Activities of Daily Living and Falls Recognition and Classification from the Wearable Sensors Data

The increasing percentage in population of the elderly and of the chronically diseased requires new solutions for tele-medicine and the continuous real-time remote health status monitoring. The present paper presents a comparison study of various machine learning and deep learning techniques for cross-person prediction of both activities of daily living and falling down. The experiments are performed on the smartphone's raw accelerometer data from the publicly available UniMiB SHAR dataset. Different cross-validation methods are tested and the performance of each classifier discussed. Deep learning method outperforms the other classifiers in many configurations, performed on the different subsets.