost principle is applied to resolve the boundary conflict between adjacent furrows;
[0022] S45, Dynamic Update and Output: Based on the local normal vector corresponding to the new point included in the current region according to the competition determination result, dynamically update the reference normal vector representing the surface orientation of the region.
[0023] Iteratively execute steps S43 to S45 until no new points satisfy the clustering conditions, and finally output the set of 3D point clouds corresponding to the independent furrows obtained by clustering.
[0024] In a further optimized approach, S43, the calculation of hybrid growth cost by combining geometric consistency and spatial weighting includes the following steps:
[0025] S431, Construct a spatial distance field based on the left and right boundary lines to determine the first... The center line of the furrow For any furrow boundary feature point Calculate its up to the first The center line of the furrow European vertical distance And based on the Euclidean vertical distance Calculate the point Spatial prior weights :
[0026]
[0027] S432, Define the current seed point To neighboring points Clustering cost function The formula is as follows:
[0028]
[0029] in, Seed point The unit normal vector, For neighborhood points The unit normal vector, Seed point Relative to the height of the ground plane, For neighborhood points The height relative to the ground plane; This is the height difference normalization factor, used to eliminate the influence of dimensions; Acting as a penalty term, when neighboring points This value surges when the cluster moves away from the center line, thus suppressing cluster out-of-bounds clustering.
[0030] In a further optimized scheme, S44, the step of resolving boundary conflicts between adjacent furrows using the minimum cost principle is as follows: for each unclassified neighboring point... It will only be accepted if its clustering cost with the current region is below a certain threshold, that is:
[0031]
[0032] in, It is the global clustering threshold;
[0033] If neighboring points If a region satisfies the clustering criteria for multiple furrow regions, then it is assigned to... The smallest region.
[0034] A LiDAR-based 3D point cloud data correction system for farmland furrows, the system being used to implement the LiDAR-based 3D point cloud data correction method for farmland furrows, the system comprising a data acquisition device and a data processing device:
[0035] The data acquisition device is used to acquire raw point cloud data of farmland by 3D lidar installed on agricultural machinery, acquire agricultural machinery attitude information in real time based on inertial measurement unit (IMU), and transmit the raw point cloud data and agricultural machinery attitude information to the data processing device.
[0036] The data processing device is embedded with a data processing module implemented by a computer program, and the data processing module includes the following data processing units:
[0037] Attitude compensation unit: used to project the original point cloud data onto a standard horizontal reference plane by constructing a rotation and translation matrix based on the agricultural machinery attitude information, so as to obtain the attitude-compensated point cloud data;
[0038] Boundary feature extraction unit: used to extract furrow boundary features from the pose-compensated point cloud data, including the following sub-units:
[0039] The first subunit is used for the local nonmaximum suppression method based on spatial windows, which relies solely on the local morphological distribution of points to achieve adaptive recognition and filter furrow boundary feature points.
[0040] The second subunit is used to cluster the feature points of the furrow boundary using the DBSCAN clustering algorithm to obtain clusters of feature points of each furrow boundary, and to fit the boundary lines of the clusters using the linear least squares method to obtain the left and right boundary lines of each furrow.
[0041] Point cloud clustering and region partitioning unit: Based on the left and right boundary lines of each furrow, an adaptive region growing algorithm that integrates global linear prior and local geometric consistency is used to perform point cloud clustering and region partitioning on the pose-compensated point cloud data to obtain a three-dimensional point cloud set for each furrow.
[0042] The beneficial effects of this invention compared to the prior art are as follows:
[0043] The method and system described in this application address the problem of geometric distortion in raw point cloud data caused by attitude changes during continuous agricultural machinery operation. By fusing lidar and agricultural machinery attitude sensing, the raw point cloud data is dynamically tilted in real time, eliminating point cloud distortion caused by pitch, roll, and other movements of the agricultural machinery. This solves the technical problem that directly collected data cannot accurately reflect the furrow morphology. At the same time, it achieves a fundamental breakthrough in furrow quality detection, moving from manual, post-event sampling to fully automatic, real-time online detection. This method overcomes the shortcomings of traditional methods, which rely on static sampling, are inefficient, and cannot provide real-time feedback during operation. It provides a distortion-free, high-fidelity real-time perception foundation for closed-loop precise control of plowing operations.
[0044] The method described in this application, based on corrected distortion-free point cloud data, employs an adaptive region growing algorithm that integrates spatial prior constraints to achieve accurate clustering and region division of 3D point cloud data of furrows in unstructured field environments. By constructing a spatial potential field to guide clustering and integrating local geometric features with minimum cost competitive decision-making, it effectively overcomes interference from complex micro-topography and adjacent furrows, solving the core problems of general point cloud processing methods being sensitive to distortion and lacking adaptability in this scenario, and achieving accurate reconstruction of the 3D morphology of furrows.
[0045] The method and system described in this application perform orthogonal section analysis based on the reconstructed three-dimensional point cloud set of furrows, achieving high-fidelity and automated extraction of key geometric parameters of furrows. It abandons the traditional approach of simplifying furrow profiles into ideal models, and can directly quantify irregular and asymmetrical real furrow morphology from distortion-free point cloud data that reflects the true morphology. This avoids fundamental errors caused by data source distortion or model simplification, and significantly improves the accuracy and reliability of tillage quality assessment.
