Spatio-temporal image processing for vehicular navigation
Philippe Burlina, Rama Chellappa · 1994
Several aspects of the problem of dynamic image analysis are investigated with the intended purpose of addressing certain issues commonly encountered in Intelligent Vehicle and Highway Systems and autonomous vehicular navigation. First, the class of motions with arbitrary smooth maneuvers (i.e., with polynomial translational components) are considered. For these dynamic situations, we study a family of temporal parametric descriptors, shown to be visually recoverable, which are relevant for vehicle guidance, enable a qualitative description of the trajectory, and follow straightforward dynamics if the detected model order is correct. For these parameters, methods based on brightness derivatives and methods based on feature trajectories are both proposed. Next we address inherent limitations in estimating these parameters. An integrated model-based estimation and decision theoretic approach enabling model order validation, collision detection and estimation is described. Experiments on both synthetic data and real imagery are presented to substantiate and test the modules. Other issues such as the use of log-polar retinas are considered as well. Finally, we investigate the use of special transforms for the analysis of image motion that includes a divergent motion component. Polynomial parallel translations are known to yield three dimensional power spectral densities that can be factored into two terms: the two dimensional spectrum of the stationary image and a spectral motion support. This useful result is extended to the case of arbitrary 3D translations by use of spatio-temporal Mellin Transforms. The relationship of time to collision to the resulting motion support is investigated, and integral methods for the computation of time to collision are derived.