A denoising method for 3D point cloud data of orchard environment integrating terrain features

Abstract

The present invention discloses a denoising method for three-dimensional point cloud data of an orchard environment that integrates terrain features. The present invention divides the point cloud into an upper and lower part by line segmentation, performs cloth filtering to fit the ground and the protective net respectively, and then calculates the elevation difference of the point cloud projected onto the two-dimensional grid plane, and filters out the ground point cloud and the protective net point cloud. Outlier noise is filtered out for other point clouds. For dynamic obstacles that are difficult to mark, a two-step marking method is performed, and the dynamic obstacles are marked using inter-frame information and region growth clustering method. In order to improve the credibility of the marking, a fusion feasibility model is determined. Finally, the dynamic obstacles are filtered out to obtain a clean point cloud of a row of fruit trees. This method effectively filters out the noise point cloud that is invalid for point cloud registration, extracts the point cloud of a row of fruit trees rich in feature information, has higher robustness, more accurate registration results, and the overall real-time performance of the method is high.

Description

Technical Field

[0001] The present invention belongs to the technical field of data processing, and in particular to a method for denoising three-dimensional point cloud data of an orchard environment by integrating terrain features. Background Art

[0002] In the field of agricultural robotics, automated production in large, standardized farms is well established. However, existing technologies are inadequate for outdoor, semi-structured orchard environments. Currently, mainstream navigation and positioning technologies rely on laser sensors. Three-dimensional laser point clouds, with their large volume and rich environmental information, are well-suited for point cloud registration. However, these point clouds also face challenges such as large data volumes, high levels of noise and interference, and difficulty processing. In outdoor, semi-structured orchard environments, noisy point clouds that cannot be used for matching and positioning generally include ground point clouds, dynamic obstacle point clouds, and optional protective net point clouds. Ground point clouds contain volatile weeds, bird-preventing protective nets can appear and disappear, and the motion of dynamic obstacles is difficult to estimate. The presence of noisy point clouds can affect subsequent point cloud registration, significantly reducing robot positioning accuracy. Therefore, developing a point cloud denoising method for orchard environments is crucial for improving robot positioning accuracy.

[0003] Chinese patent document CN110415259A discloses a point cloud processing method that processes two-dimensional laser point clouds based on laser reflection intensity. This method can filter out objects such as pedestrians, buildings, and roadways, and extract tree crowns and trunks. However, the point cloud intensity is easily affected by weather (rain affects reflection intensity) and the angle of incidence, resulting in unstable extraction results. Furthermore, the two-dimensional point cloud denoising results are difficult to use in subsequent point cloud registration and positioning processes because the two-dimensional point cloud contains too little information, making it difficult to use in engineering practice. Patent document CN116071599A discloses a mountain point cloud classification method that uses a neural network to classify mountain point clouds into residential areas, transportation, vegetation, and other areas. Although the neural network can classify point clouds, the biggest problem is that it requires training with a large dataset in advance. Changes in tree type and vegetation will have a significant impact on recognition accuracy. When faced with large-scale three-dimensional point cloud maps, the computational cost is too high, and it lacks real-time performance, making it difficult to use in engineering practice. Therefore, it is necessary to study a three-dimensional point cloud data denoising method for orchard environments that integrates terrain features with good real-time performance and high robustness. Summary of the Invention

[0004] The purpose of the present invention is to provide a denoising method for three-dimensional point cloud data of an orchard environment that integrates terrain features, so as to solve the problem that there are too many noisy point clouds and invalid point clouds in the original three-dimensional laser point cloud data, which affects the matching accuracy.

[0005] The technical solution for achieving the purpose of the present invention is as follows: On the one hand, a method for denoising three-dimensional point cloud data of an orchard environment by integrating terrain features is provided, the method comprising the following steps:

[0006] Step 1: LiDAR scans the orchard environment to collect 3D laser point cloud information, and then divides the laser point cloud into upper and lower halves according to the laser beam;

[0007] Step 2: Use cloth filtering to perform protection net fitting and ground fitting on the upper and lower halves of the laser point cloud respectively;

[0008] Step 3: Directly output the ground fitting of the lower half of the laser point cloud; determine the existence of the protection net fitting of the upper half of the point cloud. If it is determined that there is no protection net, no processing is performed on the upper half of the point cloud. Otherwise, output the protection net fitting of the upper half of the laser point cloud and proceed to the next step.