[0046] The method and system described in this invention are applicable to fields such as farmland tillage quality inspection, precision agriculture management, and intelligent operation of agricultural machinery under dynamic agricultural machinery operation conditions. Attached Figure Description
[0047] Figure 1 This is a schematic diagram of the boundary lines of each furrow obtained by fitting according to Embodiment 1, wherein the blue line represents the boundary line of furrow A, the orange line represents the boundary line of furrow B, the green line represents the boundary line of furrow C, the red line represents the boundary line of furrow D, and the purple line represents the boundary line of furrow E.
[0048] Figure 2 The candidate point selection strategy for furrow boundary based on geometric posture constraints and local morphological constraints as described in Implementation Method 4; (a) geometric constraints; (b) local morphological constraints; (c) fusion result of geometric constraints and local morphological constraints.
[0049] Figure 3 Here are schematic diagrams of adaptive region growing furrow segmentation with fusion of spatial potential field prior constraints as described in Implementation Method 5; (a) Algorithm flowchart; (b) Schematic diagram of furrow segmentation effect.
[0050] Figure 4 The point cloud data acquisition platform described in Implementation Method Nine; (a) agricultural machinery 1, lidar sensor 2, inertial measurement unit 3; (b) schematic diagram of the vertical view of lidar sensor 2; (c) schematic diagram of the point cloud data acquisition process;
[0051] Figure 5 This is a schematic diagram of the method for calculating the geometric parameters of the furrows using orthogonal slicing and cross-sectional profile analysis as described in Implementation Method Eleven; (a) Orthogonal slicing analysis; (b) Cross-sectional profile analysis;
[0052] Figure 6 This is a visual comparison of the three-dimensional point cloud reconstruction results of furrows obtained by different methods under complex field conditions as described in Implementation Method Twelve; (a) the algorithm proposed in this invention; (b) the classical region growing method; (c) region growing constraint; (d) the coarse segmentation method based on geometric envelope / shape constraint; (e) the fast segmentation method based on relative elevation threshold; in the figure, blue dots represent furrow A, orange dots represent furrow B, green dots represent furrow C, red dots represent furrow D, and purple dots represent furrow E.
[0053] Figure 7 This is a linear regression analysis comparison of the furrow structure parameters extracted from the corrected point cloud data and the manually measured values as described in Implementation Method Twelve. In the figure, the green circles represent sampling points, the horizontal axis is the manually measured value, and the vertical axis is the predicted value output by the algorithm. The solid line represents the linear regression line fitted by the sample points. The dashed line represents the ideal reference line that perfectly matches the predicted value output by the algorithm with the manually measured value.
[0054] Figure 8 The following is an analysis of the variation trend and stability of furrow structure parameters extracted from corrected point cloud data under continuous operation conditions as described in Implementation Method Twelve: (a) mean variation of furrow depth and overall coefficient of variation; (b) mean variation of ridge spacing and overall coefficient of variation; (c) mean variation of furrow width and overall coefficient of variation; (d) mean variation of slope angle and overall coefficient of variation; (e) mean variation of cross-sectional area and overall coefficient of variation. Detailed Implementation
[0055] 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.
[0056] Implementation Method 1: This implementation method provides a method for correcting three-dimensional point cloud data of farmland furrows based on lidar. The method includes the following steps:
[0057] S1 collects raw point cloud data of farmland and agricultural machinery attitude information in real time;
[0058] S2, Based on the agricultural machinery attitude information, the original point cloud data is projected onto a standard horizontal reference plane by constructing a rotation and translation matrix to obtain the attitude-compensated point cloud data;
[0059] S3, Extracting furrow boundary features from the pose-compensated point cloud data, including the following steps:
[0060] S31, a local nonmaximum suppression method based on spatial windows, relies solely on the local morphological distribution of points to achieve adaptive recognition and filter furrow boundary feature points;
[0061] S32, use the DBSCAN clustering algorithm to cluster the feature points of the furrow boundary to obtain the clusters of each furrow boundary feature point, and fit the boundary lines of the clusters using the linear least squares method to obtain the left and right boundary lines of each furrow.
[0062] The fitting process yields the left and right boundary lines of each furrow, as shown in the figure. Figure 1 As shown;
[0063] S4. Based on the left and right boundary lines of each furrow, an adaptive region growing algorithm that integrates global linear prior and local geometric consistency is used to perform point cloud clustering and region division on the pose-compensated point cloud data to obtain a three-dimensional point cloud set of a single furrow and reconstruct the three-dimensional point cloud data of farmland furrows.
[0064] In this embodiment, the raw point cloud data of the farmland can be obtained by a lidar sensor installed on the agricultural machinery. The attitude information of the agricultural machinery can be obtained by an inertial measurement unit (IMU) installed on the agricultural machinery.
[0065] Implementation Method 2: This implementation method is a further limitation of Implementation Method 1. In S2, the projection of the original point cloud data onto a standard horizontal reference plane by constructing a rotation and translation matrix is illustrated as an example.
[0066] In agricultural plowing operations, the raw point cloud data acquired is inevitably affected by changes in vehicle attitude: on the one hand, the unevenness of the farmland surface, the periodic changes in plowing resistance, and the differences in tire movement will cause low-frequency attitude tilting of the agricultural machinery in the pitch and roll directions; on the other hand, engine vibration, high-frequency contact impact between the plow and the soil, and the breaking process of irregular soil clods will lead to short-term, high-frequency attitude shaking; if no attitude compensation is performed, this dynamic tilt will cause geometric distortion in the generated point cloud, which will seriously interfere with the measurement accuracy of furrow depth and slope.
