Smartphone Sensor-Based Fall Detection Using Machine Learning Algorithms

Human Activity Recognition and particularly detection of abnormal activities such as falls have become a point of interest to many researchers worldwide since falls are considered to be one of the leading causes of injury and death, especially in the elderly population. The prompt intervention of caregivers in critical situations can significantly improve the autonomy and well-being of individuals living alone and those who require remote monitoring. This paper presents a study of accelerometer and gyroscope data retrieved from smartphone embedded sensors, using iOS-based devices. In the project framework there was developed a mobile application for data collection with the following fall type and fall-like activities: Falling Right, Falling Left, Falling Forward, Falling Backward, Sitting Fast, and Jumping. The collected dataset has passed the preprocessing phase and afterward was classified using different Machine Learning algorithms, namely, by Decision Trees, Random Forest, Logistic Regression, k-Nearest Neighbour, XGBoost, LightGBM, and Pytorch Neural Network. Unlike other similar studies, during the experimental setting, volunteers were asked to have smartphones freely in their pockets without tightening and fixing them on the body. This natural way of keeping a mobile device is quite challenging in terms of noisiness however it is more comfortable to wearers and causes fewer constraints. The obtained results are promising that encourages us to continue working with the aim to reach sufficient accuracy along with building a real-time application for potential users.