田间作物群体三维点云柱体空间分割方法

农田作物群体表型信息对于研究作物内部基因改变和培育优良品种具有重要意义.为实现田间作物群体点云数据中单个植株对象的完整提取与分割,以便于更高效地完成作物个体表型参数的自动测量,该研究提出一种田间作物柱体空间聚类分割方法.利用三维激光扫描仪获取田间油菜、玉米和棉花的三维点云数据,基于HSI(Hue-Saturation-...

Cylinder space segmentation method for field crop population using 3D point cloud

A new phenotype of crop population depends mainly on the internal genetic change of plants with environment, thereby determining new varieties of crops in farmland. A three-dimensional (3D) laser scanning technology can provide a rapid acquisition for the accurate phenotypic data of crops, compared with some traditional time-consuming and destructive measurements. However, field high-throughput phenotypic acquisition is still a major bottleneck limiting crop improvement and precision agriculture. It is also necessary to automatically acquire phenotypic traits throughout the growth cycle of crops and further to obtain target parameters with high accuracy. In this study, a cylinder space clustering segmentation was proposed for a highly efficient extraction on complete phenotypic parameters of a single plant in field crop population using a 3D point cloud. Field experiments were carried out at the Huazhong Agricultural University in Wuhan City, Hubei Province of China in 2019. Flowering rapeseed, seedling corn, and flowering cotton were selected as the research objects. The experimental procedure was: 1)A 3D laser scanner(FARO FocusS SeriesS 70) was used to collect high-precision point cloud data of field corn, rapeseed and cotton. Multiple sites were set around the experimental field for high accuracy information about the target. The measuring sites of rapeseed field were laid in the four corners and the middle of the long side of a sample plot. Four corners of a sample plot were selected to measure in corn and cotton field. Two groups of point cloud data were collected at different heights in the same measuring site. Each position was scanned once, and each scanning took 10min. At least 3 target balls were placed in the test area as the registration basis, thereby preparing for the registration of point cloud data collected by subsequent test stations.2) The crop target was then extracted from the massive point cloud, including registration, denoising, data extraction, and simplification. The point cloud registration was completed using a target ball. The noise points were eliminated using dark scan point, outlier, and edge artifact filter. A Hue Saturation Intensity(HSI) color model was utilized to extract crop group target, according to the difference between crop and soil color. Curvature sampling was selected to realize point cloud simplification. 3)A pass-through filter was used to extract the stem point clouds at a certain height, whereas, the leaf point clouds were removed according to the difference of normal vectors. Conditional Euclidian distance was selected to extract the cluster center point of each plant using stem point cloud. A cylinder spatial model with the center point was also established to segment the point cloud of each plant. The column radius and height were set according to the row spacing and growth of specific crops in farmland. The segmentation accuracies of corn, rapeseed, and cotton were 90.12%, 96.63%, and 100%, respectively. The accuracy increased by 36.42%, 61.80%, and 82.69 percentage points, respectively, while the running time shortened to to 9.98%, 16.40% and 9.04%, compared with the conventional clustering segmentation. As such, better applicability, feasibility, and universality were achieved to effectively segment and extract all three types of individual plants from crops in dense fields, compared with previous region growth. Therefore, the segmentation and recognition of a single plant in crop population can provide a promising technical approach for the accurate, rapid, and non-destructive measurement of phenotypic information of individual crop in the field.