Human Pose Estimation with Supervoxels

This thesis investigates how segmentation as a preprocessing step can reduce both the search space as well as complexity of human pose estimation in the context of smart environments. A 3D reconstruction is computed with a voxel carving algorithm. Based on a superpixel algorithm, these voxels are segmented into supervoxels that are then applied to pictorial structures in 3D to efficiently estimate the human pose. Both static and dynamic gesture recognition applications were developed.