CN-122023854-A - High-precision aphid identification and counting method in mobile computing scene

CN122023854ACN 122023854 ACN122023854 ACN 122023854ACN-122023854-A

Abstract

The invention provides a high-precision aphid identification and counting method in a mobile computing scene, which relates to the technical field of intelligent monitoring of agricultural pests, and the method comprises the steps of taking a monitored plant as a center, collecting a plurality of groups of visible light images and depth images along a circular path, and constructing a three-dimensional point cloud model of the plant through coordinate registration and point cloud fusion; the method comprises the steps of adopting a self-adaptive region growing algorithm to divide and extract independent blade units, carrying out plane fitting and two-dimensional projection to generate a blade orthographic visualization image, identifying the orthographic visualization image through a pre-training aphid identification model to obtain an aphid boundary box set and a fine granularity class, establishing a mapping relation between three-dimensional point clouds and two-dimensional images, mapping the boundary box back to the three-dimensional point clouds to obtain an aphid candidate point set, carrying out double constraint three-dimensional clustering by utilizing the consistency of the spatial Euclidean distance and the fine granularity class labels, and screening and marking independent aphid individuals. The method can realize high-precision aphid identification and counting, and can be widely applied to the intelligent monitoring field of agricultural pests.

Inventors

LI SHIRONG
TIAN XING
Qin Ehua
TIAN MIN

Assignees

延安大学
陕西艾睿生物科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260316

Claims (9)

