|M.Sc Student||Hirsch Yogev Mika|
|Subject||Skeletonization of 3D point clouds for object analysis and|
|Department||Department of Civil and Environmental Engineering||Supervisors||Professor Joshua S. Greenfeld|
|Professor Sagi Filin|
|Full Thesis text|
Modern agriculture is in need to develop new plant breeding approaches in order to cope with the challenges that climate change, population growth and the aspiration for greener and more competitive agriculture pose. These approaches should improve and accelerate plant breeding and produce ones with favorable genotypes that consume fewer resources. An important part of many such approaches is phenotyping - the process of collecting the plant’s traits such as morphology, development and biochemical or physiological properties. These measured traits are crucial for learning the influence of different genotypes and environmental factors on the plant's development. Despite the increased importance of phenotyping, the adoption of new techniques that can replace century old manual ones are still lacking. New technologies such as noninvasive imaging and image analysis can offer major improvement to the extraction of plant physical properties.
Until recently, image based phenotypic was limited to extraction of 2D information which useful as it is, does not deliver a proper set characterizing feature. The emergence of new imaging capabilities such as laser scanners, depth cameras, and stereo vision as well as advances in image processing algorithms provide new capabilities of capturing detailed geometrical data of plants as 3D point clouds. Extracting higher-level features from 3D point clouds is a mandatory step towards identification of phenotypic traits.
The aim of the present thesis is to develop means for the decomposition of 3D plant related point clouds to their elementary parts in order to facilitate phenotyping. In this regard, it has two main objectives: the first is to develop a segmentation method based on little prior knowledge that is data driven rather than model driven and is able to identify and remove noise and systematic errors and handle incomplete data caused by occlusions or shadowing. The second is to be able to generate a structural description of the plant, starting with the extraction of the leaves contour as a means to create a skeleton of the plant.
The study shows the application of the proposed method on a variety of plant models and handles different levels of noise and discontinuities. Results show that the proposed method provides an efficient and reliable segmentation and shape extraction that can be used as the basis for an improved phenotyping system.