基于立体视觉的玉米雄穗三维信息提取

摘要

The phenotypic information of maize tassel has important reference significance for maize breeding. In this study, 17 maize tassel samples were collected at the harvest stage for the purpose of obtaining the three-dimensional phenotypic information of tassels automatically. The samples were collected by the experimenter from the field and then the photogrammetry was taken indoors with them. The overlapping rate of each image is greater than 60%, and finally 72 multi-view photos were obtained for each tassel sample. Acquired tassel multi-view photos were used for three-dimensional modeling processing in the VisualSFM software. Since the reconstructed tassel three-dimensional model includes tassels and background plates, there are a large number of noise points. Therefore, it cannot be directly used for extracting phenotypic information. In this study, the three-dimensional model of tassel was first preprocessed, including point cloud thinning, noise removal, background plate separation and other steps. Then, the three-dimensional model results obtained by program were used to compute tassel phenotype information. For the obtained tassel samples, the number of branches, the main axis length, the maximum diameter of the main axis, the minimum diameter of the main axis, the maximum canopy diameter, the maximum canopy height and other information were manually measured. The artificially acquired phenotypic information is used as a verification dataset for the results of the phenotypic information calculated with the program. The number of tassel branches, tassel volume, main axis information (length, maximum diameter, minimum diameter of main axis), information of maximum canopy (diameter and height), total projected area and other parameters of information were calculated using computer methods for statistics. The statistics of the number of branches use density-based clustering method. The algorithm divides a region with sufficient density into clusters and finds arbitrarily shaped clusters in the noisy spatial database (the largest set of points with density connected), which takes full advantage of the spatial information of the three-dimensional point cloud. Compared with previous studies, this method is based on the characteristics of density and distance clustering; the algorithm can find clusters of arbitrary shape, providing a new idea for statistics of branch numbers, and has better operability. With the statistic method of tassel volume and vertical projection area, a statistical method based on convex hull is proposed. Tassel point cloud is divided into 30 layers from top to bottom, and the envelope convex surface of each point cloud is obtained. The convex surface consists of a Delaunay triangulation network. The area of each convex hull is calculated, and then multiplied by the distance between the 2 layers. After accumulating sum, it is the outer volume of the tassel that can represent the structure information of tassel more truly. It is more in line with the demand for phenotypic information of maize breeding researchers. In addition, the study also proposed the definition of tassel-related phenotypic parameters (tassel spatial aggregation, tassel plane aggregation, tassel head-to-stem ratio, tassel canopy height ratio, main axis variation coefficient, and tassel center of gravity). The final experimental results showed that the maximum absolute error of branching number was 2, the RMSE (root mean square error) was 1.03, and the nRMSE (normalized root mean square error) was 0.05. The R2 of the major axis length, the maximum/minimum diameter of the major axis, the maximum crown height and the maximum crown diameter were 0.99, 0.82, 0.83, 0.97 and 0.93, respectively; the RMSE was 0.2281, 0.2194, 0.164, 4.03, and 3.252 cm, respectively. All reach extremely significant levels. The results of the study reached the accuracy criteria for tassel phenotyping. The results provide reference for high-throughput automatic acquisition of phenotypic information and give a new method for breeding based on the phenotype information of tassel.%玉米雄穗的表型信息对玉米育种研究具有重要的参考意义.该研究以自动获取玉米雄穗三维表型信息为目的.通过对雄穗样本进行多视角摄影处理来重建其三维模型.对重建的三维点云数据运用基于密度聚类的方法统计其分枝数信息,运用Delaunays三角网方法计算其外包络体积信息,并基于点云信息对雄穗主轴和最大穗冠的结构参数进行计算,同时定义了相关表型参数.用实测结果验证计算结果:分枝数统计结果的最大绝对误差为2,RMSE(root mean square error)为1.03,nRMSE(normalized root mean square error)为0.05;主轴长度,主轴最大/最小直径,最大穗冠高度和最大穗冠直径的R2分别为0.99,0.82,0.83,0.97和0.93,均达到极显著相关水平.研究提出的相关表型参数和其提取方法在育种研究中具有应用潜力,为田间高通量雄穗信息的快速提取提供了参考.