Dynamic Visual Motion Estimation

Abstract

Visual motion is the projection of scene movements on a visual sensor. It is a rich source of information for the analysis of a visual scene. Especially for dynamic vision systems the estimation of visual motion is important because it allows to deduce the motion of objects as well as the self-motion of the system relative to the environment. Therefore, visual motion serves as a basic information for navigation and exploration tasks, like obstacle avoidance, object tracking or visual scene decomposition into static and moving parts. Despite many years of progress, visual motion processing continues to puzzle the mind of researchers involved in understanding the principles of visual perception. Basic aspects such as measuring motions of spatially local entities have been widely studied. But what is most striking about motion processing is its temporal dynamics. This is obvious, because the environment perceived by a visual observer like a video camera or the human eye is highly dynamic. Moving objects enter and leave the field of view and also change the way they move, e.g. change the direction or speed. Hence, suitable assumptions about the dynamics of the visual scene and about the correlations between local moving entities are beneficial for the estimation of the scene motion as a whole. Probabilistic machine learning techniques have become very popular for early vision problems like binocular depth and optical flow computation. The reason for that popularity is because of the possibility to explicitly consider inherent uncertainties in the measurement processes and to incorporate prior knowledge about the state to be estimated. Along this line of argumentation, we present a general approach to visual motion estimation based on a probabilistic generative model that allows to infer visual motion from visual data. We start with a definition of visual motion and point out the basic problems that come along with visual motion estimation. Then, we summaries common ideas that can be found in different state-of-the-art optical flow estimation techniques and stress the need for taking uncertainty into account. Based on the ideas of already existing models we introduce a general Bayesian framework for dynamic optical flow estimation that comprises several different aspects for solving the optical flow estimation problem into one common approach. So far, the research on optical flow has mainly concentrated on motion estimations using the observation of two frames of an image sequence isolated in time. 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 this visual