[0009] Step 4: Project the 3D laser point cloud onto a 2D grid plane, identify the point cloud based on its elevation difference, mark the laser point cloud with ground weeds and protective nets, and filter out the remaining unmarked point clouds. Record the remaining unmarked point clouds as "other point clouds" and restore them to a 3D point cloud.

[0010] Step 5: Filter outlier noise from other point clouds;

[0011] Step 6: Perform initial dynamic obstacle marking based on the inter-frame matching results;

[0012] Step 7: Project the 3D point cloud onto a 2D grid plane, build a point cloud kd-tree, select seed points based on elevation difference, and perform region growing clustering. For the clustering results, perform secondary marking of dynamic obstacles based on point cloud density and elevation difference.

[0013] Step 8: Based on the above two marking results, perform fusion credibility judgment, mark the dynamic obstacle point cloud and filter it out, and output the clean fruit tree row point cloud.

[0014] Furthermore, in step 2, the upper and lower halves of the point cloud are respectively fitted with a protective net and a ground surface by cloth filtering, specifically including:

[0015] Step 2-1, invert the lower half of the laser point cloud;

[0016] Then perform the following steps on the upper and lower halves of the laser point cloud:

[0017] Step 2-2, initialize the cloth grid and determine the number of cloth particles;

[0018] Step 2-3: Project all laser point clouds and cloth particles onto a horizontal plane. Find the laser point cloud closest to each cloth particle in the horizontal plane. The two form a point pair. Calculate the elevation difference D of each point pair. This elevation difference D is used to define the lowest position that each cloth particle can reach.

[0019] Step 2-4, calculate the displacement X(t+Δt) of each cloth particle under the influence of gravity:

[0020]

[0021] Where m is the mass of the particle, G is the acceleration due to gravity, Δt is the time step, X(t) and X(t-Δt) are the positions of the particle at time t and t-Δt, respectively;

[0022] Steps 2-5, calculate the displacement of each cloth particle under the influence of internal forces:

[0023]

[0024] in, represents the particle displacement vector, b is the particle movement constant, Represent the current position coordinate vectors of the particle and its neighboring particles, is a unit vector;

[0025] In step 2-6, the total displacement of each cloth particle under the influence of gravity and internal forces is calculated and compared with the corresponding elevation difference D. If the total displacement exceeds the elevation difference D, the iteration steps 2-4 and 2-5 are returned to until the preset maximum number of iterations is reached. Otherwise, the iteration is terminated to obtain the plane of the cloth network simulation.

[0026] Furthermore, in step 3, the existence of the protective net is determined by fitting the protective net of the upper half of the point cloud based on the elevation difference of the point cloud. If the elevation difference of the cloth particles in the protective net fitting is greater than the preset threshold H, it is determined that there is no protective net.

[0027] Furthermore, in step 4, the 3D laser point cloud is projected onto a 2D grid plane, and the point cloud is identified based on its elevation difference, with ground weed marks and protective net marks added to the laser point cloud. Specifically:

[0028] Step 4-1: Project the 3D laser point cloud onto a 2D grid plane, calculate the distance between each point cloud and the ground point cloud surface and the protective net point cloud surface output in step 3, and output the minimum distance corresponding to each point cloud by comparing the two distances;

[0029] Step 4-2: Determine whether the minimum distance of each point cloud is less than the preset threshold corresponding to the protection net point cloud surface or the ground point cloud surface. If so, mark the point cloud with a protection net mark or ground weed mark and filter it out; otherwise, do not mark it.

[0030] Furthermore, the outlier noise filtering of other point clouds described in step 5 specifically includes:

[0031] The number of point clouds within the specified radius of each point cloud is calculated. If the number of point clouds is less than the preset threshold, the point cloud is determined to be an outlier and filtered out.

