Monocular video based markerless 3D human pose estimation by using Local Multi-connected Belief Propagation with multi-cue fusion

In this paper, we put forward a concept of Local Multi-connected Belief Propagation (LMBP) and apply it on a 2D body model to capture 2D image coordinates of the human body joints. When performing the likelihood measurement, three different image features are fused adaptively to construct the likelihood function. And then, we lift the 2D tracking results to 3D human poses by utilizing Tylor method. The proposed method is then applied to several video sequences which contain different subjects, challenging human motions and complex backgrounds, and the tracking results indicate its effectiveness.