Human pose estimation in surveillance videos using temporal continuity on static pose

Human pose estimation from monocular image sequences is attracting increasing attention, and 2D (image-based) as well as 3D (joint-motion based) approaches have been proposed. The former is computationally fast, and works also for less frequent poses, but reliability is low. The latter is computationally expensive owing to the high-dimensionality of the problem, despite attempts at dimensionality-reduction. We propose to impose temporal continuity constraints on the 2D approaches in order to improve reliability and obtain spatio-temporal descriptions of the action in the image plane. In the first stage, at each frame-level spatial localisation will be done by applying background subtraction, after that static image pose estimation is performed using CRF-based probabilistic assembly of parts, based on the approach by Ramanan [1]. Next, pose continuity is imposed via first-order Markovian constraints on the CRF search. This results in improved spatial accuracy as well as significantly reduced search. We present results from the Weizmann video dataset to demonstrate how such an approach can serve as an image-plane based alternative to full 3D modeling or manifold search.