[0067] To eliminate motion distortion and unify the benchmark, this implementation uses an inertial measurement unit (IMU) to acquire the attitude information of the agricultural machinery in real time, and projects the original point cloud onto a standard horizontal reference plane by constructing a rotation and translation matrix; the external parameter relationship between its measurement coordinate system and the lidar coordinate system is obtained through static calibration before the experiment and remains unchanged throughout the experiment.
[0068] The step of projecting the original point cloud data onto a standard horizontal reference plane by constructing a rotation and translation matrix refers to projecting the original point cloud data... Through continuous transformations of rotation and translation matrices, the sensor coordinate system is transformed. To local coordinate system Mapping:
[0069] ,
[0070] in, This indicates that the original point cloud data has been transformed into three-dimensional coordinate vector data in the local coordinate system of the farmland. To collect point cloud observation vector data in the sensor coordinate system of the lidar sensor where the raw point cloud data is located, It is a dynamic rotation matrix that changes in real time with the movement of the agricultural machinery. This dynamic rotation matrix is driven by the attitude angles obtained by the inertial measurement unit that collects the attitude information of the agricultural machinery.
[0071] Implementation Method 3: This implementation method further defines Implementation Method 2, specifically the dynamic rotation matrix. Let's illustrate with examples.
[0072] The dynamic rotation matrix Used to correct the roll angle of agricultural machinery during operation. Pitch angle and yaw angle For fluctuations, the rotation order is defined using the right-hand rule, and the component rotation matrices are as follows:
[0073] ,
[0074] ,
[0075] ,
[0076] in The roll angle is the angle of the agricultural machinery about the forward direction axis, i.e., the X-axis, representing the tilt of the machinery caused by uneven soil hardness on the left and right sides. The pitch angle of the agricultural machinery is about its transverse axis, the Y-axis, representing the vertical movement of the front of the machinery due to the undulations of the ground. The yaw angle is the angle of the agricultural machinery around the vertical axis, i.e., the Z-axis, which represents the yaw of the working path.
[0077] Implementation Method 4: This implementation method is a further limitation of Implementation Method 1, and provides an example of the screening of furrow boundary feature points in S31.
[0078] In unstructured farmland environments, relying solely on local height thresholds may be affected by outliers such as overall slope, local subsidence, or inconsistent ridge height, resulting in the feature point set containing pseudo-high points that do not have ridge features. To address the soil features at furrow boundaries, this implementation proposes a local nonmaximum suppression method based on spatial windows, transforming ridge top extraction into an unsupervised clustering problem driven by the local differential geometric attributes of the point cloud, achieving adaptive recognition solely based on the local morphological distribution of points.
[0079] The process of selecting furrow boundary feature points includes the following steps:
[0080] S311, the preprocessed point cloud data Construct a Kd-tree index for any point Search for it Near-neighbor analysis and calculation of local covariance matrix. and the angle between the normal vector and the vertical axis The formula is as follows:
[0081]
[0082]
[0083] in As the center of gravity of the domain, for Perform eigenvalue decomposition to obtain eigenvalues. and the corresponding unit eigenvector , , ,in The direction corresponding to the smallest eigenvalue can be used as an estimate of the local normal vector. ;
[0084] S312, In order to accurately locate the ridge top frame, regarding the included angle... and the height Construct a joint discriminant function, such as Figure 2 As shown, a point is marked as a candidate set of furrow boundary points if and only if it simultaneously satisfies the following geometric attitude constraints and local shape constraints. :
[0085]
[0086]
[0087] in The normal angle threshold is used to determine the attitude constraint conditions and exclude slopes and noise. To filter out slope interference, this test is used to determine whether the point is part of the ridge top skeleton, ensuring that the selected points are the highest local points that conform to the characteristics of the ridge top. To preserve local high points for morphological constraints Centered on, with radius cylindrical neighborhood Within it, check whether it is a local maximum.
[0088] Implementation Method 5: This implementation method further defines Implementation Method 1 and provides an example of the adaptive region growing algorithm that integrates global linear prior and local geometric consistency in S4.
[0089] To accurately extract 3D point cloud data of furrows from unstructured farmland point cloud data, simply relying on Euclidean distance clustering or global planar segmentation cannot solve the problem of complex micro-topographic interference in the field. This implementation proposes an adaptive region growing algorithm that integrates global linear prior and local geometric consistency. Based on the extracted left and right boundary lines of the entire furrow, the algorithm constructs a spatial potential field, thereby dynamically adjusting the growth criteria and constructing a spatial prior weighted field based on the distance field, which improves the robustness and adaptability of the segmentation.
[0090] like Figure 3 As shown in (a) above, the adaptive region growing algorithm that integrates global linear prior and local geometric consistency specifically includes:
[0091] S41, Feature Space Construction: Construct a spatial prior field based on principal component analysis, and calculate the local normal vector of each point in the pose-compensated point cloud data;
[0092] S42, Seed initialization: Based on the feature space, extract the global minimum point of the region of interest in the pose-compensated point cloud data as the seed point, and initialize a parallel priority queue for managing the growth process;
[0093] S43, Hybrid cost calculation: Starting from the seed point and the parallel priority queue, the geometric consistency is evaluated through the local normal vector, the spatial weighting weight is obtained through the spatial prior field, and the hybrid clustering cost from the current seed point to its neighboring points is calculated by combining the geometric consistency and the spatial weighting weight.
