Fall Detection in Hospital Rooms with Probabilistic Graphical Models

Unintentional injury deaths, primarily caused by simple falls among elderly individuals, have consistently ranked as the second most common cause of death over the years, as documented by the World Health Organization. This not only poses a significant health concern but also imposes a substantial financial burden on hospitals, as insurance companies typically do not cover such events. This project aims to address this challenge by developing a simple Artificial Intelligence system based on Video Anomaly Detection to detect fall situations that require less computational power and protect patients' privacy. The system comprises a Keypoint Detector for estimating the patient's pose and a discrete Bayesian Network to detect whether a person has fallen or not. The approach has been trained and tested under the public dataset CAUCAFall achieving an average accuracy of 90.62%, Area Under the ROC Curve of 88.71%, Region-Based Detection Criterion of 86.23% and Track-Based Detection Criterion of 98.39%. The obtained results demonstrate highly competitive performance, approaching the levels achieved by the most recent State-of-the-art works tested on CAUCAFall. Our unsupervised methodology to fall detection is privacy-preserving, computationally lightweight, and learns patterns from the training data in seconds on a conventional 30GB memory Google Cloud Virtual Machine with 8 AMD EPYC 7B13s Intel Broadwell CPUs.