Video Grasping Classification Enhanced with Automatic Annotations

Video-based grasp classification can enhance robotics and prosthetics. However, its accuracy is low when compared to e-skin based systems. This paper improves video-based grasp classification systems by including an automatic annotation of the frames that highlights the joints of the hand. Experiments on real-world data prove that the proposed system obtains higher accuracy with respect to the previous solutions. In addition, the framework is implemented on a NVIDIA Jetson TX2, achieving real-time performances.