Athlete pose estimation by non-sequential key-frame propagation

Abstract

This paper considers the problem of estimating human pose in challenging monocular sports videos, where manual intervention is often required in order to obtain useful results. Fully automatic approaches focus on developing inference algorithms and probabilistic prior models based on learned measurements and often face challenges in generalisation beyond the learned dataset. This work expands on the idea of using an interactive model-based generative technique for accurately estimating the human pose from uncalibrated unconstrained monocular TV sports footage. A method of keyframe propagation is introduced to obtain reliable tracking from limited operator input by introducing the concepts of keyframe propagation and optimal keyframe selection assistance for the operator. Experimental results show that the approach produces results competitive with those produced with twice the number of manually annotated keyframes, halving the amount of interaction required.