EgoFall: A First-Person View Fall Detection System

Fall detection is crucial for elderly care, as over 30 million older adults experience falls annually, according to the World Health Organization. Most current fall detection systems rely on surveillance cameras to identify sudden posture changes that may indicate a fall. However, these systems often require optimal camera placement and a clear line of sight to the individual. In this work, we developed EgoFall, a wearable camera-based system that detects falls from first-person images, removing the need for fixed environmental cameras. EgoFall first applies Oriented FAST and Rotated BRIEF (ORB) to extract keypoints from the input images. It then uses optical flow methods to compute speed and direction features. Finally, a machine learning model identifies fall events. To evaluate the practicality of EgoFall, we implemented it on a Raspberry Pi 3B board. We compared three machine learning models: SVM, KNN, and DT, using the open dataset RUG-EGO-FALL. Our experimental results show that SVM performed the best, achieving an accuracy of 84.9% in detecting fall events.