[0094] S44, Competition Determination: The cost of the hybrid clustering is used as the determination criterion, and the minimum cost principle is applied to resolve the boundary conflict between adjacent furrows;
[0095] S45, Dynamic Update and Output: Based on the local normal vector corresponding to the new point included in the current region according to the competition determination result, dynamically update the reference normal vector representing the surface orientation of the region.
[0096] Iteratively execute steps S43 to S45 until no new points satisfy the clustering conditions, and finally output the set of 3D point clouds corresponding to the independent furrows obtained by clustering.
[0097] The diagram illustrates the clustering and region partitioning effects of furrow point clouds. Figure 3 As shown in (b) of the diagram.
[0098] Implementation Method Six: This implementation method is a further limitation of Implementation Method Five. It provides an example of how the calculation of hybrid growth cost in S43 combines geometric consistency and spatial weighting.
[0099] The method of calculating hybrid growth clustering by combining geometric consistency and spatial weighting includes the following steps:
[0100] S431, Construct a spatial distance field based on the left and right boundary lines to determine the first... The center line of the furrow For any furrow boundary feature point Calculate its up to the first The center line of the furrow European vertical distance And based on the Euclidean vertical distance Calculate the point Spatial prior weights :
[0101]
[0102] S432, Define the current seed point To neighboring points Clustering cost function The formula is as follows:
[0103]
[0104] in, Seed point The unit normal vector, For neighborhood points The unit normal vector, Seed point Relative to the height of the ground plane, For neighborhood points The height relative to the ground plane; This is the height difference normalization factor, used to eliminate the influence of dimensions; Acting as a penalty term, when neighboring points This value surges when the cluster moves away from the center line, thus suppressing cluster out-of-bounds clustering.
[0105] Implementation Method Seven: This implementation method is a further limitation of Implementation Method Five. It provides an example of how the principle of minimum cost is used to resolve boundary conflicts between adjacent furrows in S44.
[0106] The method of resolving boundary conflicts between adjacent furrows using the minimum cost principle is as follows: for each unclassified neighboring point... It will only be accepted if its clustering cost with the current region is below a certain threshold, that is:
[0107]
[0108] in, It is the global clustering threshold;
[0109] If neighboring points If a region satisfies the clustering criteria for multiple furrow regions, then it is assigned to... The smallest region.
[0110] Implementation Method 8: This implementation method provides a farmland furrow segmentation system based on lidar. The system is used to implement the lidar-based three-dimensional point cloud data correction method for farmland furrows. The system includes a data acquisition device and a data processing device, which are installed on the agricultural machinery.
[0111] The data acquisition device is used to acquire raw point cloud data of farmland by 3D lidar installed on agricultural machinery, acquire agricultural machinery attitude information in real time based on inertial measurement unit (IMU), and transmit the raw point cloud data and agricultural machinery attitude information to the data processing device.
[0112] The data processing device is embedded with a data processing module implemented by a computer program, and the data processing module includes the following data processing units:
[0113] Attitude compensation unit: used to project the original point cloud data onto a standard horizontal reference plane by constructing a rotation and translation matrix based on the agricultural machinery attitude information, so as to obtain the attitude-compensated point cloud data;
[0114] Boundary feature extraction unit: used to extract furrow boundary features from the pose-compensated point cloud data, including the following sub-units:
[0115] The first subunit is used for the local nonmaximum suppression method based on spatial windows, which relies solely on the local morphological distribution of points to achieve adaptive recognition and filter furrow boundary feature points.
[0116] The second subunit is used to cluster the feature points of the furrow boundary using the DBSCAN clustering algorithm to obtain clusters of feature points of each furrow boundary, and to fit the boundary lines of the clusters using the linear least squares method to obtain the left and right boundary lines of each furrow.
[0117] Point cloud clustering and region partitioning unit: Based on the left and right boundary lines of each furrow, an adaptive region growing algorithm that integrates global linear prior and local geometric consistency is used to perform point cloud clustering and region partitioning on the pose-compensated point cloud data to obtain a three-dimensional point cloud set for a single furrow.
[0118] Parameter extraction unit: Based on the three-dimensional point cloud data of a single furrow in the farmland, it performs orthogonal slicing and cross-sectional contour analysis along the furrow direction to extract the structural parameters of the furrow.
[0119] Implementation Method Nine: This implementation method further defines Implementation Method Eight and provides examples of the data acquisition device and data processing device.
[0120] like Figure 4 As shown in (a), the data acquisition device includes agricultural machinery 1, lidar sensor 2 and inertial measurement unit 3;
[0121] The agricultural machinery 1 has a height of 2.1 m and an average travel speed of 2.3 m / s;
[0122] LiDAR sensor 2 is a Robosense-Helios32 LiDAR sensor, installed directly in front of the agricultural machinery to maximize forward field of view; such as Figure 4 As shown in (b), the lidar sensor 2 has a 360° horizontal field of view and a 70° vertical field of view, a maximum detection distance of 200 m, a ranging random error of less than 2 cm, an angle random error of less than 0.05°, 32 beams, a horizontal angular resolution of 0.1°, a vertical angular resolution of 1°, and a maximum point cloud output rate of 1,200,000 points / s; the lidar sensor 2 communicates with the data processing device through a network port.
