A Method and System for 3D Point Cloud Modeling Based on Multi-View Images

Abstract

This invention discloses a method and system for 3D point cloud modeling based on multi-view images, relating to the field of agricultural intelligent sensing technology. Specifically, it includes the following steps: acquiring multi-view images of a farmland environment using a depth camera; analyzing the overlapping regions and boundary characteristics of the point cloud data; identifying plants and segmenting their phenotypic structures; adjusting geometric correction parameters; performing growth dynamic grading and yield prediction; and combining crop phenotypic information and growth dynamic grading to obtain a crop yield potential distribution dataset. In this invention, multi-view images are acquired using a depth camera, analyzing the spatial distribution characteristics of crop targets and the overlapping regions and boundary characteristics of point cloud data; optimizing geometric correction parameters; improving point cloud alignment accuracy; ensuring accuracy during the stitching process; and enhancing the quality of point cloud stitching results. In complex farmland environments, this ensures accurate extraction of crop targets and phenotypic structure segmentation, accurately assesses crop growth dynamics, and improves the level of intelligent agricultural production management.

Description

Technical Field

[0001] This invention relates to the field of agricultural intelligent sensing technology, and in particular to a three-dimensional point cloud modeling method and system based on multi-view images. Background Technology

[0002] Agricultural intelligent sensing technology is an interdisciplinary field between agricultural engineering and information science. It primarily studies how to utilize technologies such as sensors, machine vision, remote sensing imagery, deep learning, and the Internet of Things to achieve intelligent monitoring and analysis of the agricultural production environment and crop growth status. Core aspects of this field include automatic sensing of farmland environmental elements, acquisition of phenotypic information of individual crops and crop populations, dynamic assessment of crop growth, identification of pests and diseases, yield prediction, and monitoring of spatiotemporal changes in agricultural ecosystems. By integrating multi-source sensor data, such as RGB images, depth images, infrared images, and point cloud data, agricultural intelligent sensing can accurately extract crop physiological characteristic parameters, providing data support for precision agriculture management. Systematic research in this field covers the entire technology chain from data acquisition and feature extraction to model construction and parameter inversion, and is a crucial foundation for promoting the digitalization and intelligentization of agriculture.

[0003] Traditional 3D point cloud modeling methods based on multi-view images utilize image information from a single or limited number of viewpoints to generate a 3D structural model of the crop target through steps such as feature matching, spatial reconstruction, and 3D coordinate calculation. These methods rely on structured light or stereo vision systems to acquire depth information, and then achieve 3D reconstruction of the target surface through geometric correction, disparity calculation, and point cloud stitching. Under multi-view conditions, traditional methods require acquiring images from different angles separately, then extracting and registering feature points, calculating the spatial relationship between viewpoints through extrinsic parameter calibration, and finally merging multiple sets of point cloud data to form a complete 3D model. This process demands high levels of illumination consistency, image overlap, and registration accuracy, employing feature matching algorithms and coordinate transformation matrix solving to complete point cloud alignment and fusion.

[0004] Existing technologies rely on images from a limited number of viewpoints for 3D point cloud modeling. This makes it difficult to ensure high-precision feature matching and spatial reconstruction under varying lighting conditions and insufficient image overlap. The registration and point cloud stitching processes are easily affected by viewpoint differences, inconsistencies in overlapping areas, and the accuracy of geometric correction, leading to reduced accuracy and completeness of the final 3D model. This is especially true in complex farmland environments, where changes in lighting and viewpoint differences can cause image feature extraction failures, thus impacting the entire modeling process. Furthermore, existing methods have high requirements for point cloud stitching accuracy and cannot effectively handle spatial deviations between different viewpoints, resulting in unsatisfactory point cloud alignment and fusion effects, affecting the final identification and analysis of crop targets. Summary of the Invention

[0005] To address the technical problems existing in the prior art, embodiments of the present invention provide a method and system for three-dimensional point cloud modeling based on multi-view images, the specific technical solution of which is as follows:

[0006] On the one hand, a method for 3D point cloud modeling based on multi-view images is provided, including the following steps:

[0007] S1: Based on the acquisition of multi-view images of farmland environment by depth camera, extract the spatial distribution characteristics of crop targets, analyze the overlapping area and boundary characteristics of point cloud data under different perspectives, identify individual plants and segment their phenotypic structure to obtain plant segmentation distribution characteristics.

[0008] S2: Based on the plant segmentation distribution characteristics, determine the spatial distribution density of point cloud data, compare the boundary characteristics and overlapping area relationship of point cloud data from each viewpoint, adjust the geometric correction parameters for areas with inconsistent boundary characteristics, and obtain the point cloud correction optimization result.

[0009] S3: Based on the point cloud correction and optimization results, match them item by item with the point cloud stitching process, analyze the spatial deviation of the incompletely aligned areas, gradually optimize the stitching accuracy, record the point cloud alignment status, and obtain the point cloud stitching distribution feature dataset.

[0010] S4: Based on the point cloud splicing distribution feature dataset, combined with farmland environmental elements and crop growth status information, growth dynamic grading and yield prediction are performed to obtain a three-dimensional point cloud growth evaluation model for crops.

[0011] As a further embodiment of the present invention, the plant segmentation distribution features include plant number, segmentation range, and phenotypic category; the point cloud correction and optimization results include correction direction, optimization identifier, and distribution area classification; the point cloud splicing distribution feature dataset includes splicing category, distribution interval, and alignment level; and the crop three-dimensional point cloud growth evaluation model includes evaluation labels and corresponding grading interval parameters.

[0012] As a further aspect of the present invention, the step of obtaining the plant segmentation distribution characteristics specifically includes:

[0013] S101: Based on the acquisition of multi-view images of farmland environment by depth camera, extract the spatial distribution characteristics of crop targets, and by analyzing the overlapping area and boundary characteristics of point cloud data under different perspectives, construct the pairing information of point cloud distribution and boundary characteristics from one perspective to obtain point cloud boundary characteristic group.

[0014] S102: Based on the point cloud boundary characteristic group, determine the relationship between the boundary characteristics and overlapping areas of the viewpoint point cloud data, classify the areas with deviated boundary characteristics as abnormal areas, classify the areas with consistent boundary characteristics as stable areas, and classify the remaining areas as regular types to obtain point cloud response distribution labels.

[0015] S103: Based on the point cloud response distribution labels, summarize the corresponding numbers and statuses of each type of region, and gather the associated attributes of abnormal type regions through sorting and grouping to obtain plant segmentation distribution features.

[0016] As a further aspect of the present invention, the step of obtaining the point cloud correction and optimization result specifically includes:

[0017] S201: Based on the plant segmentation distribution characteristics, compare the current spatial distribution density of the multi-view point cloud data, determine the distribution difference of the point cloud data in space, and match the regions where the distribution trend deviates with the region identifiers to obtain the distribution difference characteristics.

