Multiactivity 3-D Human Pose Tracking in Incorporated Motion Model with Transition Bridges

Abstract

Multiactivity 3-D pose tracking is an extremely difficult computer vision task due to various activity types and high-dimensional state space. To solve these challenges, a novel generative framework which incorporates separate and unified motion models into one system is proposed. On one hand, pose tracking benefits from the representations of activities in low-dimensional space, but the complexity of the representations increases with the number of activities. Obviously, the separate modeling of different activities can improve the scalability of models. On the other hand, there often exist similar activities in a long motion sequence, they should be grouped and modeled in a unified model, which can enhance human pose tracking stability and reduce the optimization process over motion model for a large number of actions. Our incorporated model not only combines their advantages but also overcomes their disadvantages. Finally, a particle-based joint annealing optimization is utilized for tracking. During pose tracking, the distribution of particles depends on the outputs of action recognition system. Thus, the particles are transferred amongst different models and most of them are distributed in the model that corresponds to the currently observed action. To propagate the particles among motion models and quickly respond to the transition between different activities, a group of short and smooth transition bridges are constructed so that the interactivity motions seem to be more realistic and natural. Extensive results, via qualitative and quantitative analyses, verify the effectiveness of our model in multiactivity pose tracking.