Bayesian Multi-scale Differential Optical Flow

Eero P Simoncelli

In:
Handbook of Computer Vision and Applications
eds. B Jähne, H Haussecker, and P Geissler,
Volume II, chapter 14, pages 397-422.
Academic Press, April 1999.

The estimation of optical flow is generally assumed to be the first goal of motion processing in machine vision systems, and is also crucial for efficient representation of image sequences (e.g., MPEG). The most common methods of computing optical flow are correlation, gradient, spatiotemporal filtering, and Fourier phase or energy techniques. Correlation (usually over a local window) is by far the most prevalent technique, presumably due to a combination of intuitive directness and ease of hardware implementation. But gradient implementations are often more accurate. In addition, the gradient solution is efficient (because the solution can be computed analytically rather than via optimization), and produces sub-pixel displacement estimates. A drawback of the gradient approach is that it may only be used for small displacements, but this difficulty can be alleviated using a multi-scale coarse-to-fine algorithm. This chapter provides a practical description of a Bayesian multi-scale gradient-based optical flow estimation algorithm, based on work previously published [simoncelli-cvpr91, simoncelli-phd, Simoncelli93d].
Download Full text (ps.gz, 407k) / Full text (pdf, 510k) / EPS Online Publications