[0123] The inertial measurement unit 3 and data processing device are installed in the cockpit;
[0124] The data processing device uses an Intel Core i7-7700HQ processor (2.80 GHz), a 64-bit operating system, and 8 GB of RAM; the algorithm implementation is based on the Robot Operating System (ROS) under the Ubuntu 16.04 environment and is developed using the Python language.
[0125] A schematic diagram of the point cloud data acquisition process is shown below. Figure 4 As shown in (c) in the figure.
[0126] Implementation Method 10: This implementation method is a further limitation of Implementation Methods 1 to 9. The methods and systems described in Implementation Methods 1 to 9 also include determining the maximum allowable time threshold for single-frame point cloud processing based on the agricultural machinery operating speed, the effective look-ahead distance of the lidar, and the scanning frequency.
[0127] In the continuous operation of agricultural machinery plowing, in order to eliminate the impact of perception lag caused by the time consumption of point cloud computing on closed-loop control, the algorithm must meet strict time constraints.
[0128] From the moment the lidar sensor collects a frame of data until the algorithm outputs the corresponding phenotypic parameters or control commands for that frame, the agricultural machinery's displacement along the working direction must not exceed the algorithm's effective look-ahead distance; based on agronomic requirements and equipment parameters, the standard is the agricultural machinery's operating speed (3km / h-8km / h).
[0129] Robosense-Helios32 LiDAR sensor's forward effective scanning range 5m, output frequency The basic time constraint based on the scanning frequency is 5Hz:
[0130]
[0131] To ensure the safety and real-time performance of the operation, the maximum allowable delay in single-frame point cloud processing (including preprocessing, segmentation, and parameter extraction) is set at [number of frames]. for:
[0132]
[0133] Considering the delay processing mechanism, the maximum allowed processing time It can be extended to:
[0134]
[0135] The safety factor is 0.8, which takes into account uncertainties such as bus data transmission overhead, computational load fluctuations, and buffer management, and reserves a safety margin.
[0136] Implementation Method Eleven: This implementation method describes the analytical method for verifying the accuracy of the correction method of the present invention.
[0137] To verify that the three-dimensional point cloud data of the furrows obtained after correction by the present invention can truly reflect the actual shape of the furrows, this embodiment performs orthogonal slicing and cross-sectional contour analysis along the furrow direction based on the corrected three-dimensional point cloud data of the furrows, extracts the structural parameters of the furrows, and compares and analyzes the extracted structural parameters with the manually measured values.
[0138] The orthogonal slice analysis is as follows: Figure 5 As shown in (a) above, the cross-sectional profile analysis is as follows: Figure 5 As shown in (b) of the diagram; the structural parameters include the extraction of furrow width, furrow depth, cross-sectional area, slope steepness, and ridge spacing;
[0139] Specifically:
[0140] The point cloud was sliced at fixed intervals of 0.05m along the direction of the furrows, and the data falling into each slice was analyzed. The points within the groove are projected onto the YZ plane to form a discrete set of points representing the cross-sectional profile of the furrow. For each set of cross-section points The search has a minimum relative height. The point is defined as the bottom point of the furrow on that cross section. ;
[0141] Based on the equations of the left and right boundary lines of the furrow, calculate the theoretical boundary position. and In the set of cross-sectional profile points of the furrow Find the distance from the specified distance in each of the following steps. and The nearest actual point is used as the left and right boundary points. and The average elevation of the left and right boundary points is selected as the reference height of the ridge top for this section:
[0142] ;
[0143] Furrow width: Furrow width Defined as and Euclidean distance on the Y-axis:
[0144]
[0145] Furrow depth: Furrow depth Calculate the elevation difference between the reference plane and the bottom of the furrow:
[0146]
[0147] Cross-sectional area: Calculated using the shoelace formula from the set of points on the cross-sectional profile of the furrow. Area of the polygon formed, cross-sectional area The calculation formula is as follows:
[0148]
[0149] Slope steepness: for the point set on the left slope respectively to and right slope point set to The least squares linear regression model is as follows:
[0150]
[0151] The inclination angles of its left and right slopes are:
[0152]
[0153] Row spacing: the spacing at each section of the extracted boundary line. right boundary of the furrow With the left boundary of the furrow The distance between them is used to calculate the average of the spacing at all cross-sections, thus obtaining the ridge spacing parameter. The formula is as follows:
[0154]
[0155] Implementation Method Twelve: This implementation method uses experimental data to illustrate the beneficial effects of the present invention, in order to verify the accuracy and reliability of the correction method of the present invention.
[0156] To systematically evaluate the performance of this invention in the correction and structural parameter extraction of three-dimensional point cloud data of farmland furrows, this embodiment uses two sets of complementary indicators: one set is used to measure the consistency between the predicted furrow area in the corrected point cloud data and the manually labeled area. Precision ,recall , Another set of parameters is used to quantify the fit and error level between the predicted and measured values of structural parameters extracted from the corrected point cloud data. , In the furrow segmentation and boundary line detection experiment, this implementation method manually divided 50 groups of samples from the complete point cloud dataset, and two independent operators manually annotated them to cover different working sections and soil surface morphology changes; all indicators were calculated based on a sample-by-sample comparison with the actual values measured on the ground, and statistical summaries were performed on the test data.
