Dynamic scene analysis and applications

Summary form only given. The 4-D approach to real-time machine vision, which is based on integral spatio-temporal world models, is discussed. The method combines dynamical models for motion of and around the center of gravity with 3-D models for shape representation and the laws of perspective projection in order to arrive at recursive state estimation for objects in high-frequency image sequences. It does not require storing past frames and therefore is well suited for TV signal processing. It is an extension to image sequence understanding of well-known state estimation techniques in modern control theory. The nonunique inversion of the perspective projection is bypassed by recursive least-squares approximation exploiting a local linear relationship between image features and the internal representations of the physical 3-D state variables. The method is especially well suited for the autonomous visual guidance of vehicles, since it integrates control actuation and measurement interpretation.<>