| Title |
YOLOv11-based Real-time Hospital Safety Monitoring: Robustness to Occlusion and Distortion in Fisheye Cameras |
| DOI |
https://doi.org/10.5573/ieie.2025.62.12.22 |
| Keywords |
Hospital safety monitoring; YOLOv11; Fisheye distortion; Occlusion robustness; Class imbalance; Real-time detection |
| Abstract |
Hospital patient safety monitoring is essential for healthcare quality, but traditional approaches face limitations from workforce shortages and 24-hour surveillance challenges. While YOLO-series models show promise in video surveillance, hospital environments present unique difficulties: severe radial distortion from fisheye cameras, occlusion by medical equipment, and class imbalance. This study proposes a YOLOv11-based real-time monitoring framework addressing these challenges. We select four alarm-relevant classes (wheel-chair-patient, stretcher_patient, patient_on_bed, stretcher) from eight original classes and apply oversampling, focal loss, and class-aware mosaic augmentation to mitigate imbalance. Cutout augmentation and temporal smoothing (additional latency ≤50ms) handle occlusion, while radial distortion augmentation and F1_edge metric address fisheye distortion. On hospital-6kwxt v3 Valid split (174 images, 177 GT objects), we achieve mAP@0.5 of 85.6%, macro-F1 of 82.1%, and macro-F1_edge of 74.1% with real-time processing (≥45 FPS). This framework demonstrates effective integrated handling of imbalance, occlusion, and distortion, providing immediately applicable guidelines for hospital safety alarm systems. |