Low-cost and non-intrusive human digital twin component for task and navigation tracking through pose estimation

Abstract

Industry 5.0 strives to enhance manufacturing by prioritizing the human role, aiming that workers excel in productivity within safe and intuitive work environments. To address this need, Human Digital Twins (HDT) have gained notoriety in which human-related data is collected to feed other methodologies such as Lean manufacturing or cognitive deloading approaches. Together with concerns regarding data privacy and regulatory compliance, this imposes a formidable challenge. This work introduces a foundational HDT component designed to facilitate the human digital twinning process. With only a single camera and heavily relying on edge computing, this system uses deep learning image processing algorithms deployed on a Google Coral Dev Board, ensuring user anonymity and privacy while delivering near real-time performance. To evaluate the effectiveness of this cost-efficient solution, experiments were conducted in an industrial laboratory setting to assess its functionality as a navigation support system for warehouse picking. To the authors’ knowledge, no prior studies have utilized a single-camera depth-perception setup for human pose estimation in this context. The system achieved strong performance, estimating the third dimension of human pose keypoints with an absolute error below 0.11 meters and classifying tasks with 94% accuracy across diverse picking sequences and subjects.