Model-Based Hand Posture Estimation Using Monocular Camera

Abstract

This work studies the problem of model-based hand posture estimation via monocular camera. Generally, a model-based posture estimation method manipulates a 3D hand model whose posture is determined by a set of parameters. It looks for parameters that align the model with hands in images best. In this thesis, the method for building hand model from truncated quadrics, algorithms for silhouette and edge extraction are studied and implemented into an application of hand posture detection, which is to recognize and locate given postures in real images. Moreover, as an attempt to recover hand articulation, a multidimensional scaling (MDS) based approach is tried out. By MDS, posture templates are lain out in an embedding space where the intrinsic dimensions of templates appear to be recovered. The accuracy of articulation estimation is shown through experiments. And the discussion of future work for improving the current approach follows.