Background. The algorithms for tracking generic human motion should be able to cope with the high-dimensional state space as well as to recover complex postures with various motion types and styles. Many approaches have been proposed to address these problems [1, 3, 5] . One kind of lowdimensional approaches that learn motion models by dimensionality reduction can successfully deal with the high-dimensional problem, but it only works on specific motion types with available training data. Other

more »

... pproaches which employ smart sampling directly on high-dimensional pose space don't have that limitation. However, this kind of methods is lack of robustness, with high computational cost, and hard to recover from failures.