A local-global coupled-layer puppet model for robust online human pose tracking

We propose a new method for online tracking of articulated human body poses.Our method offers online sequential tracking from one frame to the next.Many other methods mutually optimize poses offline over all frames of a sequence.We propose a novel cross-coupled global-local model of articulated human body pose.We propose an adaptive penalty function for optimizing the pose estimates. This paper addresses the problem of online tracking of articulated human body poses in dynamic environments. Many previous approaches perform poorly in realistic applications: often future frames or entire sequences are used anticausally to mutually refine the poses in each individual frame, making online tracking impossible; tracking often relies on strong assumptions about e.g. clothing styles, body-part colours and constraints on body-part motion ranges, limiting such algorithms to a particular dataset; the use of holistic feature models limits the ability of optimisation-based matching to distinguish between pose errors of different body parts. We overcome these problems by proposing a coupled-layer framework, which uses the previous notions of deformable structure (DS) puppet models. The underlying idea is to decompose the global pose candidate in any particular frame into several local parts to obtain a refined pose. We introduce an adaptive penalty with our model to improve the searching scope for a local part pose, and also to overcome the problem of using fixed constraints. Since the pose is computed using only current and previous frames, our method is suitable for online sequential tracking. We have carried out empirical experiments using three different public benchmark datasets, comparing two variants of our algorithm against four recent state-of-the-art (SOA) methods from the literature. The results suggest comparatively strong performance of our method, regardless of weaker constraints and fewer assumptions about the scene, and despite the fact that our algorithm is performing online sequential tracking, whereas the comparison methods perform mutual optimisation backwards and forwards over all frames of the entire video sequence.