[0032] Furthermore, in step 6, initial dynamic obstacle marking is performed based on the inter-frame matching results, specifically including:

[0033] Step 6-1, with the lidar sensor as the center, obtain the two-dimensional grid map F(t) of the current frame and the two-dimensional grid map M(t-1) of the previous frame;

[0034] Step 6-2, subtract the two maps in step 6-1 and calculate the non-overlapping area of ​​F(t) and M(t-1);

[0035] Step 6-3, filtering the point cloud in the non-overlapping area;

[0036] In step 6-4, Euclidean clustering is performed on the filtered point cloud to obtain dynamic obstacle point cloud clusters, and these cloud clusters are marked as A1.

[0037] Furthermore, in step 7, the 3D point cloud is projected onto a 2D grid plane, a point cloud kd-tree is established, seed points are selected based on elevation difference, and region growing clustering is performed. Based on the clustering results, dynamic obstacles are secondary marked according to point cloud density and elevation difference, specifically including:

[0038] Step 7-1: Project the 3D point cloud onto the 2D grid plane, calculate the elevation difference of the point cloud inside the grid, and select the points in the grid whose elevation difference is greater than the set threshold as seed points;

[0039] Step 7-2: For each seed point, establish a kd-tree to search for points within its Euclidean space sphere, calculate the normal vectors between the seed point and the points within the sphere, and classify the points whose normal vectors are parallel within a preset range into the same point cloud cluster;

[0040] In step 7-3, for the clustered point cloud clusters, the number of point clouds and the elevation difference of the point clouds are calculated. The point cloud clusters with the number of point clouds less than the threshold M and the elevation difference less than the threshold N are marked as dynamic obstacles again and marked as A2.

[0041] Furthermore, step 8 performs a fusion credibility determination based on the above two marking results, marks and filters out the dynamic obstacle point cloud, and outputs a clean fruit tree row point cloud, specifically including:

[0042] Step 8-1: Establish a fusion feasibility model:

[0043] A=α·A1+β·A2

[0044] Among them: α+β=1.

[0045] For each point cloud, a corresponding A value is calculated based on the fusion feasibility model;

[0046] Step 8-2: For each point cloud, determine whether its A value is greater than the set threshold. If so, it is determined to be a dynamic obstacle, marked and filtered out;

[0047] In step 8-3, the remaining point clouds are output as the clean point cloud of the fruit tree row.

[0048] On the other hand, a system for denoising three-dimensional point cloud data of an orchard environment by integrating terrain features is provided, the system comprising the following steps executed in sequence:

[0049] The first module is used to implement laser radar scanning of the orchard environment, collect three-dimensional laser point cloud information, and then divide the laser point cloud into upper and lower half point clouds according to the laser beam;

[0050] The second module is used to implement protection net fitting and ground fitting for the upper and lower halves of the laser point cloud respectively through cloth filtering;

[0051] The third module is used to directly output the ground fitting of the lower half of the laser point cloud; the protection net fitting of the upper half of the point cloud is used to determine the existence of the protection net. If it is determined that there is no protection net, no processing is performed on the upper half of the point cloud. Otherwise, the protection net fitting of the upper half of the laser point cloud is output and the next module is executed;

[0052] The fourth module is used to project the 3D laser point cloud onto a 2D grid plane, identify the point cloud based on its elevation difference, mark the laser point cloud with ground weeds and protective nets, filter out the marks, record other unmarked point clouds as other point clouds, and restore them to a 3D point cloud.

[0053] The fifth module is used to filter outlier noise from other point clouds.

[0054] The sixth module is used to perform initial dynamic obstacle marking based on the inter-frame matching results;

[0055] The seventh module is used to project the 3D point cloud onto a 2D grid plane, establish a point cloud kd-tree, select seed points based on elevation difference, and perform region growing clustering. Based on the clustering results, dynamic obstacles are secondary marked according to the point cloud density and elevation difference.

[0056] The eighth module is used to make a fusion credibility judgment based on the above two marking results, mark the dynamic obstacle point cloud and filter it out, and output a clean fruit tree row point cloud.

