Executive Summary : | This project aims at developing techniques for three dimensional reconstruction of plants and trees targeting different types of applications in smart farming and remote sensing for vegetation analysis. Although a large body of work has been performed on 3D reconstruction of regular objects, pairwise/multiview registration of plant point cloud data has not been fully explored. Due to the complex geometrical structure, self occlusion, and deformable nature of plants, the registration problem gets extremely challenging in practical scenarios. This project will be built in two phases in two years. In the first phase of the work, the aim will be to perform 4D reconstruction of spatio-temporal sequence of growing plants by explicit consideration of deformation, as well as track individual organs of the plant throughout the temporal sequence. The plant will be scanned from multiple views using a Time of Flight Camera in each time frame, the overlapping views will be registered to form a full 3D model of the plant, and then perform sequential registration to track individual organs. In addition, the surface area, volume, and height of the plant in each time frame will be automatically computed. This framework can be useful for crop/plant growth study and experiment with new crop varieties in a fully automatic manner as the part of a robotic system for organ tracking or to monitor the growth of plants. In the next phase of the work, the plan will be to perform reconstruction of large outdoor trees and automatically extract geometric attributes for vegetation monitoring in remote sensing and forestry applications. A LiDAR camera (set on a tripod or drone depending on the size of the tree) will be used to capture point cloud data from multiple views, which will be merged together to reconstruct the full 3D structure of the tree. Since the tree may “jitter” by wind during the scanning, this will be considered during the reconstruction process, along with the occlusion factor. From the registered point cloud, surface reconstruction will be performed to build a polygonal mesh model. From the surface mesh model, several parameters will be computed including tree height, surface area, volume, diameter of the main trunk, which will be useful for automatic estimation of biomass, wood content, growth, etc. Finally a geometric model of the tree will be reconstructed with generalized cylinders, from which it will be possible to automatically extract different geometrical attributes like length of the main stem, number of main/sub branches, length and angle of each branch, distance between two branches, etc. Automatic answering of these types of questions from the “digital tree” will be useful to vegetation monitoring in different seasons for forestry and remote sensing applications. |