Monocular 3D Human Pose Estimation via Sequential Second Order Cone Programming

This paper presents an efficient method for monocular recovering and tracking 3D human pose using 3D to 2D joints correspondences. Different from previous work, its main novelty lies in several aspects: Firstly, our method does not involve any complex features, which means that it does not tend to rely on good foreground segmentation. Secondly, formulating the model as an second order cone programming (SOCP) problem has great advantages since the SOCP can be solved quite reliably and efficiently. Finally, it advocates the use of more effective prediction strategy to increase robustness. Experiments on walking sequences demonstrate that our model performs accurately and reliably.