Multi-frame structure from motion using optical flow probability distributions

  • Authors:

  • Dah-Jye Lee;Paul C. Merrell;Brent E. Nelson;Zhaoyi Wei

  • Affiliations:

  • Department of Electrical and Computer Engineering, Brigham Young University, Provo, USA;Department of Electrical and Computer Engineering, Brigham Young University, Provo, USA;Department of Electrical and Computer Engineering, Brigham Young University, Provo, USA;Department of Electrical and Computer Engineering, Brigham Young University, Provo, USA

  • Venue:
  • Neurocomputing
  • Year:
  • 2009

Quantified Score

Hi-index 0.01

Visualization

Abstract

We present a novel multi-frame structure from motion algorithm in which camera motion and object structure are calculated from optical flow probability distributions instead of a single optical flow estimate at each feature point. Optical flow distributions of the selected feature points allow us to quantify the accuracy of the optical flow estimate in any direction. With this additional knowledge, a more accurate structure from motion algorithm is created which relies on this more accurate optical flow data. This novel method is designed to use the optical flow values taken from multiple frames of video or an image sequence. It is an optimal solution to the structure from motion problem with respect to a chosen norm. We will demonstrate that this new method performs significantly better than similar methods which do not use optical flow distributions or which do not use multiple frames.