The monocular multi-view camera system is composed of one fisheye camera with four mirrors around it. This camera system is equivalent to five subcameras, with the same intrinsic parameters, at different positions. The greatest merit of this camera system is that the depth map could be obtained with only one shot, while it is impossible to perform the structure from motion (SfM) with one shot.
Stereo matching and 3D reconstruction has aroused much attention in the computer vision community, due to its potential application in real object modeling and human computer interaction. In this project, we systemically accomplished the stereo vision with the monocular multi-view camera system. A complete calibration procedure is proposed and obtains accurate precession. In the linear part of calibration, we show how the process in mean rotation outperforms the baseline method in robustness. By nonlinear refinement based on Bundle-Adjustment, the reprojection error is minimized by the iteration techniques. A Matlab toolbox, implementing the whole calibration method, is presented to make the calibration process convenient and automatic. With an accurate calibration result, we implemented the multi-baseline stereo in disparity space, which is rapid in speed and free from distortion. Sufficient experiments on both synthetic and real object experiment demonstrate the validation of calibration. The depth map generated in disparity space verified the efficiency of this approach. Finally, with the ICP algorithm, the 3D reconstruction is accomplished by aligning those depth maps captured from a moving camera.
"3D Reconstruction by a Moving Monocular Multi-View Camera System"