[0057] Compared with the prior art, the present invention has the following significant advantages:

[0058] (1) The present invention fits the ground and the protective net by flexible cloth. Compared with other plane fitting methods, in an environment with uneven terrain and large undulations such as an orchard, it can better characterize the shape of the orchard ground and the protective net, and the point cloud segmentation is more accurate.

[0059] (2) The present invention uses a two-step method to determine and fit dynamic obstacles. Compared with the commonly used dynamic obstacle determination method based on neural networks, it is faster and more reliable.

[0060] (3) The present invention filters out the ground point cloud, protection net point cloud, dynamic obstacle point cloud and fruit tree point cloud in the original point cloud, avoiding point cloud registration errors caused by such noisy and highly interfering point clouds, and provides a clean, reliable and feature-filled fruit tree point cloud for subsequent point cloud registration.

[0061] The present invention is further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 It is a schematic diagram of the ground point cloud, protection net point cloud, dynamic obstacle point cloud and fruit tree point cloud of the present invention.

[0063] Figure 2 It is a flow chart of the method for denoising three-dimensional point cloud data of an orchard environment by integrating terrain features according to the present invention.

[0064] Figure 3 It is a schematic diagram of the principle of flexible cloth fitting the ground and protective net.

[0065] Figure 4 The figure is an actual scene picture of an orchard collected in an embodiment.

[0066] Figure 5 FIG. 4 is an original point cloud image of an orchard in an embodiment.

[0067] Figure 6 It is a locally enlarged laser point cloud image of the ground point cloud in an embodiment.

[0068] Figure 7 It is a laser point cloud image of the protective net and dynamic obstacle point cloud in one embodiment.

[0069] Figure 8 The figure is a clean laser point cloud image of a row of fruit trees retained after denoising in an embodiment of the present invention. DETAILED DESCRIPTION

[0070] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0071] It should be noted that if the embodiments of the present invention involve directional indications (such as up, down, left, right, front, back, etc.), the directional indications are only used to explain the relative position relationship, movement status, etc. between the various components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indications will also change accordingly.

[0072] In addition, if there are descriptions involving "first", "second", etc. in the embodiments of the present invention, the descriptions of "first", "second", etc. are only for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features specified as "first" and "second" may explicitly or implicitly include at least one of such features. In addition, the technical solutions between the various embodiments can be combined with each other, but this must be based on the fact that ordinary technicians in this field can implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0073] The present invention uses a flexible cloth filter fitting method to fit the rugged ground in the wild and the possible protective nets, filters out the ground point cloud and the protective net point cloud, and then uses the inter-frame matching method and the region growing clustering method to perform a two-step dynamic obstacle judgment. By filtering out the dynamic obstacles, all the noise point clouds that affect the point cloud matching accuracy can be removed, realizing three-dimensional point cloud denoising of the orchard environment.

[0074] In one embodiment, combined Figure 1 and Figure 2 , provides a method for denoising three-dimensional point cloud data of an orchard environment by integrating terrain features, the method comprising the following steps:

[0075] Step 1: Use a 3D laser radar to scan the orchard environment and collect 3D laser point cloud information. The laser point cloud is then divided into upper and lower halves according to the laser beams. For example, for a 32-line laser radar, this is divided into upper 16 lines and lower 16 lines.

[0076] Step 2: Use cloth filtering to perform protection net fitting and ground fitting on the upper and lower halves of the laser point cloud respectively; Figure 3 , specifically:

[0077] Step 2-1, invert the lower half of the laser point cloud;

[0078] Then perform the following steps on the upper and lower halves of the laser point cloud:

[0079] Step 2-2, initialize the cloth grid and determine the number of cloth particles;

[0080] Step 2-3: Project all laser point clouds and cloth particles onto a horizontal plane. Find the laser point cloud closest to each cloth particle in the horizontal plane. The two form a point pair. Calculate the elevation difference D of each point pair. This elevation difference D is used to define the lowest position that each cloth particle can reach.

