Monocular 3D Hand Tracking Using Particle Filter with Partitioned Sampling

Abstract

Three dimensional model-based hand tracking always refers to search optimal hand state in high dimensional state space due to the articulated hand model has many Degrees of Freedom (DOFs). This study firstly by introducing dynamic hand constraints into finger’s movement angle to reduce local tracking DOFs. Secondly, standard Particle Filter (PF) as the tracking tool needs a huge number of samples to approximate the state’s posterior distribution, so it is very time consuming. To reduce the needed samples’ number partitioned sampling is used to divide the high dimensional state into a set of low partitions, and a tracking algorithm is built based on these partitions. Thirdly, develop a new method to extract the hand silhouette from frame image by combining Bayes classifier and background subtraction. The extracted hand silhouette is subsequently used to build an observation likelihood model.