基于多视角图像的作物果实三维表型重建  

Three-Dimensional Phenotypic Reconstruction of Crop Fruit Based on Multi-View Image

作　　者：杨会君[1,4] 王瑞萍 王增莹 王昕 Yang Huijun;Wang Ruiping;Wang Zengying;Wang Xin(College of Information Engineering,Northwest A&F University,Yangling 712100,China;Department of Foreign Languages,Northwest A&F University,Yangling 712100,China;Three Squirrels Nanjing R&D Innovation Center,Nanjing 210019,China;Key Laboratory of Agricultural Internet of Things,Ministry of Agriculture and Rural Affairs,Yangling 712100,China)

机构地区：[1]西北农林科技大学信息学院,陕西杨凌712100 [2]西北农林科技大学外语系,陕西杨凌712100 [3]三只松鼠南京研发创新中心,江苏南京210019 [4]农业农村部物联网重点实验室,陕西杨凌712100

出　　处：《南京师大学报（自然科学版）》2021年第2期92-103,共12页Journal of Nanjing Normal University(Natural Science Edition)

基　　金：陕西省重点研发计划项目(2021NY-179,2019ZDLNY07-02-01,2020NY-205);大学生创新创业训练计划项目(S202010712238、S202010712063、X202010712373).

摘　　要：针对基于激光扫描设备获取点云存在操作复杂、成本高、难以被普及等问题,本文研究了基于普通图像的复杂背景中作物果实三维表型重建.我们建立了集SFM算法、PMVS算法以及半自动化去噪方法的优势为一体的三维重建架构.以一组多视角目标作物果实二维图片为输入源,首先基于SIFT算子的比例和旋转不变性参数,提取多幅二维图像特征信息.其次,结合FLANN算法实现不同角度的数据匹配,并提出了基于二维图像关键点和相机参数等信息的稀疏点云快速生成方法.然后,基于PMVS初始特征匹配的种子面片提取、扩散获取密集面片,进一步利用可见性约束过滤不正确匹配导致的错误面片,以实现复杂果实点云模型生成.最后,我们提出了交互式选择和滤波器相结合的、半自动化的果实表型离群点去除方法,解决了作物果实模型的准确重建问题.结果表明,本文的方法能有效解决复杂实验环境中果实表型数据的低成本、准确、方便快捷获取问题.Obtaining point cloud based on laser scanning equipment has many problems such as complicated operation,high cost,and difficulty in popularization.Therefore,this paper studies the three-dimensional phenotypic reconstruction of crop fruit in complex background based on common images.We have built a 3D reconstruction architecture integrating the advantages of SFM algorithm,PMVS algorithm and semi-automatic denoising method.Taking a set of multi-view two-dimensional images of target crop fruits as input sources,we first extracted the feature information of multiple two-dimensional images based on scale of the SIFT operator and rotation invariance parameters.After that,we combined the FLANN algorithm to achieve data matching from different angles.Furthermore,we proposed a fast method of generating sparse point clouds based on the information of key points of 2D pictures and camera parameters.Then,the dense patches are obtained by seed patch extraction and diffusion based on the initial feature matching of PMVS algorithm.We further used visibility constraints to filter out the wrong patches caused by incorrect matching,and realized the generation of complex point cloud model.Finally,we proposed a semi-automatic method to remove outliers from fruit phenotypic point cloud by combining interactive selection and filter,which solved the problem of accurate reconstruction of crop fruit model.The results show that this method in this paper can effectively solve the problem of low-cost,accurate,convenient and fast acquisition of fruit phenotype data in complex experimental environment.

关 键 词：三维果实表型 稀疏点云算法 稠密点云算法 去噪 

分 类 号：O4-39[理学—物理] TP37[自动化与计算机技术—计算机系统结构]