Athlete Pose Estimation from Monocular TV Sports Footage

Abstract

Human pose estimation from monocular video streams is a challenging problem. Much of the work on this problem has focused on developing inference algorithms and probabilistic prior models based on learned measurements. Such algorithms face challenges in generalization beyond the learned dataset. We propose an interactive model-based generative approach for estimating the human pose in 2D from uncalibrated monocular video in unconstrained sports TV footage without any prior learning on motion captured or annotated data. Belief-propagation over a spatio-temporal graph of candidate body part hypotheses is used to estimate a temporally consistent pose between key-frame constraints. Experimental results show that the proposed generative pose estimation framework is capable of estimating pose even in very challenging unconstrained scenarios.