1. The high-precision aphid identification and counting method under the mobile computing scene is characterized by comprising the following specific steps of: S1, synchronously collecting a plurality of groups of visible light images and corresponding depth images at fixed intervals along a circular path on the horizontal plane around a monitoring plant serving as a center, and fusing the images with a point cloud through coordinate registration based on the collected images to construct a three-dimensional point cloud model of the plant; S2, based on a three-dimensional point cloud model of a monitoring plant, dividing and extracting each independent blade unit by adopting a self-adaptive region growing algorithm, and carrying out plane fitting and two-dimensional projection on each blade unit to generate a positive visualization image of the blade; S3, identifying the positive visualization images of all the blades through a pre-trained aphid identification model to obtain a aphid boundary box set in the positive visualization images of the blades and fine granularity categories corresponding to the aphid boundary boxes; S4, mapping the pixel areas in the boundary boxes back to the three-dimensional point cloud through the mapping relation between the three-dimensional point cloud and the two-dimensional image, obtaining aphid candidate point sets corresponding to each aphid boundary box and matching the corresponding fine granularity categories; And S5, carrying out double constraint three-dimensional clustering on all aphid candidate point sets by utilizing the consistency of the spatial Euclidean distance and the fine granularity class label, screening and marking independent aphid individuals, and completing high-precision aphid identification and counting.
2. The method for identifying and counting aphids with high precision under a mobile computing scene is characterized by constructing a three-dimensional point cloud model of plants, specifically, by taking the center of a monitored plant as an origin, establishing a world coordinate system, taking the center of the monitored plant as a circle center, arranging acquisition points at a peripheral horizontal plane along a circular path according to fixed interval angles, acquiring coordinates of each acquisition point in the world coordinate system, taking each acquisition point as the origin, establishing a corresponding camera coordinate system, acquiring a visible light image and a depth image of the plant to be monitored at each acquisition point, determining three-dimensional coordinates of each pixel point in the camera coordinate system in the visible light image and the depth image of each acquisition point, taking the coordinates of each acquisition point in the world coordinate system as a conversion basis, converting the three-dimensional coordinates of each corresponding pixel point in the camera coordinate system into the world coordinate system, and integrating all cameras and all effective pixels through the points obtained in the steps.
3. The method for identifying and counting aphids with high precision in a mobile computing scene as claimed in claim 1 or 2 is characterized in that the segmentation and flattening of the blade level are carried out based on a plant three-dimensional point cloud model, and specifically comprises the following steps: for each point in the point cloud, setting a field fixed radius by taking the point as a circle center, setting a local field, and calculating the point in the local field of each point by adopting principal component analysis, wherein the formula is as follows: Wherein, the As a covariance matrix of the points in the local neighborhood, Is the local area internal point Is defined by the point cloud coordinate vector of (a), For centroid coordinate vectors for all points in the local neighborhood, For the number of points in the neighborhood, for covariance matrix Performing feature decomposition to obtain normal vectors and curvatures of the corresponding points; The method comprises the steps of marking all points of a three-dimensional point cloud model of a plant as undivided points, setting a curvature minimum point as a first seed point, taking the first seed point as a starting point of the region growing algorithm, setting a growing condition, marking the Euclidean distance between two points as segmented points when the Euclidean distance between the two points is required to be smaller than a fixed distance threshold and the normal vector included angle between the two points is smaller than a fixed angle threshold, marking the segmented points into a growing area of the current seed point, naturally stopping growing when a point meeting the growing condition is not found in the growing area, finishing the blade segmentation of the current seed point, setting a seed point screening threshold, taking the curvature minimum point in the segmented points as a new first seed point, and repeating the segmentation step until no new seed point is generated, thereby finishing the segmentation of all the single blade units of the plant.
4. The method for identifying and counting aphids with high precision in a mobile computing scene as claimed in claim 3, wherein the method is characterized in that for the segmented blade units, the front and back three-dimensional point clouds of the blade are projected to a plane by using a fitting plane algorithm to generate a front visualization image of the blade, specifically: Smoothing each segmented blade point cloud by adopting a moving least square method, obtaining a local surface normal vector of each point, and fitting a best fitting plane of the blade by using a RANSAC algorithm, wherein the plane equation formula is as follows: Wherein, the Is that The best-fit plane of the blade unit, Is the unit normal vector of the best fit plane, Is that In blade units The point cloud coordinates of the points, Is that Setting RANSAC inner point threshold value, recording smaller than inner point threshold value In the plane of And (3) establishing a two-dimensional coordinate system, calculating the two-dimensional coordinates of each three-dimensional point in the point cloud of the blade unit under the corresponding two-dimensional coordinate system, normalizing and discretizing the two-dimensional coordinates of all the blade units to an image grid with fixed resolution, and generating a front-view image of the blade.
5. The method for identifying and counting aphids with high precision under a mobile computing scene as claimed in claim 4 is characterized in that the method comprises the steps of identifying the positive images of all the blades through a pre-trained aphid identification model to obtain a aphid boundary box set with fine granularity type in the positive images of all the blades, specifically, inputting the positive images of all the blades of a monitoring plant into the pre-trained aphid identification model, outputting all the corresponding predicted aphid boundary boxes for the positive image model of each blade, wherein each aphid boundary box comprises frame coordinates and fine granularity type labels.
6. A method for identifying and counting aphids with high precision in a mobile computing environment according to claim 5, characterized in that the specific operation of obtaining the corresponding point set of each aphid bounding box is as follows: Establishing S2 a mapping relation between pixel points in a front-view image of a leaf and a three-dimensional point cloud, obtaining all aphid bounding boxes based on an aphid recognition model, finding out an aphid area corresponding to the three-dimensional area of the leaf to which the plant leaf belongs through the mapping relation and depth image information on pixel areas of all aphid bounding boxes on the front-view image of the monitored plant leaf, and forming an aphid candidate point set by the three-dimensional points corresponding to the aphid bounding boxes , wherein, Representation of In the blade A border frame of the individual aphids is provided, Aphid bounding box The corresponding pixel area is provided with a plurality of pixel areas, In order for the mapping relationship to be a function of the mapping relationship, Is that Mapping points in a three-dimensional point cloud model of the plant, based on the point set corresponding to all boundary boxes of all blades in the obtained plant to be identified, combining all aphid candidate point sets of all blades into a total point set The formula is as follows: Wherein, the For the total number of blades, Is that The total number of aphid bounding boxes in the leaf, To develop aphid candidate point set Synchronously recording the fine granularity class label corresponding to each aphid bounding box.
7. The method for identifying and counting aphids with high precision in a mobile computing scene as claimed in claim 6, wherein the method is characterized in that, for all aphid candidate point sets, three-dimensional clustering is performed by double constraint by using consistency of spatial Euclidean distance and fine granularity class labels, namely, all aphid candidate point sets of all leaves are combined into a total point set, and clustering is performed for all aphid candidate point sets in the total point set by adopting a clustering algorithm based on Euclidean distance so as to distinguish and mark independent aphid individuals, and the method comprises the following specific steps: The method comprises the steps of obtaining mass center coordinates corresponding to all aphid candidate point sets in a total point set, initializing each aphid candidate point set into an independent cluster, setting a clustering distance threshold, traversing the independent clusters corresponding to all aphid candidate point sets in the total point set, classifying the independent clusters with the mass center coordinates smaller than the clustering distance threshold and the same fine granularity category as each other into the same cluster, wherein each combined cluster represents a new independent cluster, the independent clusters represent aphid individuals, counting the number of the final independent clusters to be the accurate count of aphids, and simultaneously outputting the three-dimensional position and the fine granularity category of each aphid individual to finish high-precision aphid identification and count.
8. The method for identifying and counting aphids with high precision under a mobile computing scene as claimed in claim 6 is characterized by comprising the steps of carrying out double constraint three-dimensional clustering on all aphid candidate point sets by utilizing consistency of spatial Euclidean distance and fine granularity category labels, specifically, carrying out preliminary three-dimensional clustering by adopting a density-based DBSCAN algorithm, setting a required neighborhood radius and minimum neighbor number of the DBSCAN algorithm, using the fine granularity category consistency as an additional constraint condition, combining all aphid candidate point sets of all blades into a total point set, obtaining centroid coordinates corresponding to all aphid candidate point sets in the total point set, and defining a double constraint distance function, wherein the formula is as follows: Wherein, the For the double constraint distance, the distance is, Is a candidate point set for aphids The euclidean distance of the centroid, Penalty component for fine granularity class, if the fine granularity class is the same If the fine grain class is different , Setting fixed distance threshold for fine granularity class weight coefficient, if And traversing all aphid candidate point sets of the total point set, and counting all the preliminary clustering clusters.
9. A high-precision aphid identification and counting method under a mobile computing scene as claimed in claim 8 is characterized in that consistency verification is carried out on each preliminary cluster, according to the number of aphid candidate point sets contained in each preliminary cluster, the occurrence times of fine grain classes in the preliminary cluster are counted, a class determination threshold is set, the occupation ratio of a main class is obtained, if the occupation ratio of the main class is larger than the class determination threshold, the fine grain class of the preliminary cluster is considered to be consistent and accepted as a final aphid individual cluster, if the occupation ratio is smaller than the class determination threshold, the aphid candidate point sets in the preliminary cluster possibly contain a plurality of aphid individuals with different classes and need to be split, the preliminary cluster candidate point sets are grouped according to fine grain class, the space distribution is checked for each fine grain class group, if the maximum distance between every two of the aphid candidate point sets in each fine grain class group is smaller than or equal to the splitting distance threshold, the fine grain class group is independent to be a new final aphid individual cluster, if the maximum distance between every two of the fine grain class candidate point sets in each fine grain class group is larger than the class determination threshold, the final cluster is further determined to be the final aphid individual cluster, and the final distance between every two aphid candidate point clusters is further determined to be the final cluster label.

