A UAV Path Planning Method for 3D Reconstruction of Fruit Trees Based on Outer Contour Guidance
By using an outer contour-guided method, the point cloud of the outer contour of the fruit tree is extracted and normal estimation and consistency correction are performed. Candidate viewpoints are generated and UAV flight trajectories are generated in combination with safety constraints. This solves the problems of severe occlusion, poor path adaptability and insufficient trajectory safety in the 3D reconstruction of fruit trees, and achieves efficient and safe data acquisition.
Patent Information
- Application Number
- CN202611399556.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-10
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2046-09-10
AI Technical Summary
Existing technologies for 3D reconstruction of fruit trees suffer from severe occlusion, poor path adaptability, excessive repeated observations, and insufficient trajectory security, resulting in incomplete data collection and low efficiency.
The method based on outer contour guidance is adopted. The original point cloud data of fruit trees is obtained, and dilation, voxelization and point number filtering are performed to extract the outer contour point cloud and construct the outer contour voxel center point set. Normal estimation and consistency correction are performed to generate candidate viewpoints and establish coverage relationship. Combined with spatial dispersion constraints and security constraints, the flight trajectory of UAV is generated.
This improved the integrity and efficiency of data acquisition for 3D reconstruction of fruit trees, ensured flight safety and path smoothness, and generated a high-coverage, low-redundancy data acquisition path.
Smart Images

Figure CN122943858B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of smart agriculture, autonomous drone operation, and 3D reconstruction technology, specifically to a drone path planning method for 3D reconstruction of fruit trees based on outer contour guidance. Background Technology
[0002] High-quality 3D reconstruction at the individual tree scale is a key foundational technology in precision agriculture, applicable to phenotypic analysis, yield assessment, and intelligent orchard management. A prerequisite for achieving high-precision 3D reconstruction is acquiring complete, accurate, and multi-angle 3D observation data of the tree. While existing ground platforms have been used for tree-scale and row-scale reconstruction, they are often limited by insufficient viewing angles, severe canopy obstruction, slow data acquisition speeds, and insufficient adaptability to complex terrain in orchard environments, easily leading to incomplete reconstructions and unstable data quality.
[0003] Multi-rotor drones, with their high maneuverability and ability to acquire tree information from multiple angles, effectively reduce ground-based occlusion and improve data collection coverage, thus gradually becoming an important technical approach for 3D reconstruction of fruit trees. However, in close-range reconstruction scenarios of single fruit trees, drones need to perform high-frequency viewpoint adjustments around irregular canopies while simultaneously meeting requirements such as sufficient coverage, obstacle avoidance safety, and path efficiency. Due to the dense foliage, severe self-occlusion, and complex depth variations in fruit trees, traditional preset flight paths or uniform circling flight strategies are difficult to adaptively adjust to the actual geometry of the tree, easily leading to problems such as missed images, excessive repeated observations, and significant path redundancy.
[0004] In existing technologies, methods that directly generate candidate viewpoints based on raw point clouds or meshes often fail to distinguish between redundant internal structures and external observation structures of fruit trees, easily introducing invalid viewpoints. Conventional normal estimation results suffer from directional ambiguity, potentially leading to incorrect orientation of candidate viewpoints and affecting coverage relationship calculations. Greedy point selection strategies based solely on coverage gain easily result in local viewpoint clustering and duplicate observations. When viewpoint selection and path planning are separated, it is difficult to simultaneously consider coverage, path length, trajectory smoothness, and flight safety. Therefore, it is necessary to provide a UAV path planning method capable of adaptive perception and planning based on the real structure of individual fruit trees, to improve the data acquisition integrity, path safety, and operational efficiency of fruit tree 3D reconstruction. Summary of the Invention
[0005] Based on the technical problems mentioned above, this invention aims to overcome the problems of severe occlusion, poor path adaptability, excessive repeated observations, and insufficient trajectory safety in the existing technology for the acquisition of data by UAVs in the three-dimensional reconstruction of single fruit trees. It provides a path planning method for UAVs in the three-dimensional reconstruction of fruit trees based on the outer contour, so as to achieve high coverage, low redundancy, safe and smooth data acquisition path generation for complex canopy structures, thereby improving the data integrity, acquisition efficiency and flight operation reliability of the three-dimensional reconstruction of fruit trees.