[0018] S202: Based on the distribution difference characteristics, compare the distribution position of each region's point cloud data in the spatial set, determine whether it is located in the boundary region of the distribution sequence, and identify the boundary regions as difference response types to obtain boundary response category groups;

[0019] S203: Based on the boundary response category group, optimize the geometric correction parameter configuration of the corresponding region for each response type region. By recording the optimization direction and parameter correction content, integrate the adjustment results of the region to obtain the point cloud correction optimization result.

[0020] As a further aspect of the present invention, the steps for obtaining the point cloud stitching distribution feature dataset are as follows:

[0021] S301: Based on the point cloud correction and optimization results, compare the distribution status of the viewpoint point cloud data with the corresponding stitching configuration, and locate the set of regions that need to be optimized by filtering out the incompletely aligned regions, and establish an optimization identification group;

[0022] S302: Based on the optimized identifier group, the splicing accuracy is adjusted step by step for the region. After each optimization adjustment, the point cloud alignment status and corresponding spatial deviation are recorded synchronously. The corresponding optimization status data is constructed according to the time series structure to obtain the optimization change trend set.

[0023] S303: Based on the optimized change trend set, analyze the alignment state changes of the region before and after continuous optimization, and perform linkage filtering according to the direction of spatial deviation change. Combine the region identifiers that meet the filtering conditions with the alignment state to obtain the point cloud stitching distribution feature dataset.

[0024] As a further aspect of the present invention, the steps for obtaining the crop three-dimensional point cloud growth evaluation model are as follows:

[0025] S401: Based on the point cloud splicing distribution feature dataset, the point cloud correction and optimization results are numbered and paired with the crop phenotypic information extracted from the viewpoint point cloud data, farmland environmental elements and crop growth status information are sorted out, and multi-information combination data of each viewpoint point cloud data is constructed to obtain an environmental information set.

[0026] S402: Based on the environmental information set, compare the distribution of information in point cloud data, determine the participation ratio of farmland environmental elements and crop growth status information in the combined data, and establish a weight allocation structure by statistically configuring the weight of difference information.

[0027] S403: Based on the weight allocation structure, normalization processing is performed on environmental information and crop phenotypic information, the influence of information is superimposed according to the weight allocation principle, and the effect of the weight allocation result on the point cloud data is judged to obtain the crop three-dimensional point cloud growth evaluation model.

[0028] As a further embodiment of the present invention, the multi-information combination data refers to point cloud data from each viewpoint as a basic unit, with the point cloud correction and optimization results corresponding to the viewpoint and the extracted crop phenotypic information numbered and paired, while the farmland environmental elements and crop growth status information corresponding to the viewpoint are associated and organized, so that the point cloud, phenotypic and environmental and growth information are combined together, and the multi-source information under the same viewpoint is structured and summarized to form multi-information combination data of viewpoint point cloud data.

[0029] The weight allocation result refers to the normalization of environmental information and crop phenotypic information under the guidance of a predetermined weight allocation structure, eliminating the influence of differential dimensions and numerical scales, and then superimposing and calculating each type of information after normalization according to the weight allocation principle to obtain the weight allocation result.

[0030] As a further aspect of the present invention, the method further includes step S5:

[0031] S5: Based on the crop three-dimensional point cloud growth evaluation model, the crop phenotypic information is merged with the weight configuration ratio, and the merged result is used to make interval judgment with the upper and lower limits of growth dynamic classification. The crop yield potential distribution dataset is obtained by classifying and organizing the data according to the interval judgment category.

[0032] The crop yield potential distribution dataset includes prediction labels, decision intervals, and classification indexes.

[0033] As a further aspect of the present invention, the steps for obtaining the crop yield potential distribution dataset are as follows:

[0034] S501: Based on the crop three-dimensional point cloud growth evaluation model, calculate the ratio of crop phenotypic information and corresponding weight configuration in the viewpoint point cloud data, analyze the distribution of the ratio merged data in the point cloud data, and determine the interval position of each point cloud data according to the upper and lower boundaries of the growth dynamic grading to obtain the signal grading interval category.

[0035] S502: Based on the signal classification interval categories, classify and organize according to the interval judgment results, record the point cloud data number and signal merging result corresponding to each category, and output the distribution of each category through structured data to obtain the crop yield potential distribution dataset.

[0036] On the other hand, a 3D point cloud modeling system based on multi-view images is provided, including:

[0037] The point cloud distribution recognition module is based on multi-view images of farmland environment collected by depth camera, extracts the spatial distribution features of crops, and combines the overlapping areas and boundary characteristics of point cloud data to segment plants, identify abnormal areas, and obtain plant segmentation distribution features.

[0038] The geometric correction optimization module evaluates the spatial distribution density of the current point cloud data based on the plant segmentation distribution characteristics, compares it with the original distribution sequence, and adjusts the geometric correction parameters for areas with high or low distribution density to obtain the point cloud correction optimization results.

[0039] The stitching accuracy dynamic filtering module compares the distribution status of the point cloud data item by item based on the point cloud correction and optimization results, optimizes the stitching accuracy, and filters the alignment areas that meet the conditions to obtain the point cloud stitching distribution feature dataset.

[0040] The multidimensional information normalization module, based on the point cloud splicing distribution feature dataset, combined with the point cloud splicing distribution feature dataset, correction and optimization results, and farmland environment and crop growth information, performs weighted normalization to obtain a crop three-dimensional point cloud growth evaluation model.

[0041] The yield prediction and discrimination module, based on the crop 3D point cloud growth evaluation model, proportionally merges crop phenotypic information with weight configuration, makes interval judgments with the upper and lower limits of growth dynamic grading based on the merged results, classifies and organizes the data according to the categories of the interval judgments, and outputs the corresponding categories to obtain the crop yield potential distribution dataset.

[0042] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0043] This invention overcomes the problem of low point cloud reconstruction accuracy caused by inconsistent viewpoints and lighting variations in traditional methods by acquiring multi-view images using a depth camera and analyzing the spatial distribution characteristics of crop targets and the overlapping areas and boundary features of point cloud data under different viewpoints. By adjusting geometric correction parameters, the alignment accuracy of the point cloud data is optimized, effectively reducing spatial deviations and ensuring accuracy during the stitching process, further improving the quality of the point cloud stitching results. Especially in complex farmland environments, it ensures more accurate crop target extraction and phenotypic structure segmentation. Simultaneously, by combining farmland environmental factors and crop growth status information, the invention accurately assesses crop growth dynamics, optimizes the accuracy of yield prediction, enhances the visualization and analysis capabilities of the crop growth process, and improves the level of intelligent agricultural production management. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0045] Figure 1 This is a schematic diagram of the steps of the present invention;

[0046] Figure 2 This is a detailed schematic diagram of S1 of the present invention;

[0047] Figure 3 This is a detailed schematic diagram of S2 of the present invention;

[0048] Figure 4 This is a detailed schematic diagram of S3 of the present invention;

[0049] Figure 5 This is a detailed schematic diagram of S4 of the present invention;

[0050] Figure 6 This is a detailed schematic diagram of S5 of the present invention;

[0051] Figure 7 This is a system module diagram of the present invention. Detailed Implementation

[0052] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0053] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0054] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0055] In this embodiment of the invention, sometimes a subscript such as W1 may be written in a non-subscript form such as W1. When the difference is not emphasized, the meaning they express is the same.

