3D Hand Pose Estimation Using a Single Camera for Unspecified Users

Abstract

The three-dimensional hand pose estimation this paper proposes uses a single camera to search a large database for the hand image most similar to the data input. It starts with coarse screening of proportional information on hand images roughly corresponding to forearm or hand rotation, or thumb or finger bending. Next, a detailed search is made for similarity among selected candidates. No separate processes were used to estimate corresponding joint angles when describing wrist’s rotation, flexion/extension, and abduction/adduction motions. By estimating sequential hand images this way, we estimated joint angle estimation error within several degrees - even when the wrist was freely rotating - within 80 fps using only a Notebook PC and high-speed camera, regardless of hand size and shape.