[0006] To achieve the above objectives, this invention provides a UAV path planning method for 3D reconstruction of fruit trees based on outer contour guidance, comprising: S1. Obtain the raw point cloud data of the target fruit tree; S2. Perform dilation, voxelization, point number filtering, and six-neighbor classification on the original point cloud data to extract the outer contour point cloud of the fruit tree and construct the voxel center point set of the outer contour. S3. Normal direction estimation is performed based on the center point set of the outer contour voxels, and the consistency of the normal direction is corrected by using the nearest neighbor graph and breadth-first search. S4. Generate candidate viewpoints based on the corrected normal direction and establish the coverage relationship between the candidate viewpoints and the target point cloud; S5. Based on the coverage relationship and marginal coverage gain of the candidate viewpoints to the target point cloud, and combined with spatial dispersion constraints, obtain the target viewpoint set. S6. Determine the access order based on the target viewpoint set, and perform straight-line connections when safety constraints are met, and execute local obstacle avoidance planning to generate obstacle bypass paths when safety constraints are not met. S7. Smooth the obtained obstacle avoidance path and perform collision checks and coverage integrity verification to obtain the data acquisition flight trajectory for fruit tree 3D reconstruction.
[0007] Preferably, step S2 includes: The outer contour extraction adopts a voxel adjacency-based point cloud filtering method. Through four steps, point cloud dilation, voxelization and filtering, six-neighborhood classification and surface point extraction, points in external voxels and isolated voxels are retained from the original point cloud to form an outer contour point cloud for characterizing the external boundary structure of fruit trees. Simultaneously, the voxel center points corresponding to the external voxels and isolated voxels are extracted to form the outer contour voxel center point set.
[0008] Preferably, step S3 includes: For any center point in the set of center points of the outer contour voxels qi A neighborhood point set is constructed within a predefined neighborhood, and a covariance matrix is constructed based on the neighborhood point set. Principal component analysis is performed on the covariance matrix, and the eigenvector corresponding to the smallest eigenvalue is taken as the center point. qi The initial normal; A nearest neighbor graph is constructed based on the center point set of the outer contour voxels. The graph is traversed and propagated starting from the seed point. When the dot product of the normals of adjacent points is less than zero, the normals of adjacent points are flipped to achieve consistency correction of the normal direction.
[0009] Preferably, step S4 includes: Candidate viewpoints are generated by translating along the corrected normal direction by a preset offset distance; Let the set of candidate viewpoints be V The target point cloud is P Each candidate viewpoint corresponds to a set of observable points, and the coverage relationship of the candidate viewpoint to the target point cloud is established through viewpoint cone geometric filtering, occlusion judgment, and observation quality evaluation.
[0010] Preferably, step S5 includes: Let the set of currently selected viewpoints be... St Then candidate viewpoints vi The marginal coverage gain is defined as the candidate viewpoint vi The set of viewpoints that can be covered but are not currently selected. St Covered target point cloud P Quantity; A near-optimal candidate set is constructed with a marginal coverage gain not lower than a preset ratio of the optimal marginal coverage gain, and a candidate viewpoint with the largest minimum distance to the currently selected viewpoint set is selected from the near-optimal candidate set to reduce local viewpoint clustering.
[0011] Preferably, step S6 includes: Let the optimized target viewpoint set be ,in, vn Represents the target viewpoint set The first in n One perspective; For the current candidate viewpoints vi The next visited viewpoint is defined as the candidate viewpoint in the set of unvisited viewpoints that is the next most visited viewpoint. vi The viewpoint with the shortest straight path between them; When the straight path between adjacent target viewpoints does not meet the safe distance constraint, the local obstacle avoidance planner is called to generate an obstacle avoidance path. The local obstacle avoidance planner uses the fast extended random tree star algorithm and selects a better path based on the path quality evaluation function, which includes a path length term and a trajectory smoothness cost term.
[0012] Preferably, step S7 includes: generating a continuous flight trajectory by using B-spline interpolation on the discrete path control point sequence, and outputting the data acquisition flight trajectory for fruit tree 3D reconstruction after collision checking and coverage integrity verification.