[0056] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0057] Please see Figure 1 This invention provides a method for modeling 3D point clouds based on multi-view images, comprising the following steps:

[0058] S1: Based on the acquisition of multi-view images of farmland environment by depth camera, extract the spatial distribution characteristics of crop targets, analyze the overlapping area and boundary characteristics of point cloud data under different perspectives, identify individual plants and segment their phenotypic structure to obtain plant segmentation distribution characteristics.

[0059] S2: Based on the plant segmentation distribution characteristics, determine the spatial distribution density of point cloud data, compare the boundary characteristics and overlapping area relationship of point cloud data from each viewpoint, adjust the geometric correction parameters for areas with inconsistent boundary characteristics, and obtain the point cloud correction optimization results.

[0060] S3: Based on the point cloud correction and optimization results, match them item by item with the point cloud stitching process, analyze the spatial deviation of the incompletely aligned areas, gradually optimize the stitching accuracy, record the point cloud alignment status, and obtain the point cloud stitching distribution feature dataset.

[0061] S4: Based on the point cloud splicing distribution feature dataset, combined with farmland environmental elements and crop growth status information, growth dynamic grading and yield prediction are carried out to obtain a three-dimensional point cloud growth evaluation model for crops.

[0062] S5: Based on the crop 3D point cloud growth assessment model, crop phenotypic information is merged with the weight configuration ratio. The merged result is used to make interval judgments with the upper and lower limits of growth dynamic grading. The data is then sorted and organized according to the interval judgment categories to obtain the crop yield potential distribution dataset.

[0063] The plant segmentation distribution features include plant number, segmentation range, and phenotypic category; the point cloud correction and optimization results include correction direction, optimization identifier, and distribution area classification; the point cloud splicing distribution feature dataset includes splicing category, distribution interval, and alignment level; the crop 3D point cloud growth assessment model includes assessment labels and corresponding grading interval parameters; and the crop yield potential distribution dataset includes prediction labels, judgment intervals, and classification index.

[0064] Please see Figure 2 The specific steps for obtaining the plant segmentation and distribution characteristics are as follows:

[0065] S101: Based on the acquisition of multi-view images of farmland environment by depth camera, extract the spatial distribution characteristics of crop targets, and by analyzing the overlapping area and boundary characteristics of point cloud data under different perspectives, construct the pairing information of point cloud distribution and boundary characteristics from one perspective to obtain point cloud boundary characteristic group.

[0066] Based on multi-view images collected in farmland environments using depth cameras, specifically, a drone equipped with a LiDAR and a visible light camera was used. The drone was set to fly at an altitude of 15 meters above the top of the crop canopy and a speed of 2 m / s. Point cloud data of corn plants were collected from three different azimuth angles: directly above, to the left front, and to the right front. Point clouds P1 (view 1), P2 (view 2), and P3 (view 3) were obtained, with 205,000, 215,000, and 210,000 points respectively. Spatial distribution characteristics of the crop targets were extracted. Subsequently, the overlapping areas and boundary characteristics of the point cloud data under different perspectives were analyzed. Specifically, P1 and P2 were spatially registered, and 5 cm × 5 cm × 5 cm three-dimensional grid cells were set. The number of points in each cell was counted, and any grid cell with at least 3 points was defined as the overlapping area R12. For the overlapping region R12, the boundary point set B1 under the view of P1 and the boundary point set B2 under the view of P2 are identified. The boundary point discrimination criterion is that if the number of points in a 2 cm radius neighborhood of a point is less than 5, it is determined to be a boundary point. Then, the Euclidean distance from each point in B1 to the nearest neighbor in B2 is calculated to form the distance set D12. The mean μ12 of D12 is calculated to be 1.8 cm and the standard deviation σ12 is 0.6 cm. Then, the pairing information of point cloud distribution and boundary characteristics is constructed for each view. The view number, overlapping region identifier, boundary point set and boundary statistical characteristics are combined into a pairing information and recorded as "(view 1, R12, B1, 1.8, 0.6)". The above operation is repeated for all view combinations (P1 and P2, P1 and P3, P2 and P3) to obtain the point cloud boundary characteristic group.

[0067] S102: Based on the point cloud boundary characteristic group, determine the relationship between the boundary characteristics and overlapping areas of the view point cloud data, classify the areas with deviated boundary characteristics as abnormal areas, classify the areas with consistent boundary characteristics as stable areas, and classify the remaining areas as regular types to obtain point cloud response distribution labels.

[0068] Based on the point cloud boundary characteristic group, which contains pairwise pairing information generated from viewpoints 1, 2, and 3, such as "(viewpoint 1, R12, B1, 1.8, 0.6)" and "(viewpoint 2, R12, B2, 1.8, 0.6)", the relationship between the boundary characteristics of the viewpoint point cloud data and the overlapping area is determined. The operation involves setting a global boundary difference benchmark, which is obtained by calculating the mean boundary distance μ and standard deviation σ from all paired information. The mean of all μ is μ_global = 1.9 cm, and the mean of all σ is σ_global = 0.7 cm. Areas with deviated boundary characteristics are classified as abnormal areas. The criterion for deviation is that if the absolute value of the difference between the mean boundary distance μx of an overlapping area Rxy at viewpoint x and the global mean μ_global is greater than twice the global standard deviation σ_global (i.e., greater than 1.4 cm), then that area Rxy is marked as abnormal at viewpoint x. In the overlapping area R13 of P1 and P3, P3 viewpoint... The mean of the boundary point set under the corner is 3.5 cm. The absolute value of the difference between it and the global mean is 1.6 cm, which is greater than 1.4 cm. Therefore, the part of region R13 under the P3 viewpoint is marked as an abnormal region, numbered A01. Regions with consistent boundary characteristics are classified as stable regions. The consistency criterion is that if the absolute value of the difference between the mean boundary distances μx and μy of an overlapping region Rxy under viewpoints x and y is less than 0.5 times the global standard deviation σ_global (i.e., less than 0.35 cm), then region Rxy is marked as a stable region. In region R12, μ1=1.8 cm and μ2=1.85 cm. The absolute value of their difference is 0.05 cm, which is less than 0.35 cm. Therefore, region R12 is marked as a stable region, numbered S01. The remaining regions are classified as normal types. The difference in the mean boundary distance of region R23 is between 0.35 cm and 1.4 cm. Therefore, it is marked as a normal region, numbered C01. The point cloud response distribution labels are obtained.

