People action recognition in image sequences using a 3D articulated object

This article deals with recognition of human action using motion tracking of articulated objects based on its 3D geometric modeling. We minimize a measure of difference between views of a real scene and the corresponding 3D ones. The minimization of the function uses colorimetric information and allows us to automatically estimate the position and the posture of the subject in the scene. We used an hybrid genetic algorithm to optimize the 3D model posture. The matching of views acquired by several video cameras displaced in the real scene and the synthesis views allows inferring the parameters that check the model. Afterward, the movement parameters obtain we will identify the people action using a fuzzy classification.