Tracking through Optical Snow

Abstract

Optical snow is a natural type of image motion that results when the observer moves laterally relative to a cluttered 3D scene. An example is an observer moving past a bush or through a forest, or a stationary observer viewing falling snow. Optical snow motion is unlike standard motion models in computer vision, such as optical flow or layered motion since such models are based on spatial continuity assumptions. For optical snow, spatial continuity cannot be assumed because the motion is characterized by dense depth discontinuities. In previous work, we considered the special case of parallel optical snow. Here we generalize that model to allow for non-parallel optical snow. The new model describes a situation in which a laterally moving observer tracks an isolated moving object in an otherwise static 3D cluttered scene. We argue that despite the complexity of the motion, sufficient constraints remain that allow such an observer to navigate through the scene while tracking a moving object.