A New View at Differential and Tensor-Based Motion Estimation Schemes

This paper contributes to the recent attempts to generalize classical differential and tensor-based motion estimation methods; it provides means for a better adaptation to the statistical structure of signal and noise. We show that conventional differential constraint equations often used for motion analysis capture the essence of translational motion only partially, and we propose more expressive formulations using higher order derivatives, finally leading to steerable nulling filters. Such filters consist of a ’static’ prefilter and a steerable filter; we show how the prefilter can be optimized for given autocovariance structure of signal and noise, and how the steerable filter can be related to classical tensor-based approaches, leading to a constrained eigensystem problem.