[0157] The extraction of furrows is described as a binary classification problem: each individual point cloud is classified as either a furrow or a non-furrow, and a true positive is defined based on the correspondence between the prediction results and the manually labeled data. ), false positives ( ) and false negatives ( )as follows: This represents the number of samples that were correctly identified as furrows. This represents the number of non-furrow samples that were misclassified as furrows; This represents the number of samples that actually have furrows but were not detected.
[0158] Based on this, the calculation formula is as follows:
[0159]
[0160]
[0161]
[0162]
[0163] For the key geometric parameters of the furrow, let the first... The true value of each sample Predicted value The sample size is The true mean is The formula is:
[0164]
[0165]
[0166] 1. Evaluation of the effectiveness of different algorithms for furrow segmentation
[0167] To verify the effectiveness and robustness of the proposed segmentation framework in real-world unstructured point clouds, this implementation method selects three representative traditional segmentation strategies for comparison: Geometric Envelope (a coarse segmentation method based on geometric envelope / shape constraints), Height Threshold (a fast segmentation method based on relative elevation thresholds), and Region Growing (a classic region growing method). Furthermore, to verify the specific contributions of morphological constraints and global linear priors, RegionGrowing restrained is introduced as a key ablation control group.
[0168] The results in Table 1 show that different algorithms exhibit significant differences in precision and recall. The fast segmentation method based on relative elevation thresholds achieves higher recall. (0.8845), but its accuracy (0.8055) and The relatively low value of (0.7288) is mainly due to the method's sensitivity to surface undulations. A simple elevation criterion can easily misclassify broken soil clods or localized uplifts on ridges as furrow areas, leading to... The increase in complexity significantly hinders the consistency of segmentation; coarse segmentation methods based on geometric envelope / shape constraints and classical region growing methods differ in... and The lack of significant advantages in the metrics indicates that, in the absence of global structural constraints, coarse segmentation methods based on geometric envelope / shape constraints tend to over-smooth and weaken boundary details. Classical region growing methods, which mainly rely on local consistency, are prone to leakage or breakage at furrow boundaries and lack robustness to outliers and morphological abrupt changes. In contrast, region growing constraints show a leap in all four metrics, indicating that explicitly incorporating spatial priors based on boundary lines into the growth criteria can construct effective soft boundaries, thereby limiting the diffusion of the segmentation process into non-target regions and significantly reducing the risk of erroneous merging caused by local noise.
[0169] Table 1
[0170]
[0171] The algorithm proposed in this invention achieves optimal overall performance. Compared with the best-performing fast segmentation method based on relative elevation thresholds among the three traditional baselines, the method of this invention... and The improvements of 0.0899 and 0.1457 respectively indicate that the method of this invention not only better reflects the actual furrow morphology in terms of point-level consistency, but also has a significant advantage in handling unstructured farmland disturbances. Furthermore, compared to region growth constraints, the algorithm proposed in this invention achieves higher accuracy. , ,recall and The gains were increased by 0.0251, 0.0256, 0.0246, and 0.0148 respectively. This further gain indicates that, under the premise of establishing basic structural constraints, the refined processing that incorporates local morphological consistency still has a significant positive effect on suppressing minor anomalies at boundary details.
[0172] In terms of computational efficiency, the fast segmentation method based on relative elevation threshold achieves the shortest running time due to its simple decision rule based on altitude, while the coarse segmentation method based on geometric envelope / shape constraints also maintains low computational cost. Region growth and region growth constraint require additional processing time for neighborhood consistency checks and restricted growth, respectively. Although the proposed method introduces boundary fitting and spatial prior-guided growth of the constrained region, it is still less than the estimated maximum processing time of 1.6s when processing a single frame point cloud. Therefore, the algorithm proposed in this invention is sufficient to meet the requirements of real-time applications.
[0173] Figure 6 This section presents a visual comparison of the 3D point cloud reconstruction results of furrows obtained by different methods under complex field conditions. For example... Figure 6 As shown in (e), the height threshold often over-cuts out local soil blocks and surface undulations, leading to false alarms at ridge tops; as Figure 6 As shown in (d), geometric envelopes can produce smoother groove regions, but they typically cannot preserve boundary details; as Figure 6 As shown in (c), regional growth improves local continuity, but leakage problems still exist near the blurred ditch boundaries; in contrast, as Figure 6 As shown in (a), the method proposed in this invention achieves more complete and consistent furrow segmentation, clearer boundary delineation, and reduced misclassification, benefiting from the integration of boundary-based spatial priors and constrained region growth.
[0174] 2. Results of extraction of furrow geometric structure parameters
[0175] To quantitatively evaluate the accuracy of structural parameter extraction, following a manual data collection process, 20 typical locations along the work route were manually measured using stratified random sampling. The measurements were then compared sample-by-sample with the furrow structure parameters extracted from corrected point cloud data at the same locations. The regression fitting results for each parameter are shown below. Figure 7 As shown.
