A Fall Detection System using Hybrid Inertial and Physiological Signal Classifiers for Dynamic Environments
Tamonash Bhattacharyya, Prasun Ghosal · 2023
In this modern day and age, passive health monitoring and event alert systems are crucial in redefining healthcare for future generations. Among other health events, Accidental Fall is one of the leading causes of injuries in the elderly, who make up about three-quarters of the total victim pool. This study focuses on low latency, power-efficient implementation of an accidental Fall Detection system in dynamic environments (indoor and outdoor) from Accelerometer and PPG sensor data, using a hybrid CNN-kNN estimator for computation at the edge. Convolution Neural Networks (CNNs), the foundation for almost all image processing and classification models, are used here to independently estimate the representative probability of motion in individual axes. The k-Nearest Neighbour (kNN) algorithm classifies if the event is a fall considering the features from the CNN inter-dependently, correlating them as it would in case of an accidental fall, with a shallow computational requirement. Our proposed design thus focuses on an efficient and custom Independent Axis Inertial Analysis (2IA) design for motion and vital signal analysis, along with a Dependent Analysis (DA) model for the final fall estimation.