[0069] S103: Based on the point cloud response distribution labels, summarize the corresponding numbers and statuses of each type of region, and gather the associated attributes of abnormal type regions through sorting and grouping to obtain the plant segmentation distribution characteristics;

[0070] Based on the point cloud response distribution labels, which contain the classification results of each region, such as "(R13, Anomalous, A01)", "(R12, Stable, S01)", and "(R23, Normal, C01)", the corresponding numbers and states of each type of region are summarized to form a list "[A01: Anomalous, S01: Stable, C01: Normal]". Through sorting and grouping, first, regions are grouped according to their type (Anomalous, Stable, Normal). Then, regions of the Anomalous type are sorted based on the difference between the region's boundary distance mean and the global mean; regions with larger differences are sorted first. The difference in region A01 is 1.6 cm, and the difference in another abnormal region A02 is 1.5 cm. Therefore, A01 is ranked before A02. The associated attributes of the abnormal regions are aggregated. The operation is to extract the ranked list of abnormal regions "(A01, A02)" and associate each abnormal region with its corresponding viewpoint number, original point cloud data index and specific value that determines it to be abnormal. For region A01, its associated attribute is "(viewpoint 3, R13, mean boundary distance = 3.5 cm)". The ranked abnormal regions and their associated attributes are integrated to obtain the plant segmentation distribution characteristics.

[0071] Please see Figure 3 The specific steps for obtaining the point cloud correction and optimization results are as follows:

[0072] S201: Based on the plant segmentation distribution characteristics, compare the current spatial distribution density of multi-view point cloud data, determine the spatial distribution differences of point cloud data, and match the regions with distribution trend deviations with region labels to obtain distribution difference characteristics.

[0073] Based on the plant segmentation distribution characteristics, this characteristic indicates that the abnormal region A01 is located in the point cloud data P3 at viewpoint 3, with the associated attribute "(viewpoint 3, R13, mean boundary distance = 3.5 cm)". By comparing the current spatial distribution density of the multi-view point cloud data, the operation is as follows: within the overlapping region R13, the space is divided into 10 cm × 10 cm × 10 cm cube units. The number of points in each unit for P1 and P3 is calculated, obtaining density distribution matrices M1 and M3. In a certain cell Ci of R13, the value of M1 is 85, and the value of M3 is 30, thus determining the spatial distribution of the point cloud data. For the difference, a density difference threshold of 20% is set. This threshold is obtained by calculating the average difference of point cloud density from each viewpoint within the stable region S01 and rounding it up. The density difference percentage between M1 and M3 in cell Ci is calculated to be 64.7%, which is greater than 20%. Regions with deviations in distribution trend are matched with region labels. Since cell Ci is located in region R13 and its density difference of 64.7% exceeds the threshold of 20%, cell Ci and the surrounding set of continuous cells with similar high differences are marked as deviation regions and matched with region label A01 to obtain the distribution difference characteristics.

[0074] S202: Based on the distribution difference characteristics, compare the distribution position of each region's point cloud data in the spatial set, determine whether it is located in the boundary region of the distribution sequence, and identify the boundary region as the difference response type to obtain the boundary response category group;

[0075] Based on the distribution difference feature, this feature records the sub-regions with density deviations within the abnormal region A01. The distribution position of each region's point cloud data in the spatial set is compared to determine whether it is located in the boundary region of the distribution sequence. The operation is as follows: the point set belonging to region A01 in point cloud P3 of viewpoint 3 is extracted, and the spatial geometric center G3 of this point set is calculated. Simultaneously, the spatial geometric center G1 of the corresponding region R13 point set in point cloud P1 of viewpoint 1 is calculated. Using G1 as a reference, the distribution position of the point set in region A01 is represented by the vector difference between G3 and G1. Regions located at the boundary are respectively identified as the difference response type, "boundary". This refers to the extreme case of deviation from the distribution location. The deviation distance threshold is set to 2.4 cm. This threshold is set based on 3 times the average distance of 0.8 cm between the geometric centers of the corresponding point sets from different perspectives within the entire stable region S01. If the magnitude of the vector (G3-G1) is greater than 2.4 cm, the point set in region A01 is determined to be on the "outer boundary" of the distribution sequence and is marked as "outer bias response". If the magnitude is between 1.2 cm and 2.4 cm, it is marked as "edge response". The distance between G3 and G1 is calculated to be 2.9 cm, which is greater than 2.4 cm. Therefore, region A01 is marked as "outer bias response", and the boundary response category group is obtained.

[0076] S203: Based on the boundary response category group, optimize the geometric correction parameter configuration of the corresponding region for the response type region, and integrate the adjustment results of the region by recording the optimization direction and parameter correction content to obtain the point cloud correction optimization result;

[0077] Based on the boundary response category group, this group identifies the abnormal region A01 as "outside bias response". For each response type region, the geometric correction parameter configuration of the corresponding region is optimized. The operation is as follows: the point cloud data of region A01 is identified as originating from the point cloud P3 of viewpoint 3. Its initial geometric correction parameters are the rotation matrix R_init and the translation vector T_init. The optimization objective is to reduce the distance between the geometric center G3 of the point set in region A01 and the reference center G1. By recording the optimization direction and parameter correction content, the optimization direction vector V = G1 - G3 is first calculated. Then, the translation vector T_init is corrected according to the vector V. The correction step size is set to 10% of the magnitude of vector V, that is, the correction amount ΔT = 0.1 × V. The new translation vector T_new = T_init + ΔT. At the same time, the rotation matrix R_init is fine-tuned, with an adjustment angle of 0.1 degrees. The rotation axis is a vector perpendicular to V and the Z-axis of the coordinate system. The new... The point cloud with parameters (R_new, T_new) in region A01 is recalculated, its geometric center G3' is recalculated, and the distance to G1 is recalculated. The optimized distance value is recorded as decreasing from 2.9 cm to 2.6 cm, and the parameter correction content is recorded as "(ΔT_x, ΔT_y, ΔT_z, ΔR_angle=0.1°)". Here, the symbols (ΔT_x, ΔT_y, ΔT_z, ΔR_angle=0.1°) represent the specific correction content of translation and rotation parameters in this geometric correction optimization process: ΔT_x, ΔT_y, and ΔT_z are the correction components of the translation vector in the X, Y, and Z coordinate axes, respectively, used to describe the fine adjustment of the point cloud model in spatial position; ΔR_angle=0.1° represents the angular fine adjustment magnitude applied to the rotation matrix, that is, a rotation correction of 0.1 degrees around the rotation axis perpendicular to the optimization direction vector V and the Z axis of the coordinate system. This parameter combination collectively reflects the refined correction operation for the abnormal region A01 during the spatial geometric alignment process, providing a quantitative basis for the accurate stitching and overall optimization of point cloud data. The adjustment results of the integrated region are combined with the region identifier A01, the response type "outer bias response," the content of this parameter correction, and the optimized distance value into a single record, yielding the point cloud correction optimization result.

