Region-Based Tracking in an Image Sequence

1 I n t r o d u c t i o n Digitized time-ordered image sequences provide an actually rich support to analyze and interpret temporal events in a scene. Obviously the interpretation of dynamic scenes has to rely somehow on the analysis of displacements perceived in the image plane. During the 80's, most of the works have focused on the two-frame problem, that is recovering the structure and motion of the objects present in the scene either from the opticM flow field derived between time t and time t -t1, or from the matching of distinguished features (points, contour segments, ...) previously extracted from two successive images. Both approaches usually suffer from different shortcomings, like intrinsic ambiguities, and above all numerical instability in case of noisy data. It is obvious that performance can be improved by considering a more distant time interval between the two considered frames (by analogy with an appropriate stereo baseline). But matching problems become then overwhelming. Therefore, an attractive solution is to take into account more than two frames and to perform tracking over time using recursive temporal filtering [1]. Tracking thus represents one of the central issues in dynamic scene analysis. First investigations were concerned with tracking of points, [2], and contour segments, [3, 4]. However the use of vertices or edges lead to a sparse set of trajectories and can make the procedure sensitive to occlusion. The interpretation process requires to group these features into consistent entities. This task can be more easily achieved when working with a limited class of a priori known objects [5]. It appears that the ability of directly tracking complete and coherent entities should enable to more efficiently solve for occlusion problems, and also should make the further scene interpretation step easier. This paper addresses this issue. Solving it requires to deal with a dense spatio-temporal information. We have developed a new tracking method which takes into account regions as features and relies on 2D motion models. * This work is supported by MRT (French Ministry of Research and Technology) in the context of the EUREKA European project PROMETHEUS, under PSA-contract VY/85241753/14/Z10.