[0081] Step 2-4, calculate the displacement X(t+Δt) of each cloth particle under the influence of gravity:

[0082]

[0083] Where m is the mass of the particle, G is the acceleration due to gravity, Δt is the time step, X(t) and X(t-Δt) are the positions of the particle at time t and t-Δt, respectively;

[0084] Steps 2-5, calculate the displacement of each cloth particle under the influence of internal forces:

[0085]

[0086] in, represents the particle displacement vector, b is the particle movement constant, Represent the current position coordinate vectors of the particle and its neighboring particles, is a unit vector;

[0087] In step 2-6, the total displacement of each cloth particle under the influence of gravity and internal forces is calculated and compared with the corresponding elevation difference D. If the total displacement exceeds the elevation difference D, the iteration steps 2-4 and 2-5 are returned to until the preset maximum number of iterations is reached. Otherwise, the iteration is terminated to obtain the plane of the cloth network simulation.

[0088] Step 3: Directly output the ground fitting of the lower half of the laser point cloud; determine the existence of the protection net fitting of the upper half of the point cloud. If it is determined that there is no protection net, no processing is performed on the upper half of the point cloud. Otherwise, output the protection net fitting of the upper half of the laser point cloud and proceed to the next step.

[0089] Here, the existence of the protective net is determined based on the point cloud elevation difference. If the elevation difference of the cloth particles in the protective net fitting is greater than the preset threshold H, it is determined that there is no protective net.

[0090] Step 4: Project the 3D laser point cloud onto a 2D grid plane, identify the point cloud based on its elevation difference, mark the laser point cloud with ground weeds and protective nets, and filter out the remaining unmarked point clouds. Record the remaining unmarked point clouds as "other point clouds" and restore them to a 3D point cloud. Specifically:

[0091] Step 4-1: Project the 3D laser point cloud onto a 2D grid plane, calculate the distance between each point cloud and the ground point cloud surface and the protective net point cloud surface output in step 3, and output the minimum distance corresponding to each point cloud by comparing the two distances;

[0092] Step 4-2: Determine whether the minimum distance of each point cloud is less than the preset threshold corresponding to the protection net point cloud surface or the ground point cloud surface. If so, mark the point cloud with a protection net mark or ground weed mark and filter it out; otherwise, do not mark it.

[0093] Step 5: Filter outlier noise from other point clouds; specifically, the following steps are involved:

[0094] The number of point clouds within the specified radius of each point cloud is calculated. If the number of point clouds is less than the preset threshold, the point cloud is determined to be an outlier and filtered out.

[0095] Step 6: Perform initial dynamic obstacle marking based on the inter-frame matching results. This specifically includes:

[0096] Step 6-1, with the lidar sensor as the center, obtain the two-dimensional grid map F(t) of the current frame and the two-dimensional grid map M(t-1) of the previous frame;

[0097] Step 6-2, subtract the two maps in step 6-1 and calculate the non-overlapping area of ​​F(t) and M(t-1);

[0098] Step 6-3, filtering the point cloud in the non-overlapping area;

[0099] In step 6-4, Euclidean clustering is performed on the filtered point cloud (the appropriate clustering radius is set according to the obstacle type (dynamic obstacles in the orchard are generally people)) to obtain dynamic obstacle point cloud clusters, and these cloud clusters are marked as A1.

[0100] Step 7: Project the 3D point cloud onto a 2D grid plane, build a point cloud kd-tree, select seed points based on elevation difference, and perform region growing clustering. For the clustering results, perform secondary marking of dynamic obstacles based on point cloud density and elevation difference. This includes:

[0101] Step 7-1: Project the 3D point cloud onto the 2D grid plane, calculate the elevation difference of the point cloud inside the grid, and select the points in the grid whose elevation difference is greater than the set threshold as seed points;

[0102] Step 7-2: For each seed point, establish a kd-tree to search for points within its Euclidean space sphere, calculate the normal vectors between the seed point and the points within the sphere, and classify the points whose normal vectors are parallel within a preset range into the same point cloud cluster;

[0103] In step 7-3, for the clustered point cloud clusters, the number of point clouds and the elevation difference of the point clouds are calculated. The point cloud clusters with the number of point clouds less than the threshold M and the elevation difference less than the threshold N are marked as dynamic obstacles again and marked as A2.