[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention removes redundant points inside fruit trees by extracting the outer contour based on voxel adjacency, retaining only the boundary structure relevant to external observation, thereby reducing the computational burden of subsequent viewpoint generation and path planning. Simultaneously, it estimates the normal direction based on the voxel center point set of the outer contour and corrects the normal direction consistency through nearest neighbor graphs and breadth-first search, effectively reducing the candidate viewpoint orientation error caused by normal ambiguity. Furthermore, by combining marginal coverage gain and spatial dispersion constraints into a viewpoint selection mechanism, it reduces local viewpoint clustering and duplicate observations while ensuring coverage efficiency. Finally, by employing a hierarchical path connection strategy combining straight-line connection, local obstacle avoidance planning, and trajectory smoothing, it can generate safe, continuous flight trajectories suitable for complex canopy environments that satisfy UAV motion constraints. Attached Figure Description
[0014] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0015] Figure 1 This is a flowchart illustrating the overall process of an embodiment of the present invention. Figure 2 This is a system structure block diagram according to an embodiment of the present invention; Figure 3 A sub-flowchart for candidate viewpoint generation and coverage relationship establishment in this embodiment of the invention. Detailed Implementation
[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0018] Example 1: This embodiment provides a UAV path planning method for 3D reconstruction of fruit trees based on outer contour guidance. (See attached diagram) Figure 1As shown, this method generally includes steps such as acquiring the original point cloud, extracting the outer contour, estimating the normal and correcting for consistency, generating candidate viewpoints and establishing coverage relationships, selecting target viewpoints, connecting paths and planning local obstacle avoidance, and smoothing and validating the trajectory. Figure 3 As shown, the method first performs point cloud dilation, voxelization, low-density voxel filtering, and six-neighbor classification on the original point cloud data to extract the outer contour point cloud and voxel center point set. Then, based on the voxel center point set, PCA normal estimation and consistency correction are performed, and candidate viewpoints are generated and coverage relationships are established. Specifically, the method includes the following steps: S1. Obtain the original point cloud data of the target fruit tree.
[0019] This embodiment selects target trees located on a university campus in a city in my country as the point cloud acquisition object, and uses a drone platform equipped with a Livox Mid-360 LiDAR to collect the raw point cloud data of the target fruit trees. During the acquisition process, the drone flies around the target fruit trees at different heights and directions according to a preset flight path. The LiDAR continuously scans the canopy, branches, and surrounding environment of the target fruit trees, and combines the drone's pose information to perform coordinate transformation and fusion on the acquired laser point cloud to generate the original three-dimensional point cloud map of the target area.
[0020] The original 3D point cloud data is imported into point cloud processing software. In this embodiment, CloudCompare software is used to perform coordinate transformation, point cloud registration, outlier noise point filtering, and target area clipping in sequence. Point cloud data of the ground, surrounding vegetation, and other non-target objects are removed to obtain the 3D point cloud data of the target single fruit tree. This provides basic data for subsequent fruit tree outline extraction, candidate viewpoint generation, viewpoint screening, and UAV flight path planning.
[0021] S2. Perform dilation, voxelization, point number filtering, and six-neighbor classification on the original point cloud data to extract the outer contour point cloud of the fruit tree and construct the voxel center point set of the outer contour.
[0022] The outer contour extraction adopts a point cloud filtering method based on voxel adjacency. Through four steps, point cloud dilation, voxelization and filtering, six-neighborhood classification and surface point extraction, points in external voxels and isolated voxels are retained from the original point cloud to form an outer contour point cloud for characterizing the external boundary structure of the fruit tree. At the same time, the voxel center points corresponding to the external voxels and isolated voxels are extracted to form the outer contour voxel center point set, which serves as the geometric basis for subsequent normal estimation and candidate viewpoint generation.
[0023] Let the original point cloud be Inflated point cloud is The set of voxels is V The minimum number of points threshold is 0. mThe effective set of voxels after voxel filtering is: , Each effective voxel is classified according to its six-neighborhood occupancy, and the original points corresponding to the external voxels and isolated voxels are retained to form an outer contour point cloud.
[0024] S3. Normal direction estimation is performed based on the center point set of the outer contour voxels, and the consistency of the normal direction is corrected by using the nearest neighbor graph and breadth-first search.
[0025] Let the set of center points of the outer contour voxels be: , Where M represents the total number of voxels.
[0026] For any center point qi Construct a set of neighborhood points within a preset neighborhood: , in, qj Represents the set of voxel centers of the outer contour. Q The first in j One central point, r This indicates the preset neighborhood search radius.
[0027] The covariance matrix is: , in, It represents the average value of all points within a local neighborhood.
