Industrial Scene Analysis from a Human-centered Perspective through Intelligent Systems

The transition from Industry 4.0 to Industry 5.0 emphasizes the critical role of human-technology synergy in industrial environments, where artificial intelligence and robotics increasingly share workspaces with human operators. This research examines the advancement of human-centered industrial analysis through computer vision, investigating Human-Robot Interaction (HRI) techniques and introducing the novel concept of Human-Ambient Interaction (HAI). While HRI focuses on direct human-machine interaction, collaboration, and safety protocols in shared workspaces, HAI extends beyond immediate operative zones to analyze broader human interactions with their surroundings, offering new perspectives on safety, behavioral analysis, and environmental awareness. Our investigation in the HRI domain encompasses three distinct research directions: external overhead monitoring for comprehensive space analysis, robot-mounted sensors for dynamic interaction, and human-centric first-person perspectives for intuitive control and gesture recognition. Exploring the external surveillance-like viewpoint, we implemented SeS-GCN, a lightweight pose forecasting methodology capable of dynamically representing the adjacency matrix of a human pose, giving flexibility to the understanding of motion. This is crucial for tasks of human-robot collaboration due to the need for robots to understand human motion. Along with that, we release CHICO, a human pose forecasting dataset specifically designed for human-robot collaboration. Considering the robot's perspective, understanding the motion requires a good human pose estimation methodology that can handle the limits of the robot's onboard sensors view while allowing for human-robot interaction and collaboration. To this end, we created HARPER, a benchmark for human pose estimation, forecasting, and collision prediction from the robot's perspective. Finally, the human-robot firstperson interaction viewpoint has been investigated by developing a novel real-time online hand gesture recognition algorithm, OO-dMVMT, capable of allowing a natural and immediate way to interact with robots. Within the concept of HAI, we study the indirect interaction of humans with the ambient using only human motion and orientation. From the motion perspective, we propose a method, MICRO-TRACK, to effectively implement distributed multi-camera tracking and re-identification in industrial settings. To exploit the orientation of people, we developed a methodology to estimate the human visual focus of attention, creating a real-time dynamic map of the visual attention in a scene, taking into consideration social signals such as conversations and social interactions between multiple subjects. Finally, we investigate efficient deployment strategies, including split computing approaches and distributed processing architectures, to make these systems practical and scalable for real-world industrial implementation The contributions of this work have been implemented in the Industrial Engineering Laboratory (ICE Lab) of the University of Verona, where they were tested in practical applications and are currently deployed.