[0104] Step 8: Based on the above two marking results, perform fusion credibility judgment, mark and filter out the dynamic obstacle point cloud, and output the clean fruit tree row point cloud; specifically, it includes:

[0105] Step 8-1: Establish a fusion feasibility model:

[0106] A=α·A1+β·A2

[0107] Among them: α+β=1.

[0108] For each point cloud, a corresponding A value is calculated based on the fusion feasibility model;

[0109] Step 8-2: For each point cloud, determine whether its A value is greater than the set threshold. If so, it is determined to be a dynamic obstacle, marked and filtered out;

[0110] In step 8-3, the remaining point clouds are output as the clean point cloud of the fruit tree row.

[0111] In one embodiment, the present invention provides a denoising system for three-dimensional point cloud data of an orchard environment integrating terrain features, characterized in that the system comprises the following steps executed in sequence:

[0112] The first module is used to implement laser radar scanning of the orchard environment, collect three-dimensional laser point cloud information, and then divide the laser point cloud into upper and lower half point clouds according to the laser beam;

[0113] The second module is used to implement protection net fitting and ground fitting for the upper and lower halves of the laser point cloud respectively through cloth filtering;

[0114] The third module is used to directly output the ground fitting of the lower half of the laser point cloud; the protection net fitting of the upper half of the point cloud is used to determine the existence of the protection net. If it is determined that there is no protection net, no processing is performed on the upper half of the point cloud. Otherwise, the protection net fitting of the upper half of the laser point cloud is output and the next module is executed;

[0115] The fourth module is used to project the 3D laser point cloud onto a 2D grid plane, identify the point cloud based on its elevation difference, mark the laser point cloud with ground weeds and protective nets, filter out the marks, record other unmarked point clouds as other point clouds, and restore them to a 3D point cloud.

[0116] The fifth module is used to filter outlier noise from other point clouds.

[0117] The sixth module is used to perform initial dynamic obstacle marking based on the inter-frame matching results;

[0118] The seventh module is used to project the 3D point cloud onto a 2D grid plane, establish a point cloud kd-tree, select seed points based on elevation difference, and perform region growing clustering. Based on the clustering results, dynamic obstacles are secondary marked according to the point cloud density and elevation difference.

[0119] The eighth module is used to make a fusion credibility judgment based on the above two marking results, mark the dynamic obstacle point cloud and filter it out, and output a clean fruit tree row point cloud.

[0120] Regarding the specific limitations of the orchard environment three-dimensional point cloud data denoising system that integrates terrain features, please refer to the limitations of the orchard environment three-dimensional point cloud data denoising method that integrates terrain features above, which will not be repeated here. Each module in the above-mentioned orchard environment three-dimensional point cloud data denoising system that integrates terrain features can be implemented in whole or in part through software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0121] As a specific example, the present invention is further verified in one of the embodiments.

[0122] In this embodiment, an Ouster 32-line laser radar was used to collect laser point cloud maps in an apple orchard for experimental verification. Figure 4 、 5 , 6, 7, and 8 for specific explanations. Figure 4 This is a collected real-life orchard scene picture. Figure 5 It is an unprocessed 3D laser point cloud map of the orchard, which contains a large number of noisy point clouds (protective nets and ground), making it difficult to extract valid point clouds. Figure 6 This is the effect after the method of the present invention removes the ground point cloud and the protection network point cloud. Figure 7 It is a point cloud map of pedestrian trajectories left when dynamic obstacles (pedestrians) walk in the orchard. Figure 8This is the clean point cloud of the fruit tree row left after the denoising method of the present invention removes the dynamic obstacle point cloud. It can be seen that the method of the present invention can effectively remove various useless noise point clouds in the orchard and retain the required clean point cloud of the fruit tree row.

[0123] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only illustrative of the principles of the present invention. Without departing from the spirit and scope of the present invention, any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included in the scope of protection of the present invention.

