(英) |
Self-occlusions occur in 3D hand pose estimation with a single camera, which will cause improper behaviors in a robot hand, especially in its grasping and pinching due to the occlusion of the thumb. This study therefore proposes an optimal usage of two cameras for a 3D hand pose estimation, so as to prevent the occlusion and improve the behavior of the robot. Our system adopts two high-speed cameras as a major and minor, whose directions are perpendicularly intersecting. Hand images, image features, and joint angle data are recorded from two cameras and paired in every dataset of the database for searching the most similar hand image with an inputted hand image. The experimental results show that the estimation errors decrease in all finger joints as well as wrist rotation, compared with those with single camera, which indicates the effectiveness of the proposed method. |