A vision-based observation layer framework for manufacturing digital twins: detection, tracking, and zone analytics from ceiling-mounted production video

Introduction Manufacturing digital twins require reliable observations of evolving production‐floor state, yet many camera-based studies stop at detection or tracking and do not formalize how video becomes twin‐ready state variables. Methods This study presents a vision-based observation-layer framework that links video acquisition, detection, tracking, spatial semantics, observation-variable extraction, and a digital twin interface. The framework was evaluated on a real food‐production floor recorded by a ceiling‐mounted camera. A time‐separated design kept the final production hour fully held out for detector and tracker evaluation. Results The selected detector (YOLO11x, freeze depth 22) reached mAP50 = 0.939 and mAP50‐95 = 0.754 on the primary held‐out set; ByteTrack reached HOTA = 0.783 at 27.6 frames/s on two manually reviewed tracking clips. The resulting stack produced zone occupancy, entries, exits, and dwell-time summaries over a 60 min held-out recording. Discussion These results support the proposed component as an observation layer for a manufacturing digital twin. The study does not claim to implement a complete twin with simulation, prediction, bidirectional synchronization, or closed-loop control.