Model-based tracking using motion from motion

The static perception of scene structure contributes in part to the perception of 3D motion, in the human visual system. In order to examine this process, a model is used to make explicit the fixed and changing properties of rigid body structure. The changing state of the model is continuously evaluated using features from a sequence of 2D images. This is also known as the model-based visual tracking problem. We seek to reproduce this ability to perceive and understand motion by building a machine vision system. This system uses multi-resolution concepts from visual information processing of biological systems. It then combines these concepts with computational schemes that express spatio-temporal transformations from a sequence of images, in the form of mathematical relations. Such a system tracks perceived motion using spatio-temporal sequences. Our work uses a motion-searching paradigm combined with a new optical flow estimation algorithm. A parallel object recognition scheme generates model pose hypotheses, until correspondence with the configuration of image features is verified. The models used are those of polyhedral objects. The spatio-temporal assumption is exploited in computing the optical flow, which when used with the object recognition scheme, allows for chosen features on the object to be tracked and used to obtain a depth estimate of the object, along the line of sight. The parallel implementation of the tracking scheme on a fine-grained data parallel machine, the Connection Machine, is an additional issue of this research, aiming at real-time realization of the tracking system