Computer Vision-Based Human Activity Detection for Intensive Care Patients: A Case Study Based on Region of Interest (ROI)

Human activity detection (HAD) poses a significant challenge in computer vision, offering diverse applications in human-computer interactions (HCIs), ambient assisted living and human-robot interactions. Over the past decade, remarkable advancements have been made in HAD applications, where automated systems play a crucial role in improving quality of life by deciphering and anticipating user needs, identifying unusual behaviours and preventing hazardous situations. In this book chapter, we employ a deep neural network architecture integrated with a motion tracking model to detect patient movements, aiming to alleviate the continuous monitoring burden on hospital staff. To achieve this, we adopt the Faster Region Convolutional Neural Network (Faster R-CNN) architecture. Additional layers are introduced to the existing Faster R-CNN structure to recognize patient activities on the bed. We deploy a pretrained Faster R-CNN model on our customized dataset to classify objects in video frames, determining whether a patient is on the bed or is unoccupied. Subsequently, the region of interest (ROI) was extracted from the frames, encompassing different patients on the bed. The method involves observing gradient changes among pixels in these ROIs, enabling activity monitoring in video frames. This book chapter introduces a novel gradient-based motion detection method using temporal frames over a predetermined duration. The Faster R-CNN model exhibits rapid response and high accuracy, achieving 91% accuracy on our dataset without the need for explicit feature extraction methods.