Scalability in Human Shape Analysis

Abstract

This paper proposes a new approach for the human motion analysis. The main contribution comes from the proposed representation of the human body. Most of already existing systems are based on a model. When this one is a priori known, it may not evolve automatically according to user needs, or to the detail level that is actually possible to extract, or to restrictions due to the processing time. In order to propose a more flexible system, a hierarchical representation of the human body is implemented. It aims at providing a multi-resolution description and results at different levels of accuracy. An explanation about the model construction and the method used to map it onto features extracted from an image sequence are presented. Relations between the different body limbs and some physical constraints are then integrated. The transition from a model level to the next one is also explained and results on frames coming from a video sequence give an illustration of the proposed strategy