Abstract: Dicriminative approaches to 3D human pose estimation directly learn a mapping from image observations to pose,which requires large training sets. Gaussian process regression(GPR) to learn this mappings has been limited for high computational complexity, so we proposed a incrementally learning mappings based on GPR and Locally Weighted Projection Regression(LWPR). The approach utilized GPR to learn individual local models and LWPR to update existing models or learn a new local model for pose estimation. The experiment showed that the approach could greatly decrease computational complexity and exactly estimate the poses.

Key words: Pose estimation, Gaussian process regression, Locally weighted projection regression, Incremental learning