[0078] Please see Figure 4 The specific steps for obtaining the point cloud stitched distribution feature dataset are as follows:

[0079] S301: Based on the point cloud correction and optimization results, compare the distribution status of the viewpoint point cloud data with the corresponding stitching configuration, and locate the set of regions that need to be optimized by filtering out the incompletely aligned areas, and establish an optimization label group;

[0080] Based on the point cloud correction optimization results, which include the geometric correction parameter adjustment records of region A01 under viewpoint 3, the distribution status of the viewpoint point cloud data and the corresponding stitching configuration are compared. The operation is as follows: the updated correction parameters (R_new, T_new) are applied to the entire point cloud data P3 of viewpoint 3 to generate the optimized point cloud P3'. Then, P3' and P1 are stitched together. By filtering out regions that are not fully aligned, the filtering criterion is to calculate the average distance between the point originating from P1 and the nearest neighbor point originating from P3' within the overlapping region R13 after stitching. The good alignment threshold is set to 1.0 cm. If the average distance is greater than this threshold, it is considered not fully aligned. The average distance of region R13 is calculated to be 1.5 cm, which is greater than 1.0 cm. The set of regions that need to be optimized is located. Therefore, region R13 is located as the region to be optimized, and an optimization label group is established. This group contains a list of region labels that need further optimization. Currently, it is "{R13}", and it is associated with its current alignment status value of 1.5 cm.

[0081] S302: Based on the optimized identifier group, the stitching accuracy is adjusted step by step for the region. After each optimization adjustment, the point cloud alignment status and corresponding spatial deviation are recorded synchronously. The corresponding optimization status data is constructed according to the time series structure to obtain the optimization change trend set.

[0082] Based on the optimized identifier group, the current group contains region R13 to be optimized. For this region, a step-by-step optimization method is used to adjust the stitching accuracy. The operation is as follows: The correction parameters (R_new, T_new) of the viewpoint 3 point cloud P3' corresponding to region R13 are optimized in a second iteration. The iteration step size is reduced to half of the previous one, i.e., the new translation correction ΔT' = 0.05 × (G1 - G3'), and the rotation angle is adjusted to 0.05 degrees. A new point cloud P3'' is generated. After each optimization adjustment, the point cloud alignment status and corresponding spatial deviation are recorded synchronously. After the second optimization, region R13 is recalculated. The average nearest neighbor distance between P1 and P3'' decreases to 1.2 cm, which is the new alignment state. The spatial deviation is this distance value. Construct the corresponding optimization state data according to the time series structure, and take each optimization iteration as a node of the time series, recording "(iteration number = 2, average distance = 1.2 cm, parameter correction = (ΔT'_x, ΔT'_y, ΔT'_z, ΔR'_angle = 0.05°)". Concatenate this record with the record of the first iteration "(iteration number = 1, average distance = 1.5 cm)" to obtain the optimization change trend set.

[0083] S303: Based on the optimization trend set, analyze the alignment status changes of the region before and after continuous optimization, and perform linkage filtering according to the direction of spatial deviation change. Combine the region identifiers that meet the filtering conditions with the alignment status to obtain the point cloud stitching distribution feature dataset.

[0084] Based on the optimization trend set, which records the state of region R13 after two consecutive optimizations: "[(Iteration 1, 1.5 cm), (Iteration 2, 1.2 cm)]", the alignment state changes of the region before and after continuous optimization are analyzed. The operation is as follows: the rate of change of spatial deviation between two adjacent iterations is calculated. From iteration 1 to iteration 2, the change is 0.3 cm, and the rate of change is 20%. Linked filtering is performed according to the direction of change of spatial deviation. The filtering conditions are set as follows: for two consecutive iterations, the spatial deviation value is monotonically decreasing and the decrease rate is greater than 5%. The current change rate is 20%, which is greater than 5%, and the deviation value decreases from 1.5 cm to 1.2 cm, which satisfies the monotonically decreasing condition. Therefore, region R13 meets the filtering conditions. The region identifier that meets the filtering conditions is combined with the alignment state. The region identifier "R13" is combined with the final alignment state "average distance = 1.2 cm" to form the record "(R13, 1.2 cm)". This process is repeated for all optimized regions to obtain the point cloud stitching distribution feature dataset.

[0085] Please see Figure 5 The specific steps for obtaining the crop 3D point cloud growth evaluation model are as follows:

[0086] S401: Based on the point cloud stitching distribution feature dataset, the point cloud correction and optimization results are numbered and paired with the crop phenotypic information extracted from the viewpoint point cloud data. Farmland environmental elements and crop growth status information are sorted out, and multi-information combination data of each viewpoint point cloud data is constructed to obtain the environmental information set.

[0087] Multi-information combined data refers to point cloud data from each viewpoint as the basic unit, which is used to number and pair the point cloud correction and optimization results corresponding to the viewpoint with the extracted crop phenotypic information. At the same time, it is associated and organized with farmland environmental elements and crop growth status information corresponding to the viewpoint, so that point cloud, phenotypic and environmental and growth information are combined together. The multi-source information under the same viewpoint is structured and summarized to form multi-information combined data of viewpoint point cloud data.

[0088] Based on a point cloud stitched distribution feature dataset containing the optimized alignment status of each region, such as "(R13, 1.2 cm)," crop phenotypic information is extracted from the self-corrected multi-view stitched point cloud. For maize plant P-001, its plant height of 2.5 meters, leaf area index of 3.8, and canopy coverage of 85% are extracted. Next, the point cloud correction and optimization records corresponding to plant P-001 are associated and paired to form a combination of "(P-001, 2.5 meters, 3.8, 85%, correction record)." Farmland environmental elements and crop growth status information are then organized, and data corresponding to the location of plant P-001 are collected. The soil nitrogen content was 1.2 g / kg, soil moisture content was 22%, light duration was 10 hours / day, and the crop was in the "tasseling stage" growth stage. Multi-information combination data was constructed for each viewpoint point cloud data, integrating all the aforementioned information into a single record: "{Plant ID: P-001, Point Cloud Data Index, Correction and Optimization Results, Phenotypic Information: {Plant Height: 2.5, Leaf Area Index: 3.8, Canopy Coverage: 0.85}, Environmental Elements: {Nitrogen Content: 1.2, Moisture: 0.22, Light: 10}, Growth Status: Tasseling Stage}". This process was repeated for all plant point clouds to obtain the environmental information set.

