Human Body Tracking by Monocular Vision

Abstract

This article describes a tracking method of 3D articulated complex objects (for example, the human body), from a monocular sequence of perspective images. These objects and their associated articulations must be modelled. The principle of the method is based on the interpretation of image features as the 3D perspective projections points of the object model and an iterative Levenberg-Marquardt process to compute the model pose in accordance with the analysed image. This attitude is filtered (Kalman filter) to predict the model pose relative to the following image of the sequence. The image features are extracted locally according to the computed prediction. Tracking experiments, illustrated in this article by a cycling sequence, have been conducted to prove the validity of the approach.