Learning-based tracking of complex non-rigid motion

This paper describes a novel method for tracking complex non-rigid motions by learning the intrinsic object structure. The approach builds on and extends the studies on non-linear dimensionality reduction for object representation, object dynamics modeling and particle filter style tracking. First, the dimensionality reduction and density estimation algorithm is derived for unsupervised learning of object intrinsic representation, and the obtained non-rigid part of object state reduces even to 2-3 dimensions. Secondly the dynamical model is derived and trained based on this intrinsic representation. Thirdly the learned intrinsic object structure is integrated into a particle filter style tracker. It is shown that this intrinsic object representation has some interesting properties and based on which the newly derived dynamical model makes particle filter style tracker more robust and reliable. Extensive experiments are done on the tracking of challenging non-rigid motions such as fish twisting with self-occlusion, large inter-frame lip motion and facial expressions with 'global head rotation. Quantitative results are given to make comparisons between the newly proposed tracker and the existing tracker. The proposed method also has the potential to solve other type of tracking problems.