Description

High-precision aphid identification and counting method in mobile computing scene Technical Field The invention relates to the technical field of intelligent monitoring of agricultural pests, in particular to a high-precision aphid identification and counting method in a mobile computing scene. Background Along with the continuous improvement of the modern agriculture precise management demand, the intelligent monitoring, prevention and control technology of crop diseases and insect pests has become an important direction of agricultural informatization development. Aphids, one of the most damaging pests in agricultural production, pose a serious threat to crop yield and quality due to their rapid propagation and broad spread characteristics. The traditional aphid monitoring method mainly relies on manual visual inspection and a simple counting tool, so that the efficiency is low, and large-scale and continuous accurate monitoring is difficult to realize. At present, various aphid identification and counting techniques have been proposed. The prior art with publication number CN115620050A discloses an improved YOLOv aphid identification and counting method based on a climate chamber environment, which uses a YOLOv model as a basic network, and introduces CBAM attention mechanism and a transducer module to treat the influence of small aphid scale, aggregation degree and illumination intensity. The prior art with publication number CN119206780A discloses a rapid field aphid counting method, which is based on a deformable convolution network and a self-attention module to construct an aphid counting model, and the aphid counting efficiency is improved by identifying and directly counting aphids in an image. The prior art with the publication number of CN121353925A discloses a small target aphid identification method, which generates a bounding box label and a continuous density map label through double-label collaborative labeling, and improves the monitoring capability of small target aphids under a complex background by utilizing multi-scale feature extraction and fusion capability. In the field of plant counting, the prior art with the publication number of CN115424257A discloses a plant seedling stage plant counting method based on an improved multi-column convolutional neural network, which extracts different scale features through a multi-column convolutional branch attention encoder and performs feature fusion through a multi-branch fusion module. The prior art with publication number CN113392775 discloses an automatic recognition and counting method for sugarcane seedlings based on a deep neural network, which learns plant characteristics through the deep convolutional neural network and splices and de-duplicates large-size aerial monitoring results. However, the prior art still has the following defects that the existing aphid identification method based on deep learning still has obvious technical defects that firstly, two-dimensional images which are dependent on a single visual angle are seriously influenced by leaf shielding, attitude change and illumination conditions, accurate counting in a field environment is difficult to realize, secondly, when aphids are densely distributed, detection frames in the two-dimensional images are seriously overlapped, individuals cannot be accurately distinguished, the counting accuracy is greatly reduced, thirdly, the aphids in different growth stages cannot be distinguished due to the lack of fine granularity classification capability and a three-dimensional space verification mechanism, the differential hazard degree of the aphids in different growth stages to crops is difficult to evaluate, and the establishment of an accurate prevention and control strategy is limited. The above information disclosed in the background section is only for enhancement of understanding of the background of the disclosure and therefore it may include information that does not form the prior art that is already known to a person of ordinary skill in the art. Disclosure of Invention The invention aims to provide a high-precision aphid identification and counting method in a mobile computing scene, so as to solve the problems in the background technology. In order to achieve the above purpose, the present invention provides the following technical solutions: a high-precision aphid identification and counting method under a mobile computing scene comprises the following specific steps: S1, taking a monitoring plant as a center, synchronously collecting a plurality of groups of visible light images and corresponding depth images along a circular path on the horizontal plane around the monitoring plant at fixed interval angles, and constructing a three-dimensional point cloud model of the plant through coordinate registration and point cloud fusion; S2, based on a three-dimensional point cloud model of a monitoring plant, dividing and extracting each independent blade unit by adopting