Motion tracking on the spatiotemporal surface

Abstract

The spatiotemporal (ST) surface has been shown to be a useful representation of projected scene dynamics. The authors previous use of this representation has focused on geometric recovery of scene static structure from the analysis of relative motions on the moving image plane. That earlier work (Int. J. of Comput. Vis., vol.2, no.1, p.51-72 (1989); p.33-50, (1989); vol.1, no.1, p.7-55 (1987)), exploited the implicit partitioning of motions along epipolar lines to enable search-free feature tracking and position estimation. The ST manifolds provide explicit information about feature 3D contiguity, and their use leads to the recovery of feature 3D position, object 3D contours, and scene 3D surfaces. The authors have turned their attention to the task of interpretating non-static scenes, and track and estimate motions of independently moving objects and background by their appearance and behavior on the ST surface. Selecting the most reliable and discriminating information in the scene, the system demonstrates robust feature tracking over a large range of feature sizes and velocities. When coupled with the more mature epipolar-plane image analysis system, this motion analysis capability will enable camera solving, dynamics tracking, and scene reconstruction within a unified framework.<>