Regresar

Articulated Human Motion Tracking with Online Appearance Learning

Abstract:

Automatic tracking of the articulations of human from a video sequence is a difficult task due to complex motions of the limbs, dynamic background, and varieties of poses. These challenges make it difficult to train a generative motion and appearance model to be used in different scenarios. In our work, we employ particle swarm optimization framework to avoid the need of motion model. Particularly, we propose a novel appearance learning strategy to learn the appearance of each body part in real time. Furthermore, we also propose an appearance model to represent the shape of each body part. Samples from UIUC dataset had been used in experiments. The results had shown that our method performed well on complex activities without motion model and online appearance training. It also showed the robustness of our method to recover from tracking failure in an occluded video. © Springer-Verlag Berlin Heidelberg 2013.