Claims

1. A method for denoising three-dimensional point cloud data of an orchard environment by integrating terrain features, characterized in that: The method comprises the following steps: Step 1: LiDAR scans the orchard environment to collect 3D laser point cloud information, and then divides the laser point cloud into upper and lower halves according to the laser beam; Step 2: Use cloth filtering to perform protection net fitting and ground fitting on the upper and lower halves of the laser point cloud respectively; Step 3: Directly output the ground fitting of the lower half of the laser point cloud; determine the existence of the protection net fitting of the upper half of the point cloud. If it is determined that there is no protection net, no processing is performed on the upper half of the point cloud. Otherwise, output the protection net fitting of the upper half of the laser point cloud and proceed to the next step. Step 4: Project the 3D laser point cloud onto a 2D grid plane, identify the point cloud based on its elevation difference, mark the laser point cloud with ground weeds and protective nets, and filter out the remaining unmarked point clouds. Record the remaining unmarked point clouds as "other point clouds" and restore them to a 3D point cloud. Step 5: Filter outlier noise from other point clouds; Step 6: Perform initial dynamic obstacle marking based on the inter-frame matching results; Step 7: Project the 3D point cloud onto a 2D grid plane, build a point cloud kd-tree, select seed points based on elevation difference, and perform region growing clustering. For the clustering results, perform secondary marking of dynamic obstacles based on point cloud density and elevation difference. Step 8: Based on the above two marking results, the fusion credibility is determined, the dynamic obstacle point cloud is marked and filtered out, and the clean fruit tree row point cloud is output; In step 6, the initial dynamic obstacle marking is performed based on the inter-frame matching results, specifically including: Step 6-1, with the lidar sensor as the center, obtain the two-dimensional grid map F(t) of the current frame and the two-dimensional grid map M(t-1) of the previous frame; Step 6-2, subtract the two maps in step 6-1 and calculate the non-overlapping area of ​​F(t) and M(t-1); Step 6-3, filtering the point cloud in the non-overlapping area; Step 6-4, perform Euclidean clustering on the filtered point cloud to obtain dynamic obstacle point cloud clusters, and mark these cloud clusters as A1; In step 7, the 3D point cloud is projected onto a 2D grid plane, a point cloud kd-tree is constructed, seed points are selected based on elevation difference, and region growing clustering is performed. Based on the clustering results, dynamic obstacles are secondary marked according to point cloud density and elevation difference, specifically including: Step 7-1: Project the 3D point cloud onto the 2D grid plane, calculate the elevation difference of the point cloud inside the grid, and select the points in the grid whose elevation difference is greater than the set threshold as seed points; Step 7-2: For each seed point, establish a kd-tree to search for points within its Euclidean space sphere, calculate the normal vectors between the seed point and the points within the sphere, and classify the points whose normal vectors are parallel within a preset range into the same point cloud cluster; In step 7-3, for the clustered point cloud clusters, the number of point clouds and the elevation difference of the point clouds are calculated. The point cloud clusters with the number of point clouds less than the threshold M and the elevation difference less than the threshold N are marked as dynamic obstacles again and marked as A2.

2. The method for denoising orchard environment three-dimensional point cloud data by integrating terrain features according to claim 1, characterized in that: In step 2, the upper and lower halves of the point cloud are fitted with a protective net and a ground plane respectively through cloth filtering, specifically including: Step 2-1, invert the lower half of the laser point cloud; Then perform the following steps on the upper and lower halves of the laser point cloud: Step 2-2, initialize the cloth grid and determine the number of cloth particles; Step 2-3: Project all laser point clouds and cloth particles onto a horizontal plane. Find the laser point cloud closest to each cloth particle in the horizontal plane. The two form a point pair. Calculate the elevation difference D of each point pair. This elevation difference D is used to define the lowest position that each cloth particle can reach. Step 2-4, calculate the displacement X(t+Δt) of each cloth particle under the influence of gravity: Where m is the mass of the particle, G is the acceleration due to gravity, Δt is the time step, X(t) and X(t-Δt) are the positions of the particle at time t and t-Δt, respectively; Steps 2-5, calculate the displacement of each cloth particle under the influence of internal forces: in, represents the particle displacement vector, b is the particle movement constant, Represent the current position coordinate vectors of the particle and its neighboring particles, is a unit vector; In step 2-6, the total displacement of each cloth particle under the influence of gravity and internal forces is calculated and compared with the corresponding elevation difference D. If the total displacement exceeds the elevation difference D, the iteration steps 2-4 and 2-5 are returned to until the preset maximum number of iterations is reached. Otherwise, the iteration is terminated to obtain the plane of the cloth network simulation.

