Multi-image gradient-based algorithms for motion estimation
Samson J. Timoner · Optical Engineering · 2001
Image registration algorithms based on gradient methods provide quantitative motion measurements from sequences of video im- ages. Although such measurements can be degraded by image noise, larger degradations typically result from systematic bias in the algorithms that is present even if the images are noise-free. To improve the accu- racy of motion measurements, we develop a new class of multi-image algorithms based on multidimensional digital filters. The new algorithms provide better estimates of spatial and temporal gradients and also com- pensate for motion blur caused by the nonzero acquisition time of the imager. We optimize filters to measure arbitrary motions, and we illus- trate the results when those filters are used to estimate constant velocity movements. We also show results for filters that are optimized for har- monic analysis of periodic motions. Using these algorithms, systematic bias in the amplitude of sinusoidal motion is less than 0.001 pixels for motions smaller than 1 pixel in amplitude. This represents a hundredfold decrease in bias compared to existing methods. © 2001 Society of Photo-