A recursive approach to the design of adjustable linear models for complex motion analysis

Parametric models are widely used in motion analysis. Traditionally, affine or learned models are adopted. Here, we propose the use of a set of linear models that dynamically adjust their properties to approximate first-order structures in noisy optic flow fields. Each model is generated by the evolution of a recursive network that can be used as a process equation of a multiple model Kalman Filter. The presence of a model is checked by computing the consistence between the observations (data) and the predictions (model). In each image region, for each model, a probability value can be computed, on which to base motion analysis. Experimental results on multiple motion detection problems and facial expressions analysis validate the approach. The algebraic transformations relating our linear descriptors with the traditional affine models are discussed