[0089] S402: Based on the environmental information set, compare the distribution of information in point cloud data, determine the participation ratio of farmland environmental elements and crop growth status information in the combined data, and establish a weight allocation structure by statistically configuring the weight of difference information.

[0090] Based on an environmental information set containing multi-information combination data of multiple plants, a subset of N=100 plant samples was selected to analyze the contribution of each information item to the variance of plant height. Through principal component analysis, it was found that 70% of the variation in plant height is explained by three variables: leaf area index, nitrogen content, and water content. Among them, leaf area index contributes 40%, nitrogen content contributes 20%, and water content contributes 10%. The remaining 30% can be handled in two ways: First, it can be treated as the overall contribution of "other information items," and distributed proportionally and normalized according to the marginal explanatory power of each item on plant height (such as the partial correlation coefficient with plant height or the comprehensive score of the standardized coefficient × principal component loading in multiple regression), so that the sum of the weights in "other information items" is 0.3. Second, a separate weight of "residual / unexplained term" W_res=0.3 is set to characterize environmental disturbances, measurement noise, or model structure biases not captured by the current variable system. This proportion can be gradually eroded by adding new variables or nonlinear terms (interactive, kernel methods). This results in a weight vector: leaf area index (W_LAI=0.4), nitrogen content (W_N=0.2), and water content (W_W=0.1), with {W_other_i} or W_res constituting the remainder of 0.3. The weights are then converged and calibrated using cross-validation or information criteria. The principal component analysis (PCA) process is briefly described as follows: First, standardize each variable in the environmental information set (zero mean, unit variance), construct a correlation coefficient matrix, and perform eigenvalue decomposition to obtain the principal components and their variance contribution rates and loading matrices. Select several principal components with high cumulative contribution rates, use loadings to map the original variables to the principal component space, and then combine this with plant height (target phenotype) for correlation / regression analysis to calculate the proportion of each original variable that explains the plant height variance. Finally, set initial weights according to the explanation proportions and normalize them to 1.0 to form the weight allocation structure.

[0091] S403: Based on the weight allocation structure, normalization processing is performed on environmental information and crop phenotypic information. The influence of information is superimposed according to the weight allocation principle, and the effect of the weight allocation result on the point cloud data is judged to obtain the crop three-dimensional point cloud growth evaluation model.

[0092] The weight allocation result refers to the normalization of environmental information and crop phenotypic information under the guidance of a predetermined weight allocation structure, eliminating the influence of different dimensions and numerical scales, and then superimposing and calculating each type of information after normalization according to the weight allocation principle to obtain the weight allocation result.

[0093] Based on the weighted allocation structure, the weights of each piece of information are defined, such as W_LAI=0.4, W_N=0.2, and W_W=0.1. For plant height, the minimum value in the sample is 2.0 meters, and the maximum value is 2.8 meters. Normalizing the plant height of plant P-001 (2.5 meters) yields (2.5-2.0) / (2.8-2.0)=0.625. Similarly, normalization is performed on all information such as leaf area index and nitrogen content within the [0,1] interval. Based on the weighted allocation principle, the influence of the information is superimposed, and the normalized information values ​​are multiplied by their corresponding weights. The growth assessment score S of plant P-001 was calculated by weighted summation as follows: 0.7 × 0.4 + 0.6 × 0.2 + 0.55 × 0.1 = 0.455. Here, 0.7, 0.6, and 0.55 represent the normalized values ​​of leaf area index, nitrogen content, and water content, respectively. The final weighted summation of all information items yielded a total score of 0.85. The remaining information items were set as photosynthetically active radiation (photosynthetically active radiation weight W_PAR = 0.12, normalized value x_PAR = 0.98, contribution 0.12 × 0.12). 98=0.1176), soil water stability (soil water stability weight W_SM=0.08, normalized soil water stability value x_SM=0.92, contribution 0.08×0.92=0.0736), temperature suitability (temperature suitability weight W_Temp=0.05, temperature suitability weight x_Temp=0.95, contribution 0.05×0.95=0.0475), pH suitability (pH suitability weight W_pH=0.03, normalized pH suitability value x_pH=0.90, contribution 0. ...98=0.1176), soil water stability (soil water stability weight W_SM=0.08, normalized soil water stability value x_SM=0.92, contribution 0.98=0.0736), soil water stability (soil water stability weight W_Temp=0.05, normalized pH suitability value x_pH=0.90, contribution 0.98=0.92, contribution 0.98=0.92, contribution 0.98=0.92, contribution 0.98=0.92, contribution 0.98=0.92, contribution 0.98=0.92, contribution The weights of the plant P-001 are calculated as follows: 0.0270 (0.03 × 0.90 = 0.0270), 0.0270 (Pest and Disease Stress Suitability Weight W_Pest = 0.02, Pest and Disease Stress Suitability Normalized Value x_Pest = 0.88, Contribution 0.02 × 0.88 = 0.0176), and the sum of the weights of the remaining items is 0.30. The total contribution is 0.1176 + 0.0736 + 0.0475 + 0.0270 + 0.0176 = 0.2833. Therefore, the final growth assessment score for plant P-001 is S = 0.455 + 0.2833 = 0.7383. To determine the effect of the weight allocation on the point cloud data, the calculated growth assessment score of 0.7383 for each plant is attached as an attribute to the point cloud data corresponding to plant P-001, resulting in a three-dimensional point cloud growth assessment model for the crop.

[0094] Please see Figure 6 The specific steps for obtaining the crop yield potential distribution dataset are as follows:

[0095] S501: Based on the crop three-dimensional point cloud growth assessment model, calculate the proportional merging result of crop phenotypic information and corresponding weight configuration in the point cloud data, analyze the distribution of proportionally merged data in the point cloud data, and determine the interval position of each point cloud data according to the upper and lower boundaries of the growth dynamic grading to obtain the signal grading interval category.