[0028] Principal component analysis is performed on the covariance matrix, and the eigenvector corresponding to the smallest eigenvalue is taken as the center point. qi The initial normal direction.
[0029] To achieve consistency correction of normal direction, a nearest neighbor graph is constructed based on the center point set, and traversal propagation is performed starting from the seed point; when the dot product of the normals of adjacent points is less than zero, the normals of adjacent points are flipped, and the correction rule is as follows: , in, ni and nj These represent the normal vectors of adjacent center points.
[0030] S4. Generate candidate viewpoints based on the corrected normal direction and establish the coverage relationship between the candidate viewpoints and the target point cloud.
[0031] Candidate viewpoints are generated by translating along the corrected normal direction by a preset offset distance. For any center point... qi and its modified normal direction ni Corresponding candidate viewpoints vi Defined as: , in, d The preset offset distance is determined based on the safe distance between the UAV body, the effective imaging distance of the sensor, and the requirements for image distortion control.
[0032] Let the set of candidate viewpoints be The target point cloud is P Each candidate viewpoint corresponds to a set of observable points. vi The coverage relationship of candidate viewpoints to the target point cloud is established through viewpoint cone geometric filtering, occlusion judgment, and observation quality evaluation.
[0033] S5. Based on the coverage relationship of candidate viewpoints to the target point cloud and the marginal coverage gain, combined with spatial dispersion constraints, obtain the target viewpoint set.
[0034] Let the set of currently selected viewpoints be... St Then candidate viewpoints vi The marginal coverage gain is defined as: , Based on this, a near-optimal candidate set is constructed with a marginal coverage gain not lower than the preset ratio of the optimal marginal coverage gain. The candidate viewpoint with the largest minimum distance from the current selected viewpoint set is selected from the near-optimal candidate set to reduce local viewpoint clustering.
[0035] S6. Determine the access order based on the target viewpoint set, and perform straight-line connections when safety constraints are met, and execute local obstacle avoidance planning to generate obstacle bypass paths when safety constraints are not met.
[0036] Let the optimized target viewpoint set be ,in, vn Represents the target viewpoint set The first in n Each viewpoint. For the current viewpoint... vi The next access viewpoint is defined as: , in, This represents the set of unvisited viewpoints. Indicate viewpoint vi with viewpoint vj The straight line segment between them.
[0037] When the straight-line path between adjacent target viewpoints does not meet the safe distance constraint, the local obstacle avoidance planner is invoked to generate an obstacle avoidance path. Preferably, the local obstacle avoidance planner employs the Fast Extended Random Tree Star algorithm and selects a better path based on the following path quality evaluation function: , in, Indicates the path length item. This represents the cost term for trajectory smoothness. and These are the weighting coefficients.
[0038] S7. Smooth the obtained obstacle avoidance path and perform collision checks and coverage integrity verification to obtain the data acquisition flight trajectory for fruit tree 3D reconstruction.
[0039] Continuous flight trajectories are generated from discrete path control point sequences using B-spline interpolation, and the expression is as follows: , in, For B-spline basis functions, For discrete path control points, k Let be the degree of the B-spline basis function.
[0040] Finally, after collision checks and coverage integrity verification, the data acquisition flight trajectory for fruit tree 3D reconstruction is output.
[0041] Example 2: This invention also provides a UAV path planning system for 3D reconstruction of fruit trees based on outer contour guidance, including a data acquisition module, an outer contour extraction module, a normal estimation and correction module, a candidate viewpoint generation module, a viewpoint filtering module, a path planning module, and a trajectory smoothing and verification module, used to execute the above methods. Figure 2 As shown, the system takes the original point cloud data from the front end as input, and after data acquisition, outer contour extraction, normal estimation and correction, candidate viewpoint generation, viewpoint screening, path planning and trajectory smoothing and verification, it outputs the flight trajectory of the fruit tree 3D reconstruction data acquisition.
[0042] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.
