Data-Driven Video Grasping Classification for Low-Power Embedded System

Abstract

Video-based hand grasp analysis can support both robotics and prosthetics. Indeed, computational aspects represent a major issue, as hand grasp analysis is expected to support grasping systems that are hosted on low-power embedded systems. This paper proposes a framework for video-based grasping classification that is designed for implementation on resource-constrained devices. The framework adopts a fully data-driven strategy and relies on deep learning to deal with advanced analysis of video signals. Nonetheless, the overall design takes advantage of CNN architectures that can cope with the constraints imposed by embedded systems. The experimental session involved a real-world dataset containing daily life activities collected using egocentric perspective. In addition, the complete inference system is implemented on a NVIDIA Jetson-TX2 obtaining real time performances. The results confirm that the proposed system can suitably balance the trade off between accuracy and computational costs.