[0096] Based on a crop 3D point cloud growth assessment model, this model assigns a growth assessment score to each plant point cloud, such as 0.7383 for P-001. This operation directly calls the growth assessment score obtained in the previous process to analyze the distribution of proportionally merged data in the point cloud data. "Proportionally merged data" refers to the comprehensive data result formed by normalizing and weighting environmental and crop phenotypic information according to a weighted allocation structure during the construction of the crop 3D point cloud growth assessment model. This data reflects the comprehensive influence of different information sources (such as leaf area index, nitrogen content, water content, light, temperature, and soil moisture) on crop growth status; its essence is a weighted comprehensive index after multi-information fusion. Specifically, the various information of each plant sample is first normalized to eliminate dimensional differences, and then proportionally weighted to obtain the plant's comprehensive growth assessment score. The score is assigned as an additional attribute to each plant point cloud in the point cloud space, forming "proportionally merged data" with both physiological and environmental characteristics. Therefore, the proportionally merged data not only reflects the growth response ratio of crops under different environmental factors, but also provides a quantitative basis for subsequent growth grading and yield potential analysis at the point cloud level. Based on the upper and lower boundaries of the growth dynamic grading, the interval position corresponding to each point cloud data point is determined. First, the grading boundary is set, which is determined by statistical analysis of the growth assessment scores of all sample plants. The scores are divided into three levels according to percentiles: scores below the 33rd percentile (score < 0.60) are "low potential," scores between the 33rd and 66th percentiles (0.60 ≤ score < 0.80) are "medium potential," and scores above the 66th percentile (score ≥ 0.80) are "high potential." Plant P-001 has an assessment score of 0.85, which is greater than or equal to 0.80, therefore its corresponding interval position is "high potential," thus obtaining the signal grading interval category.

[0097] S502: Based on the signal classification interval category, classify and organize according to the interval judgment result, record the point cloud data number and signal merging result corresponding to each category, and output the distribution of each category through structured data to obtain the crop yield potential distribution dataset;

[0098] Based on the signal grading interval category, plant P-001 is classified into the "high potential" level. A "high potential" classification list is created, recording the point cloud data number and signal merging result corresponding to each category. The data number of plant P-001 and its growth evaluation score of 0.85 are added to this list, forming the record "(P-001, 0.85)". This operation is repeated for all plants. Plant P-002 has a score of 0.72, so it is recorded in the "medium potential" list: "(P-002, 0.72)". Plant P-003 has a score of 0.55, so it is recorded in the "low potential" list: "(P-003, 0.55)". The distribution of each category is output through structured data, and finally, three datasets are compiled. Each dataset contains all plant numbers and their evaluation scores for the corresponding potential level, resulting in a crop yield potential distribution dataset.

[0099] Please see Figure 7 This invention also provides a 3D point cloud modeling system based on multi-view images, comprising:

[0100] The point cloud distribution recognition module is based on multi-view images of farmland environment collected by depth camera, extracts the spatial distribution features of crops, and combines the overlapping areas and boundary characteristics of point cloud data to segment plants, identify abnormal areas, and obtain plant segmentation distribution features.

[0101] The geometric correction optimization module evaluates the spatial distribution density of the current point cloud data based on the plant segmentation distribution characteristics, compares it with the original distribution sequence, and adjusts the geometric correction parameters for areas with high or low distribution density to obtain the point cloud correction optimization results.

[0102] The stitching accuracy dynamic filtering module compares the distribution status of point cloud data item by item based on the point cloud correction and optimization results, optimizes the stitching accuracy, and filters the alignment areas that meet the conditions to obtain the point cloud stitching distribution feature dataset.

[0103] The multidimensional information normalization module, based on the point cloud splicing distribution feature dataset, combines the point cloud splicing distribution feature dataset, correction and optimization results, and farmland environment and crop growth information to perform weighted normalization, thereby obtaining a three-dimensional point cloud growth evaluation model for crops.

[0104] The yield prediction and discrimination module, based on the crop 3D point cloud growth assessment model, proportionally merges crop phenotypic information with weight configurations, makes interval judgments with the upper and lower limits of growth dynamic grading based on the merged results, classifies and organizes the data according to the interval judgments, and outputs the corresponding classifications to obtain the crop yield potential distribution dataset.

[0105] For ease of explanation, Figure 7 Only the main components of the system are shown. The system of this embodiment can be used to perform... Figure 1The technical solutions of the method embodiments shown are similar in principle and in effect, and will not be described again here.

[0106] In this embodiment of the invention, depth camera acquisition of multi-view images is used to analyze the spatial distribution characteristics of crop targets and the overlapping areas and boundary characteristics of point cloud data under different perspectives. This overcomes the problem of low point cloud reconstruction accuracy caused by inconsistent perspectives and lighting changes in traditional methods. By adjusting geometric correction parameters, the alignment accuracy of point cloud data is optimized, effectively reducing spatial deviations and ensuring accuracy in the stitching process. This further improves the quality of point cloud stitching results, especially in complex farmland environments, ensuring more accurate crop target extraction and phenotypic structure segmentation. Simultaneously, by combining farmland environmental factors and crop growth status information, crop growth dynamics are accurately assessed, the accuracy of yield prediction is optimized, the visualization and analysis capabilities of the crop growth process are enhanced, and the level of intelligent agricultural production management is improved.

[0107] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for modeling 3D point clouds based on multi-view images, characterized in that, Includes the following steps: S1: Based on the acquisition of multi-view images of farmland environment by depth camera, extract the spatial distribution characteristics of crop targets, analyze the overlapping area and boundary characteristics of point cloud data under different perspectives, identify individual plants and segment their phenotypic structure to obtain plant segmentation distribution characteristics. S2: Based on the plant segmentation distribution characteristics, determine the spatial distribution density of point cloud data, compare the boundary characteristics and overlapping area relationship of point cloud data from each viewpoint, adjust the geometric correction parameters for areas with inconsistent boundary characteristics, and obtain the point cloud correction optimization result. S3: Based on the point cloud correction and optimization results, match them item by item with the point cloud stitching process, analyze the spatial deviation of the incompletely aligned areas, gradually optimize the stitching accuracy, record the point cloud alignment status, and obtain the point cloud stitching distribution feature dataset. S4: Based on the point cloud splicing distribution feature dataset, combined with farmland environmental elements and crop growth status information, growth dynamic grading and yield prediction are performed to obtain a three-dimensional point cloud growth evaluation model for crops. The specific steps for obtaining the point cloud stitching distribution feature dataset are as follows: S301: Based on the point cloud correction and optimization results, compare the distribution status of the viewpoint point cloud data with the corresponding stitching configuration, and locate the set of regions that need to be optimized by filtering out the incompletely aligned regions, and establish an optimization identification group; S302: Based on the optimized identifier group, the splicing accuracy is adjusted step by step for the region. After each optimization adjustment, the point cloud alignment status and corresponding spatial deviation are recorded synchronously. The corresponding optimization status data is constructed according to the time series structure to obtain the optimization change trend set. S303: Based on the optimized change trend set, analyze the alignment state changes of the region before and after continuous optimization, and perform linkage filtering according to the direction of spatial deviation change. Combine the region identifiers that meet the filtering conditions with the alignment state to obtain the point cloud stitching distribution feature dataset.

