Pre-Impact Fall Detection Based on Wearable Inertial Sensors using Hybrid Deep Residual Neural Network
Sakorn Mekruksavanich, Ponnipa Jantawong, Anuchit Jitpattanakul
Abstract
Falls are the primary cause of fatal and nonfatal injuries among the elderly. Consequently, pre-impact fall detection that identifies a fall before the body’s collision with the ground is of essential importance. In the latest days, academics have moved their attention from post-impact fall detection to pre-impact fall detection. Pre-impact fall detection systems typically use either a threshold-based or machine learning-based approach, and the thresholding is challenging to determine accurately for threshold-based methods. In addition, while additional features can sometimes assist in categorizing falls and non-falls more precisely, the value computation of prominent features will be too time-intensive, using too much of the algorithm’s operating time. We proposed a pre-impact fall detection method employing wearable inertial sensors and a deep residual model to address the limitations of feature extraction, threshold definition, and algorithm sophistication. After training on a large-scale motion dataset known as the KFall, the suggested deep learning model could identify with 91.87% accuracy in 0.5 second.