A Divide-and-Conquer Approach for Visual Odometry with Minimally-Overlapped Multi-Camera Setup

Pose estimation of a moving camera rig from the images alone has been investigated by the computer vision community for decades, because the location and direction information of the cameras are the basis for more advanced applications, such as 3D reconstruction and Simultaneous Localization and Mapping (SLAM). Visual Odometry (VO) is the accumulation of the relative pose estimation while the camera rig moves. There are some visual odometry methods for mono view, stereo, omnidirectional and multi-cameras that require additional sensor input(odometry, compass, e.g,) and/or synchronized cameras. However, in our Virtual Exercise Environment (VEE) system, and other low-cost multi-camera setups, none of the above methods can be applied. The Bundle Adjustment(BA) is the general approach for a non-regular case like the VEE system. BA puts all the known variables in one huge matrix and solves the unknowns at once with Levenberg-Marquardt iterations. Thus, the BA is computationally expensive in nature with the complexity of O(n3), and sometimes infeasible. In this thesis, I propose a `divide and conquer\u27 approach that generates additional observations from consecutive images in neighboring camera pairs. I show this approach to solve the critical condition and drastically speed up pose estimation when compared to BA. The performance with different conditions and sub-algorithms are also tested and discussed