next up previous
Next: The Dynamical System Up: 3D Structure from 2D Previous: Filtered and Recursive Multi-Frame

The Proposed Nonlinear Recursive Framework

In the following we review and discuss the formulation proposed by Azarbayejani and Pentland [2] for a recursive recovery of 3D structure, 3D motion and camera geometry from feature correspondences over an image sequence. The emphasis here is that the Structure from Motion is cast into a dynamical system framework with important representational improvements that provide an accurate and stable implementation. In addition, the system integrates information over a complete sequence of images in a probabilistic framework. Therefore, it is robust to errors in 2D feature tracking. The pertinent issue is the parameterization of the geometric concepts and the representation which bring forth numerical advantages, real-world reliability and new functionality. For instance, we demonstrate the recursive recovery of focal length which is free to span the full range from both perspective projection to orthographic projection. The recursive framework allows online and real-time recovery of structure, motion and focal length for true flexibility and applicability to real-time problems. Finally, the approach provides a fully metric 3D recovery of structure (versus a pseudo-structure recovery). Thus, it has important applications in graphics and post-production which traditionally operate in such representations.




next up previous
Next: The Dynamical System Up: 3D Structure from 2D Previous: Filtered and Recursive Multi-Frame

1999-05-17