Plants are three-dimensional (3D) organisms whose morphology is complex and highly variable between individuals, making them challenging to fully characterize with standard 2D image-based methods. Owing to the lack of depth information in images, 3D data has attractive for analysis of plants’ organs and traits. 3D data holds promise for measuring the geometrical features of plants but 3D modeling and processing remain problematic in plant science. This thesis introduces a high-throughput yet accurate pipeline for objective measurement of plants’ attributes directly in 3D space. The pipeline consists of two main phases including i) 3D modelling and ii) analysis of the 3D models in the point cloud domain. In this study, we have adopted Brassica and wheat as the case studies and measured their different attributes, including the number of branches and their coordinates, branch angles and the diameter/thickness of different organs in different locations in Brassica; and the number of ears and their length in wheat. The texture pattern of plant data, here Brassica and wheat, is thin, repetitive and complex, which makes them challenging to phenotype with existing methods. The goal of the proposed pipeline is to fully-automatically measure all the attributes and therefore need fast, reliable and accurate techniques for this purpose. In the first phase of the proposed framework, the 3D models of plants are reconstructed from multiview images. To this end, we first set up two in-cabin cameras in the National Plant Phenomics Centre (NPPC) for capturing 60-74 images per plant from different perspectives. The stereo images are then fed into structure-from-motion (SfM) for producing 3D models. In the SfM pipeline, we propose a fast feature detector (FFD) and an interwoven texture-based feature descriptor (InterTex) for improving the contents of 3D models. Both FFD and InterTex are fast techniques that are developed to cope with texture-less, repetitive and complex structures of plant imagery. The second phase measures the attributes of biological organisms in the 3D point cloud domain. Direct analysis of 3D models in the point cloud domain, however, is not straightforward, owing to its discrete nature, imaging noise and cluttered background. We have introduced a novel method for accurately phenotyping the Brassica point clouds using random sample consensus (RANSAC). It is shown that the proposed technique can approximate diameters and angles of organs accurately even in the presence of noise and irrelevant data. The developed framework has been applied to point cloud models of Brassica samples, captured at different intervals. The results show the success of the proposed pipeline in dealing with irregular point clouds of Brassica contaminated with noisy and irrelevant outliers. We have also demonstrated that wheat point clouds can be segmented in the 3D point cloud domain. To this end, a novel deep-learning network is proposed that efficiently handles highly complex point clouds of wheat. The proposed network is the first study to segment and analyse 3D organs of wheat plants in the point cloud domain via deep learning. The proposed framework has been validated on wheat point clouds, captured at different times during the growth cycle. The results indicate that our deep learning method is robust and can accommodate irregular point clouds that are noisy and contain irrelevant outliers
3D Object Modelling and Its Application for Trait Measurement of Plants
Ghahremani Boozandani, M. (Author). 2021
Student thesis: Doctoral Thesis › Doctor of Philosophy