Modeling View and Posture Manifolds for Tracking.pdfVIP

Modeling View and Posture Manifolds for Tracking Chan-Su Lee and Ahmed Elgammal Department of Computer Science Rutgers University Piscataway, NJ, 08854, USA {chansu,elgammal}@cs.rutgers.edu Abstract In this paper we consider modeling data lying on multi- ple continuous manifolds. In particular, we model the shape manifold of a person performing a motion observed from dif- ferent view points along a view circle at fixed camera height. We introduce a model that ties together the body configura- tion (kinematics) manifold and the visual manifold (observa- tions) in a way that facilitates tracking the 3D configuration with continuous relative view variability. The model exploits the low dimensionality nature of both the body configuration manifold and the view manifold where each of them are repre- sented separately. 1. Introduction Despite the high dimensionality of the human body config- uration space, many human activities lie intrinsically on low dimensional manifolds. Exploiting such property is essen- tial to constrain the solution space for many problems such as tracking, posture estimation, and activity recognition. Re- cently, there have been increasing interests in learning low dimensional representations of the body configuration mani- folds, as in [3, 15, 18, 2, 8], for tracking and posture estima- tion. The goal of this paper is to model the visual manifold of an articulated object observed from different view points. Mod- eling visual manifolds is a challenging task. In particular, we focus on modeling human motion as observed from dif- ferent view points. Traditionally, generative model-based ap- proaches have been used for tracking and posture estimation, where a 3D body model and a camera model are used, and the problem is formulated as a search problem in high dimen- sional spaces (articulated body configuration and geometric transformation), e.g. [13]. Alternatively, discriminative map- pings have also been introduced, e.g. [10, 1]. The model intro- d