Phenotyping involves the measurement, ideally objectively, of characteristics or traits, usually in the context of living organisms, including plants. Traditionally, this is limited to either tedious and sparse manual measurements, often acquired destructively, or coarse image-based 2-D measurements. A dynamic model of 3-D architecture through developmental time could capture useful geometric characteristics representing phenotypic information on morphology and also record the adaptive response to environmental conditions. In recent years, many emerging 3-D imaging technologies (based on laser scanners, structured light, multi-view stereo, etc.) provide the potential to capture quantitatively morphological features and have been proposed as non-destructive phenotyping alternatives to current cost-intensive commercial phenotyping platforms. However, available 3-D resolutions are limited in various ways. For example, they may focus on a specific organ (e.g. leaf or stem) or tend to be qualitative rather than providing quantitative information and often lack any estimation of accuracy. This thesis investigates the existing methods and aim to build a cost-effective accurate 3-D reconstruction framework that could cope with a diversity of plant forms and sizes, while using equipment available to most biology labs. I firstly developed a multi-view image acquisition method to obtain highresolution 2-D image sequences, using low-cost turntable and a consumer-grade digital camera. To deal with the self-occlusion problem, the proposed method is based on short baseline and multiple viewpoint photography, and usually acquired approximate 60-120 images. To improve the flexibility and save time of data acquisition, the camera is equipped with a variable focal length lens and does not need extra calibration processing. I proposed an efficient Structure-From-Motion method to estimate cameras' parameters and poses from the uncalibrated images of the plant. Finally, I developed an accurate multi-view stereo 3-D reconstruction method to yield a dense and detailed 3-D point cloud of plant. The method takes both accuracy and efficiency into account, and therefore is faster than other state-of-the-art methods. The accuracy of the proposed method was evaluated with limited ground truths as well as the measurement deviations of test specimens. The experimental results show that the proposed methods are highly reliable for phenotyping of various plants during their entire growth cycles. This promising generalized 3-D imaging technique has the potential to be implemented in automated procedure.
Date of Award | 29 Jul 2016 |
---|
Original language | English |
---|
Awarding Institution | |
---|
Supervisor | Yonghuai Liu (Supervisor) & Chris Price (Supervisor) |
---|
Cost-Effective Accurate 3-D Reconstruction Based on Multi-View Images for Plant Phenotyping
Lou, L. (Author). 29 Jul 2016
Student thesis: Doctoral Thesis › Doctor of Philosophy