3D Point Cloud-based Small Plant Phenotypic Characteristics Extraction

In recent years,with the rapid development of modern information technology,many new technical means had come out to meeting the need of measuring agriculture,life sciences and other disciplines.So plant phenotyping was proposed to meet the urgent need of the agricultural field and developed rapidly.However,plant phenotype measurement technology is relatively lagging,when it was compared with the more rapid advancement of plant genomics.And it has become a bottleneck hindering the process of plant breeding.Therefore,the key issue of modern agriculture is researching high-throughput,high-precision and non-destructive plant phenotype monitoring systems.Individual plants and plant organs block each other seriously in the plant canopy because of the diversity and irregular morphology of the plant.Moreover,with the complex growth environment of plants,extracting plant phenotypic features from two-dimensional images will be subject to many restrictions.Therefore,extracting plant phenotypic characteristics based on three-dimensional information has become a popular research direction in modern digital agriculture.This paper will do a further study to get plant phenotypic parameters from the acquisition and processing of 3D point cloud data and the application method of 3D point cloud deep learning.The main work is as follows:1.A method of crop group 3D point cloud reconstruction and phenotypic parameter extraction based on multi-view stereo is proposed.In this method,phenotypic parameters such as plant height,leaf length and leaf width of each plant in the population were obtained by taking cucumber population plants as an example.Firstly,the multi-view image of the crop group is acquired by a digital camera,and then the three-dimensional dense point cloud with color information is reconstructed by the SVM-MVS method.The reconstructed point cloud data is preprocessed to generate a point cloud containing only plant parts.Euclidean clustering and region growth algorithms are used to segment point clouds,and finally point cloud data of crop leaves are obtained.The leaf point cloud surfaces are reconstructed by triangulation,and then plant phenotypic parameters such as leaf length and leaf width are obtained through projection calculation.The experimental results show that the calculated value is highly correlated with the manual measurement value(R~2=0.80?0.93,RMSE=0.79cm?1.36cm),which indicates that the SFM-MVS method provides a feasible solution for the rapid and low-cost measurement of plant phenotypic parameters.2.Research on the automatic segmentation technology of plant three-dimensional point cloud based on deep learning has been carried out and the segmentation of single crops from crop groups and the segmentation of leaf organs from single crops have been realized.Firstly,the 3D point cloud is manually annotated,and then input to the point cloud instance segmentation network based on 3D bounding box regression to complete the segmentation task of the above two scenes.Data augmentation of 3D point cloud is realized by rotation.In the test of individual cucumber leaf segmentation,the best average precision rate was 0.809,and the average recall rate was 0.884;in the test of individual plant population segmentation,the best average precision rate was 0.875,and the average recall rate was 0.897.