Automatic Estimation of Manual Sorting Efficiency in Non-Ferrous Scrap Recycling Using Dual-Camera Computer Vision and Deep Learning

Manual picking is a critical refining stage in non-ferrous scrap recycling, yet it is rarely measured continuously. This study presents a dual-camera computer-vision system that estimates manual sorting efficiency by counting materials before and after picking. YOLO detectors trained on an in-plant dataset (4361 images, four classes) are combined with a Kalman-filter tracker and line-crossing counters to yield per-class and hourly efficiency indicators. YOLOv11m reached a higher maximum F1-score (0.82 vs. 0.79), and selected models attained mAP@50 values of 0.85–0.86. Image-level splitting may make detector metrics optimistic; stage comparisons are indicative only. In a six-hour test, 3753 input objects and 690 output objects were counted, yielding a count-derived global efficiency estimate of 81.61% (95% Wilson interval, sampling only: 80.34–82.82%). Estimated efficiency exceeded 90% for circuit boards and rubber but was only 75.95% and 73.75% for copper and brass, the most valuable fractions. This demonstrates end-to-end operation, not validated accuracy: no manual reference counts or real output-camera test images were available, so all efficiency values are conditional on the automatic counts. Sensitivity analysis quantifies error propagation into the indicator. The results show the feasibility of detection, tracking, and input–output comparison as a monitoring prototype for Industry 4.0 recycling lines.