Synthetic-to-Real Transfer Learning for Image-Based Recognition of Unsafe Work States at Scaffold Platform Edges

Automated recognition of worker posture–structure relationships at scaffold platform edges is constrained by the scarcity of unsafe work-state images and bias associated with imaging conditions. We re-audited a single public dataset comprising staged working-at-height scenarios and constructed a four-class dataset containing 96,603 images, 246 sequences, and 57 identifier-derived event proxies. The classes were normal work (C0), climbing/contact (C1), upper-body leaning beyond the guardrail (C2), and edge work (C3). Real-only training was compared with synthetic pretraining followed by real-data head adaptation (PFT) using 3000 and 9000 low-fidelity images generated with Blender and BlenderProc2. A dual-input ConvNeXt-Tiny combining the full frame with an annotation-derived worker-context crop was evaluated using three-fold event-proxy-grouped cross-validation and two random seeds, with no overlap in event-proxy, sequence, or source identifiers across folds. Mean event-level Macro-F1 was 0.4158 ± 0.0205 for Real-only, 0.5392 ± 0.0534 for PFT-3k, and 0.6011 ± 0.0131 for PFT-9k. PFT-9k exceeded Real-only by 0.1853 and achieved higher F1 for all four classes, although C1 and C3 remained comparatively challenging. Because PFT-3k and PFT-9k also differed in synthetic-family composition, training exposure, and pretraining seed, their difference cannot be attributed to image count alone. Synthetic pretraining was associated with higher performance under the evaluated internal-validation setting. However, the results represent offline classification under reliable annotation-derived ROI inputs within a single dataset and do not establish end-to-end CCTV performance or generalizability to independent construction sites, workers, or camera systems.