Claims
1. A method for UAV path planning in 3D reconstruction of fruit trees based on outer contour guidance, characterized in that, include: S1. Obtain the raw point cloud data of the target fruit tree; S2. Perform dilation, voxelization, point number filtering, and six-neighbor classification on the original point cloud data to extract the outer contour point cloud of the fruit tree and construct the outer contour voxel center point set. The steps of S2 include: the outer contour extraction adopts a point cloud filtering method based on voxel adjacency. Through four steps of point cloud dilation, voxelization and filtering, six-neighbor classification, and surface point extraction, the points in the external voxels and isolated voxels are retained from the original point cloud to form the outer contour point cloud used to characterize the external boundary structure of the fruit tree; at the same time, the voxel center points corresponding to the external voxels and isolated voxels are extracted to form the outer contour voxel center point set. S3. Normal direction estimation is performed based on the center point set of the outer contour voxels, and the consistency of the normal direction is corrected by using the nearest neighbor graph and breadth-first search. S4. Generate candidate viewpoints based on the corrected normal direction and establish the coverage relationship between the candidate viewpoints and the target point cloud; S5. Based on the coverage relationship and marginal coverage gain of the candidate viewpoints to the target point cloud, and combined with spatial dispersion constraints, obtain the target viewpoint set; the steps of S5 include: assuming the currently selected viewpoint set is... St Then candidate viewpoints vi The marginal coverage gain is defined as the candidate viewpoint vi The set of viewpoints that can be covered but are not currently selected. St Covered target point cloud P The number of near-optimal candidate points; construct a near-optimal candidate set with a marginal coverage gain not lower than a preset ratio of the optimal marginal coverage gain, and select the candidate viewpoint with the largest minimum distance from the current selected viewpoint set from the near-optimal candidate set to reduce local viewpoint clustering; S6. Determine the access order based on the target viewpoint set, and perform straight-line connections when safety constraints are met, and execute local obstacle avoidance planning to generate obstacle bypass paths when safety constraints are not met. S7. Smooth the obtained obstacle avoidance path and perform collision checks and coverage integrity verification to obtain the data acquisition flight trajectory for fruit tree 3D reconstruction.
2. The UAV path planning method for three-dimensional reconstruction of fruit trees based on outer contour guidance according to claim 1, characterized in that, The steps in S3 include: For any center point in the set of center points of the outer contour voxels qi A neighborhood point set is constructed within a predefined neighborhood, and a covariance matrix is constructed based on the neighborhood point set. Principal component analysis is performed on the covariance matrix, and the eigenvector corresponding to the smallest eigenvalue is taken as the center point. qi The initial normal; A nearest neighbor graph is constructed based on the center point set of the outer contour voxels. The graph is traversed and propagated starting from the seed point. When the dot product of the normals of adjacent points is less than zero, the normals of adjacent points are flipped to achieve consistency correction of the normal direction.
3. The UAV path planning method for three-dimensional reconstruction of fruit trees based on outer contour guidance according to claim 1, characterized in that, The steps in S4 include: Candidate viewpoints are generated by translating along the corrected normal direction by a preset offset distance; Let the set of candidate viewpoints be V The target point cloud is P Each candidate viewpoint corresponds to a set of observable points, and the coverage relationship of the candidate viewpoint to the target point cloud is established through viewpoint cone geometric filtering, occlusion judgment, and observation quality evaluation.
4. The UAV path planning method for three-dimensional reconstruction of fruit trees based on outer contour guidance according to claim 1, characterized in that, Step S6 includes: Let the optimized target viewpoint set be ,in, vn Represents the target viewpoint set The first in n One perspective; For the current candidate viewpoints vi The next visited viewpoint is defined as the candidate viewpoint in the set of unvisited viewpoints that is the next most visited viewpoint. vi The viewpoint with the shortest straight path between them; When the straight path between adjacent target viewpoints does not meet the safe distance constraint, the local obstacle avoidance planner is called to generate an obstacle avoidance path. The local obstacle avoidance planner uses the fast extended random tree star algorithm and selects a better path based on the path quality evaluation function, which includes a path length term and a trajectory smoothness cost term.
5. The UAV path planning method for three-dimensional reconstruction of fruit trees based on outer contour guidance according to claim 1, characterized in that, The steps in S7 include: generating a continuous flight trajectory by using B-spline interpolation on the discrete path control point sequence, and outputting the data acquisition flight trajectory for fruit tree 3D reconstruction after collision checking and coverage integrity verification.
Citation Information
Patent Citations
Method and system for adjusting, installing and positioning special-shaped stone curtain wall based on unmanned aerial vehicle
CN120180576A
Fruit tree image acquisition method and system based on two-dimensional skeleton and viewpoint planning
CN122265609A