2. The 3D point cloud modeling method based on multi-view images according to claim 1, characterized in that, The plant segmentation distribution features include plant number, segmentation range, and phenotypic category; the point cloud correction and optimization results include correction direction, optimization identifier, and distribution area classification; the point cloud splicing distribution feature dataset includes splicing category, distribution interval, and alignment level; and the crop 3D point cloud growth evaluation model includes evaluation labels and corresponding grading interval parameters.

3. The 3D point cloud modeling method based on multi-view images according to claim 1, characterized in that, The specific steps for obtaining the plant segmentation and distribution characteristics are as follows: S101: Based on the acquisition of multi-view images of farmland environment by depth camera, extract the spatial distribution characteristics of crop targets, and by analyzing the overlapping area and boundary characteristics of point cloud data under different perspectives, construct the pairing information of point cloud distribution and boundary characteristics from one perspective to obtain point cloud boundary characteristic group. S102: Based on the point cloud boundary characteristic group, determine the relationship between the boundary characteristics and overlapping areas of the viewpoint point cloud data, classify the areas with deviated boundary characteristics as abnormal areas, classify the areas with consistent boundary characteristics as stable areas, and classify the remaining areas as regular types to obtain point cloud response distribution labels. S103: Based on the point cloud response distribution labels, summarize the corresponding numbers and statuses of each type of region, and gather the associated attributes of abnormal type regions through sorting and grouping to obtain plant segmentation distribution features.

4. The 3D point cloud modeling method based on multi-view images according to claim 3, characterized in that, The specific steps for obtaining the point cloud correction and optimization results are as follows: S201: Based on the plant segmentation distribution characteristics, compare the current spatial distribution density of the multi-view point cloud data, determine the distribution difference of the point cloud data in space, and match the regions where the distribution trend deviates with the region identifiers to obtain the distribution difference characteristics. S202: Based on the distribution difference characteristics, compare the distribution position of each region's point cloud data in the spatial set, determine whether it is located in the boundary region of the distribution sequence, and identify the boundary regions as difference response types to obtain boundary response category groups; S203: Based on the boundary response category group, optimize the geometric correction parameter configuration of the corresponding region for each response type region. By recording the optimization direction and parameter correction content, integrate the adjustment results of the region to obtain the point cloud correction optimization result.

5. The 3D point cloud modeling method based on multi-view images according to claim 4, characterized in that, The specific steps for obtaining the crop three-dimensional point cloud growth evaluation model are as follows: S401: Based on the point cloud splicing distribution feature dataset, the point cloud correction and optimization results are numbered and paired with the crop phenotypic information extracted from the viewpoint point cloud data, farmland environmental elements and crop growth status information are sorted out, and multi-information combination data of each viewpoint point cloud data is constructed to obtain an environmental information set. S402: Based on the environmental information set, compare the distribution of information in point cloud data, determine the participation ratio of farmland environmental elements and crop growth status information in the combined data, and establish a weight allocation structure by statistically configuring the weight of difference information. S403: Based on the weight allocation structure, normalization processing is performed on environmental information and crop phenotypic information, the influence of information is superimposed according to the weight allocation principle, and the effect of the weight allocation result on the point cloud data is judged to obtain the crop three-dimensional point cloud growth evaluation model.

6. The 3D point cloud modeling method based on multi-view images according to claim 5, characterized in that, The multi-information combined data refers to point cloud data from each viewpoint as the basic unit, which is to number and pair the point cloud correction and optimization results corresponding to the viewpoint with the extracted crop phenotypic information, and at the same time associate and organize the farmland environmental elements and crop growth status information corresponding to the viewpoint, so that the point cloud, phenotypic and environmental and growth information are combined together, and the multi-source information under the same viewpoint is structured and summarized to form multi-information combined data of viewpoint point cloud data. The weight allocation result refers to the normalization of environmental information and crop phenotypic information under the guidance of a predetermined weight allocation structure, eliminating the influence of differential dimensions and numerical scales, and then superimposing and calculating each type of information after normalization according to the weight allocation principle to obtain the weight allocation result.

7. The 3D point cloud modeling method based on multi-view images according to claim 1, characterized in that, The method also includes step S5: S5: Based on the crop three-dimensional point cloud growth evaluation model, the crop phenotypic information is merged with the weight configuration ratio, and the merged result is used to make interval judgment with the upper and lower limits of growth dynamic classification. The crop yield potential distribution dataset is obtained by classifying and organizing the data according to the interval judgment category. The crop yield potential distribution dataset includes prediction labels, decision intervals, and classification indexes.

8. The 3D point cloud modeling method based on multi-view images according to claim 7, characterized in that, The specific steps for obtaining the crop yield potential distribution dataset are as follows: S501: Based on the crop three-dimensional point cloud growth evaluation model, calculate the ratio of crop phenotypic information and corresponding weight configuration in the viewpoint point cloud data, analyze the distribution of the ratio merged data in the point cloud data, and determine the interval position of each point cloud data according to the upper and lower boundaries of the growth dynamic grading to obtain the signal grading interval category. S502: Based on the signal classification interval categories, classify and organize according to the interval judgment results, record the point cloud data number and signal merging result corresponding to each category, and output the distribution of each category through structured data to obtain the crop yield potential distribution dataset.

9. A 3D point cloud modeling system based on multi-view images, characterized in that, The system is used to implement the 3D point cloud modeling method based on multi-view images as described in any one of claims 1-8, and the system comprises: The point cloud distribution recognition module is based on multi-view images of farmland environment collected by depth camera, extracts the spatial distribution features of crops, and combines the overlapping areas and boundary characteristics of point cloud data to segment plants, identify abnormal areas, and obtain plant segmentation distribution features. The geometric correction optimization module evaluates the spatial distribution density of the current point cloud data based on the plant segmentation distribution characteristics, compares it with the original distribution sequence, and adjusts the geometric correction parameters for areas with high or low distribution density to obtain the point cloud correction optimization results. The stitching accuracy dynamic filtering module compares the distribution status of the point cloud data item by item based on the point cloud correction and optimization results, optimizes the stitching accuracy, and filters the alignment areas that meet the conditions to obtain the point cloud stitching distribution feature dataset. The multidimensional information normalization module, based on the point cloud splicing distribution feature dataset, combined with the point cloud splicing distribution feature dataset, correction and optimization results, and farmland environment and crop growth information, performs weighted normalization to obtain a crop three-dimensional point cloud growth evaluation model. The yield prediction and discrimination module, based on the crop 3D point cloud growth evaluation model, proportionally merges crop phenotypic information with weight configuration, makes interval judgments with the upper and lower limits of growth dynamic grading based on the merged results, classifies and organizes the data according to the categories of the interval judgments, and outputs the corresponding categories to obtain the crop yield potential distribution dataset.

Images