Abstract
In this thesis, we present a 3D reconstruction system that integrates multi-view RGBD images acquired with Kinect for an object sitting on a turntable. In the proposed system, we first segment the object from images by using a simple background model with depth and produce the 3D point cloud from the RGBD image in a single view. Next, we compute feature correspondences between each pair of successive frames by SURF, and remove the false feature correspondences by applying RANSAC affine matching for each pair of adjacent views. Then, we use all the verified 3D feature correspondences to determine the geometric transformation that transforms the 3D coordinates from the corresponding camera coordinates to a unified world coordinate centered at the turntable. Because of the cumulative error of rotation angles and other environmental variables, we propose a two-step refinement process by using the LM Optimization. Finally, we propose a novel point set simplification algorithm to simplify the integrated point dataset that reduces the density of 3D points in the overlapped regions. Experimental results are given to demonstrate superior 3D reconstruction results by using the proposed method on both real data and simulation data.