[0176] Overall, furrow depth ( Figure 7 (a) in the middle), furrow width ( Figure 7 (c) in the middle and the slope angle ( Figure 7 The regression consistency of (d) is relatively high. They reached 0.956, 0.963, and 0.953 respectively, and All are controlled within 2.5cm (including the slope angle). (Approximately 1°), indicating that even under unstructured micro-topography and echo noise interference in the field, this method can still stably recover the core geometric features of a single furrow; in contrast, the extraction of composite parameters is slightly more difficult, particularly the ridge spacing ( Figure 7 (b) of It is 0.895. The value was 4.76 cm. Although its accuracy was slightly lower than that of the single boundary parameter mentioned above, the level of variation remained within a low range. The regression slope of 0.844 and the intercept of 7.98 cm reflect a systematic bias in some sampling locations. This is mainly due to the fact that the calculation of the ridge spacing involves the joint positioning of the boundaries of two adjacent furrows, and the small positional errors in the process of double boundary identification have accumulated.
[0177] Cross-sectional area of furrows ( Figure 7 (e) in the text comprehensively reflects the combined indicators of furrow depth, width, and internal micro-topography. It is 0.868. It is 88.30cm 2 The regression slope of this parameter is 0.795, showing a relatively obvious decreasing trend. This is mainly attributed to the dual effects of error propagation and sampling discretization. In addition, the non-uniformity of point cloud distribution caused by irregular soil clods makes polygon area estimation quite sensitive to sampling density. The cross-sectional area calculation involves the integration of all sampling points on the contour, and small deviations in boundary positioning and depth estimation will be amplified in the area calculation. Despite the above physical limitations, this level of accuracy is still sufficient to support the detection of tillage quality differences and consistency evaluation at the field scale.
[0178] 3. Analysis of Variation and Stability of Furrow Structure Parameters Extracted from Corrected Point Cloud Data under Continuous Operation Conditions: Analysis of Variation Trends and Stability along the Furrow Flow Flow.
[0179] To evaluate the variation trend and stability of furrow structure parameters extracted from corrected point cloud data under continuous operation conditions, 100 consecutively acquired point cloud frames were analyzed within a stable tillage zone with an actual length of approximately 19.14 m (Y-axis coordinate range of 473 cm to 2387 cm). This zone covers typical conditions of smooth straight-line tractor travel and can fully reflect the algorithm's comprehensive performance in dealing with soil micro-topographic undulations, mechanical vibrations, and continuous changes in sensor attitude. Based on the proposed algorithm, the mean changes and overall coefficient of variation of furrow top width, depth, cross-sectional area, left and right slope angles, and row spacing were extracted frame by frame. The results are as follows: Figure 8 As shown.
[0180] from Figure 8 As can be seen from (a) in the figure, the mean standard deviation and coefficient of variation (CV) of the furrow depth are 26.37 ± 0.12 cm and 6.17%, respectively, indicating that the algorithm can effectively filter out the interference of surface undulations on relative depth measurement and accurately quantify the consistency of the plow body's penetration depth. Figure 8 (b) and Figure 8 In (c), the ridge spacing and furrow top width characterize the lateral structural features of the ridge, which are relatively stable with coefficients of variation of 7.08% and 3.53%, respectively. The fluctuation of furrow top width (SD = 0.16 cm) is mainly affected by the degree of soil fragmentation and the irregularity of soil clod edges. Figure 8 (d) shows that the mean inclination angles of the left and right sidewalls are 44.42° and 43.08°, respectively. The slight difference between the two may be related to the direction of plowing and soil anisotropy. Although the slope parameter is most sensitive to local soil clod loss, its coefficient of variation is still controlled at about 4.57%, indicating that the ditch wall formation quality after plowing is relatively uniform and no large-scale collapse or backfilling phenomenon has occurred. Figure 8 The cross-sectional area of the furrow (e) comprehensively reflects the depth of penetration of the plow body into the soil, the width of the tillage, and the degree of sidewall breakage. Its mean and standard deviation are 741.88 cm². 2 and 69.45cm 2 The area variation coefficient was slightly higher at 11.34%, which is because it accumulates small fluctuations in various boundary features compared to a single linear feature index. Nevertheless, the variation level is still below 15%, indicating that the amount of soil turning during the operation maintained good uniformity, which is conducive to the uniform growth of subsequent crop roots.
Claims
1. A method for correcting laser radar-based three-dimensional point cloud data of a farmland furrow, characterized by, The method includes the following steps: S1 collects raw point cloud data of farmland and agricultural machinery attitude information in real time; S2, Based on the agricultural machinery attitude information, the original point cloud data is projected onto a standard horizontal reference plane by constructing a rotation and translation matrix to obtain the attitude-compensated point cloud data; S3, Extracting furrow boundary features from the pose-compensated point cloud data, including the following steps: S31, a local nonmaximum suppression method based on spatial windows, relies solely on the local morphological distribution of points to achieve adaptive recognition and filter furrow boundary feature points; S32, use the DBSCAN clustering algorithm to cluster the feature points of the furrow boundary to obtain the clusters of each furrow boundary feature point, and fit the boundary lines of the clusters using the linear least squares method to obtain the left and right boundary lines of each furrow. S4. Based on the left and right boundary lines of each furrow, an adaptive region growing algorithm that integrates global linear prior and local geometric consistency is used to perform point cloud clustering and region division on the pose-compensated point cloud data to obtain a three-dimensional point cloud set of a single furrow and reconstruct the three-dimensional point cloud data of a single furrow in the farmland.
