Traitement de séquences d'images aériennes en visée latérale pour la reconstruction 3D

We study the estimation of digital surface model of a 3D scene using aerial image sequences taken from a lateral viewpoint. When the motion is known, we show that multi-baseline processing leads to a better precision than stereoscopy. In this calibrated context, we use a pixel-wise multi-baseline correlation criterion, penalized with a L1-norm smoothing term, which is edge-preserving. When the motion is not precisely known, we consider motion estimation : two methods are presented and experimented. This work is illustrated by results on a synthetic sequence.