Thesis abstract:

«p»The primary objective of this project is to develop new methods for building accurate and consistent 3D models of structures in urban environments such as buildings, trees, and power lines that are difficult to reconstruct using traditional photogrammetry methods from multi-angle imagery. This involves incorporating semantic information and generating a streamlined representation suitable for export as a CAD-style model, as opposed to a textured triangular mesh. The potential key points includes:1. Construct the 3D photo-textured mesh of the scenes from the multi-angle aerial photography. The images may be captured by the high-resolution cameras.2. Apply 3D-data-based deep learning algorithm to process the constructed 3D phototextured mesh, identify and perform semantics segmentation geometric objects (e.g. buildings, trees, power lines) with distinctive features (e.g. RGB color, geometry and normal vector).3. use supervised deep learning methods, incorporating the semantic segmentation of a urban object to infer a better structures.4. Ultimately, develop an algorithm for 3D reconstruction of complex urban objects that is more accurate than traditional photogrammetry techniques.«/p»