Spatiotemporal Filters for Optical Flow

Research Overview

Filtering plays a significant role in many approaches to image sequence analysis. In particular, spatiotemporal filtering can be particularly time- and memory-consuming if done with inseparable 3d FIR filters.

My work in 1985, with Allan Jepson (University of Toronto), was one of the early examples of spatiotemporal filtering with Gabor FIR filters. We also show how one can build spatiotemporal filters with cascades of very simple FIR filters.

More recently, with Keith Langley (University College London) I concentrated some efforts on designing and implementing recursive (IIR) filters for motion analysis. The goal was to save both the storage and the computation requirements. These goals are relevant to gradient-based methods and to phase-methods. In both cases one needs both a smoothed version of the signal and its derivatives in space and time. Langley and I addressed these issues with a class of filters (band-pass and low-pass) that admit efficient ways of simultaneously smoothing images and obtaining their partial derivatives. For phase-based methods, which require even larger numbers of filters, we have shown how one might adapt the frequency tuning of the filters to both reduce the required number of filters and to improve signal-to-noise ratios and hence the robustness of velocity estimates.

Related Publications

Return to David Fleet's home page.