• 3dcnn
• deep learning
• openpose
• patient monitoring
• pose estimation
• skeleton extraction
• yolo

The adoption of artificial intelligence in healthcare, particularly in Patient Monitoring PM, is improving the way healthcare professionals interact with patients. PM system usually relies on Internet of Things devices and Deep Learning models to monitor patient health, offering support for clinical decision-making. This thesis investigates the application of DL-based Human Action Recognition (HAR) for in-bed patient movement monitoring, which is an important task for patients with psychiatric conditions or those recovering from surgery. The challenge is to perform accurate HAR in a clinical setting, where non-intrusive, markerless solutions are preferred.
The work aim to enhance the accuracy of remote patient monitoring for action recognition in clinical environments, to do so investigates two primary deep learning based approaches: end-to-end convolutional neural networks applied to raw video data, and a hybrid pipeline based on skeleton-based feature extraction. Two dataset have been considered to evaluate the performances: a dataset collected within a hospital simulation and the art action recognition dataset.