Human Pose Estimation for RGBD Imagery with Multi-Channel Mixture of Parts and Kinematic Constraints

Abstract

In this paper, we present a approach that combines monocular and depth information with a multichannel mixture of parts model that is constrained by a structured linear quadratic estimator for more accurate estimation of joints in human pose estimation. Furthermore, in order to speed up our algorithm, we introduce an inverse kinematics optimization that allows us to infer additional joints that were not included in the original solution. This allows us to train in less time and with only a subset of the total number of joints in the final solution. Our results show a significant improvement over state of the art methods on the CAD60 and our own dataset. Also, our method can be trained in less time and with smaller fraction of training samples when compared to state of the art methods.