Small 3-D Reconstructions from Calibrated Sequential Stereo Images

Abstract

A growing variety of emerging technologies such as Virtual Reality, Augmented Reality and Unmanned Aerial Vehicle (UAV) systems, have incredible potential in tandem with a robust and accurate approach to acquiring 3-D geometry from cameras. Camera based scene geometry acquisition through Structure from Motion (SfM) has been around since the 1840's, yet the technology is still unable to attain the reliability and accuracy necessary to satisfy modern demands. The most typical camera configuration when performing scene acquisition is a single aperture based camera that moves freely through the scene. This configuration is problematic since such a camera cannot acquire scene depth directly. As a result, new range imaging cameras that actively interrogate scenes with laser or infra-red light are being developed. This work explores the use of wide-angle aperture based cameras in a stereo configuration to improve SfM. The use of calibrated stereo cameras greatly simplifies the process and makes it significantly more robust and accurate, yet this approach appears to be relatively unexplored in literature, with the main focus being on monocular systems. The research outlined in this dissertation looks specifically at SfM from a sequence of images capturing a small to medium object. The steps undertaken for the acquisition of a 3-D reconstruction are Camera Calibration, Stereo Matching, Egomotion estimation, and 3-D model construction. A new adaptive Tsai calibration approach outperforming current state-of-the-art is introduced. A PatchMatch based guided stereo matching system is proposed to generate more accurate depth-maps. A cluster feature based egomotion estimation strategy is proposed to improve camera trajectory estimation and overall 3-D reconstruction. Finally, a CPU based Truncated Signed Distance Function (TSDF) is proposed to merge images into 3-D models. All these steps come together to form a new 3-D reconstruction algorithm that is more robust than its predecessors. Throughout this research, accuracy and robustness have been preferred to speed. While Real-Time systems are often required to compromise through approximation, the proposed system focuses on reliability over speed, thus while some effort has been made to make the system fast, it does not claim to be real-time.