So far, the research on optical flow has mainly concentrated on motion estimations using the observation of a small number of temporal consecutive frames of an image sequence. The dynamics of the flow field evolution is mostly neglected. Our main concern is to stress that visual motion is a dynamic feature of an image input stream and the more visual data has been observed the more precise and detailed we can estimate and predict the motion contained in the visual data. In this paper, we present a probabilistic dynamical system that is suitable to recurrently infer visual motion. The assumed flow dynamics fuses spatial smoothness constraints and smoothness constraints along time and scale. We propose a certain class of transition probability functions which satisfy a probability mixture model and allow for an efficient approximate inference based on Belief Propagation. We arrive at a compact and general algorithm for optical flow filtering and realize one instance using factored Gaussian belief representations.
展开▼
