Plant Organ Growth Tracking Based On Point Cloud Data

Plant phenotypic analysis is of great significance to the development of agricultural engineering and is one of the core issues in crop science and breeding.Because of the spatial and temporal nature and synchronization of plant growth,understanding the growth and development of individuals helps to reveal the growth of the whole block and improve the planting methods.In recent years,the use of 3D point cloud data for automated phenotypic analysis of plant growth has received extensive attention from the industry and academia.Regularly scanning the same plant to obtain spatio-temporal point cloud data and processing and phenotypic analysis of the data,it is a common technical idea to monitor the growth and measure the relevant data of organs by observing the changes of three-dimensional morphology of plants.The most difficult task in spatio-temporal point cloud data is to match and track the same organ during the growing period.Because plant growth has phototropism,its topological structure may change with time.And because crop growth is unpredictable,the size of the organ may increase,the number may increase,or organ wither may occur,so the growing crop for single-organ threedimensional tracking is a big challenge.In order to monitor the three-dimensional growth of plant organs and reduce the tedious manual labor in agricultural breeding,this master ’s degree thesis takes three-dimensional automatic crop phenotype analysis as the research field,and proposes a plant organ growth tracking method based on spatio-temporal point cloud data.The main contributions of this thesis are as follows:(1)This thesis proposes a graph-based plant organ matching(GPOM)algorithm for 3D crop point clouds,which corresponds to organ matching relationships under three different organ growth events.This algorithm uses the idea of graph theory to solve the many-to-many matching problem of organs between time series point clouds.It takes plant organs as vertices and the association between organs as edges,and then constructs a cost matrix.Each element in the cost matrix represents the degree of association between corresponding organs.Aiming at three different organ growth events(budding,withering or disappearing,normal growth)in the process of plant growth,a method of expanding cost matrix is proposed.The original cost matrix is extended to four submatrices,and each sub-matrix is used to represent different matching situations.A new calculation method for cost matrix elements is proposed.We design a weighted matching cost calculation formula.It can accurately measure the comprehensive similarity between different organs in two adjacent periods by integrating three distances.We have determined a set of optimal weight configurations through a large number of experiments.The ablation analysis for this matching distance shows that each item in the matching cost calculation formula has a positive effect on improving the tracking accuracy.(2)A two-stage method for multi-to-multi-tracking of plant organ point clouds is proposed,which is named as a two-step Organ Growth Tracking Method(TOGTM)based on crop point cloud data.This method first reduces the matching difficulty caused by rigid and non-rigid changes in plant growth through spatial registration of plant point clouds at two time points,and then stably tracks plant organs through the proposed three-dimensional crop point cloud organ matching algorithm based on graph theory.We established a crop 3D point cloud spatio-temporal data set based on two public crop point cloud data sets,and manually labeled point-level organ growth instance labels on this data set(each organ on the point cloud of the same plant at different times has an independent label).On the self-built dataset,the proposed two-stage method performs quantitative and qualitative tracking experiments,and the organ matching accuracy reaches 82.89 %,which has a good growth tracking effect.Our method is compared with a current mainstream organ growth tracking method,and the effect is beyond the comparison method,which proves that the proposed method has a good ability to find new plant organs and withering organs.(3)The above two-stage organ tracking method is applied to the output backend of two popular plant organ segmentation deep networks,respectively,and a fully automated phenotypic analysis framework from plant point cloud organ segmentation,registration to tracking is explored.The implementation logic of the framework is as follows: Firstly,each organ instance is segmented from the whole plant point cloud through the deep learning network;then the proposed growth tracking method is used for organ registration and growth tracking.Several samples were qualitatively tested on the self-built 3D point cloud spatio-temporal data set.The test results show that good organ growth tracking results can be obtained when the segmentation results are accurate.Even in the case of serious errors in the segmentation results,such as merging multiple instance classes into one instance class,or wrongly dividing an instance class into multiple instance class fragments,our framework can still find relatively optimal organ matching,indicating that our framework has good robustness and variety adaptability.