Particle swarm optimization on low dimensional pose manifolds for monocular human pose estimation

摘要: Automatic assessment of situations with modern security and surveillance systems requires sophisticated discrimination capabilities. Therefore, action recognition, e.g. in terms person-person or person-object interactions, is an essential core component any system. A subclass recent recognition approaches are based on space time volumes, which generated from trajectories multiple anatomical landmarks like hands shoulders. general prerequisite these methods the robust estimation body pose, i.e. a simplified model consisting several landmarks. In this paper we address problem estimating 3D poses monocular person image sequences. The first stage our algorithm localization parts 2D image. For this, part object detection method used, previous work has been shown to provide sufficient basis for landmark single step. output processing step probability distribution each indicating possible locations coordinates. second searches pose estimates that best t 15 distributions. resolving ambiguities introduced by uncertainty landmarks, perform optimization within Particle Swarm Optimization (PSO) framework, where hypothesis represented particle. Since search high-dimensional needs further guidance deal inherently restricted input information, propose new compact representation motion sequences provided capture databases. Poses sequence embedded low-dimensional manifold. We represent referred as splines using small number supporting point poses. PSO can be extended process directly splines. Results proposed UMPM benchmark.