Full Body Pose Estimation During Occlusion using Multiple Cameras

Automatic estimation of the human pose enables many interesting applications and has therefore achieved much attention in recent years. One of the most successful approaches for estimating unconstrained poses has been the pictorial structures framework. However, occlusions between interacting people is a very challenging problem for methods based on pictorials structure as for any other monocular pose estimation method. In this report we present work on a multi-view approach based on pictorial structures that integrate low level information from multiple calibrated cameras to improve the 2D pose estimates in each view. The proposed method in shown to work under heavy occlusions but does not improve the pose estimates in the non-occluded cases in it's current form.