Context-Sensitive Recurrent Filters for Visual Motion Analysis

Real-world motion field patterns contain intrinsic statistic properties that allow us to define Motion Primitives as groups of pixels that share the same motion property. By checking the presence of such primitives in optic flow field we can make its interpretation more conident. We propose a context-sensitive recurrent ilter capable of evidencing motion primitives corresponding to 1st-order flow descriptors. A higher-level motion description emerges from a noisy flow as a solution of an iterative process of spatially interacting nodes that correlates the properties of the visual context with those of a structural model of the motion primitive. A systematic comparison of the proposed approach with respect to several standard methods have been performed through benchmarking sequences. Moreover, real-world situations have been used to validate the approach. By proper speciication of the interconnection scheme, the approach can be straightforwardly extended to model any type of multimodal spatio-temporal relationships (i.e., multimodal spatiotemporal context)