2. The method for correcting three-dimensional point cloud data of farmland furrows based on lidar according to claim 1, characterized in that, The method also includes determining the maximum allowable time threshold for single-frame point cloud processing based on the agricultural machinery operating speed, the effective look-ahead distance of the lidar, and the scanning frequency.
3. The method for correcting three-dimensional point cloud data of farmland furrows based on lidar according to claim 1, characterized in that, In S2, the step of projecting the original point cloud data onto a standard horizontal reference plane by constructing a rotation and translation matrix refers to projecting the original point cloud data... Through continuous transformations of rotation and translation matrices, the sensor coordinate system is transformed. To local coordinate system Mapping: ， in, This indicates that the original point cloud data has been transformed into three-dimensional coordinate vector data in the local coordinate system of the farmland. To collect point cloud observation vector data in the sensor coordinate system of the lidar sensor where the raw point cloud data is located, It is a dynamic rotation matrix that changes in real time with the movement of the agricultural machinery. This dynamic rotation matrix is driven by the attitude angles obtained by the inertial measurement unit that collects the attitude information of the agricultural machinery.
4. The method for correcting three-dimensional point cloud data of farmland furrows based on lidar according to claim 1, characterized in that, In S4, the adaptive region growing algorithm that integrates global linear prior and local geometric consistency specifically includes: S41, Feature Space Construction: Construct a spatial prior field based on principal component analysis, and calculate the local normal vector of each point in the pose-compensated point cloud data; S42, Seed initialization: Based on the feature space, extract the global minimum point of the region of interest in the pose-compensated point cloud data as the seed point, and initialize a parallel priority queue for managing the growth process; S43, Hybrid cost calculation: Starting from the seed point and the parallel priority queue, the geometric consistency is evaluated through the local normal vector, the spatial weighting weight is obtained through the spatial prior field, and the hybrid clustering cost from the current seed point to its neighboring points is calculated by combining the geometric consistency and the spatial weighting weight. S44, Competition Determination: The cost of the hybrid clustering is used as the determination criterion, and the minimum cost principle is applied to resolve the boundary conflict between adjacent furrows; S45, Dynamic Update and Output: Based on the local normal vector corresponding to the new point included in the current region according to the competition determination result, dynamically update the reference normal vector representing the surface orientation of the region. Iteratively execute steps S43 to S45 until no new points satisfy the clustering conditions, and finally output the set of 3D point clouds corresponding to the independent furrows obtained by clustering.
5. The method for correcting three-dimensional point cloud data of farmland furrows based on lidar according to claim 4, characterized in that, In S43, the calculation of hybrid clustering cost by combining geometric consistency and spatial weighting includes the following steps: S431, Construct a spatial distance field based on the left and right boundary lines to determine the first... The center line of the furrow For any furrow boundary feature point Calculate its up to the first The center line of the furrow European vertical distance And based on the Euclidean vertical distance Calculate the point Spatial prior weights : S432, Define the current seed point To neighboring points Clustering cost function The formula is as follows: in, Seed point The unit normal vector, For neighborhood points The unit normal vector, Seed point Relative to the height of the ground plane, For neighborhood points The height relative to the ground plane; This is the height difference normalization factor, used to eliminate the influence of dimensions; Acting as a penalty term, when neighboring points This value surges when the cluster moves away from the center line, thus suppressing cluster out-of-bounds clustering.
6. The method for correcting three-dimensional point cloud data of farmland furrows based on lidar according to claim 4, characterized in that, In S44, the method of resolving boundary conflicts between adjacent furrows using the minimum cost principle is as follows: for each unclassified neighboring point... It will only be accepted if its clustering cost with the current region is below a certain threshold, that is: in, It is the global clustering threshold; If neighboring points If a region satisfies the clustering criteria for multiple furrow regions, then it is assigned to... The smallest region.
7. A three-dimensional point cloud data correction system for farmland furrows based on lidar, characterized in that, The system is used to implement the method described in any one of claims 1 to 6, and the system includes a data acquisition device and a data processing device: The data acquisition device is used to acquire raw point cloud data of farmland by 3D lidar installed on agricultural machinery, acquire agricultural machinery attitude information in real time based on inertial measurement unit (IMU), and transmit the raw point cloud data and agricultural machinery attitude information to the data processing device. The data processing device is embedded with a data processing module implemented by a computer program, and the data processing module includes the following data processing units: Attitude compensation unit: used to project the original point cloud data onto a standard horizontal reference plane by constructing a rotation and translation matrix based on the agricultural machinery attitude information, so as to obtain the attitude-compensated point cloud data; Boundary feature extraction unit: used to extract furrow boundary features from the pose-compensated point cloud data, including the following sub-units: The first subunit is used for the local nonmaximum suppression method based on spatial windows, which relies solely on the local morphological distribution of points to achieve adaptive recognition and filter furrow boundary feature points. The second subunit is used to cluster the feature points of the furrow boundary using the DBSCAN clustering algorithm to obtain clusters of feature points of each furrow boundary, and to fit the boundary lines of the clusters using the linear least squares method to obtain the left and right boundary lines of each furrow. Point cloud clustering and region partitioning unit: Based on the left and right boundary lines of each furrow, an adaptive region growing algorithm that integrates global linear prior and local geometric consistency is used to perform point cloud clustering and region partitioning on the pose-compensated point cloud data to obtain a three-dimensional point cloud set for each furrow.