3. The method for denoising orchard environment three-dimensional point cloud data by integrating terrain features according to claim 1 or 2, characterized in that: In step 3, the existence of the protective net is determined by fitting the protective net of the upper half of the point cloud based on the elevation difference of the point cloud. If the elevation difference of the cloth particles in the protective net fitting is greater than the preset threshold H, it is determined that there is no protective net.

4. The method for denoising orchard environment three-dimensional point cloud data by integrating terrain features according to claim 3, characterized in that: In step 4, the 3D laser point cloud is projected onto a 2D grid plane, and the point cloud is identified based on its elevation difference, with ground weed marks and protective net marks added to the laser point cloud. Specifically: Step 4-1: Project the 3D laser point cloud onto a 2D grid plane, calculate the distance between each point cloud and the ground point cloud surface and the protective net point cloud surface output in step 3, and output the minimum distance corresponding to each point cloud by comparing the two distances; Step 4-2: Determine whether the minimum distance of each point cloud is less than the preset threshold corresponding to the protection net point cloud surface or the ground point cloud surface. If so, mark the point cloud with a protection net mark or ground weed mark and filter it out; otherwise, do not mark it.

5. The method for denoising orchard environment three-dimensional point cloud data by integrating terrain features according to claim 4, characterized in that: Step 5 is to filter out outlier noise from other point clouds, specifically including: The number of point clouds within the specified radius of each point cloud is calculated. If the number of point clouds is less than the preset threshold, the point cloud is determined to be an outlier and filtered out.

6. The method for denoising orchard environment three-dimensional point cloud data by integrating terrain features according to claim 1, characterized in that: In step 8, based on the above two marking results, the fusion credibility is determined, the dynamic obstacle point cloud is marked and filtered out, and the clean fruit tree row point cloud is output, which specifically includes: Step 8-1: Establish a fusion feasibility model: A=α·A1+β·A2 Where: α + β = 1; For each point cloud, a corresponding A value is calculated based on the fusion feasibility model; Step 8-2: For each point cloud, determine whether its A value is greater than the set threshold. If so, it is determined to be a dynamic obstacle, marked and filtered out; In step 8-3, the remaining point clouds are output as the clean point cloud of the fruit tree row.

7. A 3D point cloud data denoising system for an orchard environment integrating terrain features based on the method according to any one of claims 1 to 6, characterized in that: The system includes the following steps: The first module is used to implement laser radar scanning of the orchard environment, collect three-dimensional laser point cloud information, and then divide the laser point cloud into upper and lower half point clouds according to the laser beam; The second module is used to implement protection net fitting and ground fitting for the upper and lower halves of the laser point cloud respectively through cloth filtering; The third module is used to directly output the ground fitting of the lower half of the laser point cloud; the protection net fitting of the upper half of the point cloud is used to determine the existence of the protection net. If it is determined that there is no protection net, no processing is performed on the upper half of the point cloud. Otherwise, the protection net fitting of the upper half of the laser point cloud is output and the next module is executed; The fourth module is used to project the 3D laser point cloud onto a 2D grid plane, identify the point cloud based on its elevation difference, mark the laser point cloud with ground weeds and protective nets, filter out the marks, record other unmarked point clouds as other point clouds, and restore them to a 3D point cloud. The fifth module is used to filter outlier noise from other point clouds; The sixth module is used to perform initial dynamic obstacle marking based on the inter-frame matching results; The seventh module is used to project the 3D point cloud onto a 2D grid plane, establish a point cloud kd-tree, select seed points based on elevation difference, and perform region growing clustering. Based on the clustering results, dynamic obstacles are secondary marked according to the point cloud density and elevation difference. The eighth module is used to make a fusion credibility judgment based on the above two marking results, mark the dynamic obstacle point cloud and filter it out, and output a clean fruit tree row point cloud.

Images