Research On 3D Reconstruction Of Rice Plant

Agriculture is the most important industry in our country,it is the foundation of economy construction and development.Rice is the main food crop in China.Genome research is critical for crop breeding and food security.The important target of modern rice breeding and cultivation is to select varieties with high-yield,high tolerance and high nutrient utilization.It is significant to provide geneticists and breeders with a set of automatic,efficient and accurate methods for phenotype measuring and analysis.In recent years,computer vision has played an important role in plant phenotyping.A common routine which is adopted in numerous researches is to extract traits of given samples through image processing and following with mathematical modeling analysis.Image-based phenotyping has gone through a long way,but only a few researches considered problems on the basis of 3D plant models.In general,more comprehensive information can be extracted from 3D models than 2D images.In this paper,we proposed a method to batch reconstruct the 3D point clouds of potted rice based on multi view images.The proposed work could be the foundation of subsequent 3D phenotyping research.The paper summarized the merits,demerits and applicability of existing 3D reconstruction methods which can be divided into hardware-based and multi-view image-based ways.Shape-from-silhouette based on the concept of visual hull has been adopted to reconstruct 3D point clouds of rice plants.The definition of visual hull was elaborated,error source and optimal results of reconstruction were analyzed as well.The proposed algorithm was performed on the basis of an improved calibration method.Large amount of rice plants can be reconstructed with just one calibration result.Using GPU optimization,the algorithm can reconstruct the 3D point cloud with RGB information of a rice plant in about 8 to 10 minutes.Plant heights,macro axis length,the area and length-width ratio of minimum enclosing rectangle in top view,volume of bounding box,canopy density and leaf number were extracted from 100 reconstructed rice 3D point clouds.The mean absolute percentage error of plant height,macro axis length and leaf number were evaluated by comparing with ground truth values.The mean absolute percentage error of plant height and macro axis length was lower than 4.50% and 10.47%,respectively,and of leaf number could achieve 11.58%.Besides,we compared the performance of our algorithm and classic structure-from motion method.The results demonstrated that our algorithm is faster,more accurate and more suitable for rice plants of different kinds and growth periods.Our work provides a solid foundation for 3D rice phenotyping.