Advancing AI-driven surveillance systems in hospital: A fine-grained instance segmentation dataset for accurate in-bed patient monitoring.

In the era of digital health, artificial intelligence (AI)-driven patient monitoring systems have attracted growing interest for their potential to prevent accidents in clinical settings. However, the advancement of these systems requires the availability of high-quality and annotated datasets that enable accurate detection and recognition of real-world situations. This study aims to present a new dataset for recognizing and tracking patients in bed positions proper to advance surveillance systems. From the free available Fall Simulation Dataset (FSD), videos depicting bed-related actions were selected, and each frame was meticulously annotated with segmentation labels for 'person' and 'bed' instances. A total of 1487 frames were manually annotated to create a fine-grained dataset for in-bed patient monitoring. A YOLOv8 instance segmentation model was fine-tuned and evaluated to benchmark the dataset's effectiveness. The model demonstrated high segmentation performance following data augmentation techniques and a rigorous 5-fold cross-validation protocol. For bounding box detection, the model achieved a precision of 0.97, a recall of 0.97, and a mAP50 of 0.99. Similarly, segmentation masks reached a precision of 0.98, recall of 0.98, and mAP50 of 0.98. This dataset contributes to advancing non-invasive, in-bed patient surveillance systems, enabling more effective and reliable patient tracking and monitoring within clinical and residential care environments using AI techniques.