A cross-dataset deep learning-based classifier for people fall detection and identification

Abstract

This paper addresses the issue of fall detection, particularly for elderly individuals who may live alone and be unable to call for help after a fall. The objective is to develop a deep learning-based approach that can detect falls and identify individuals without needing model fine-tuning for different datasets. The proposed method uses a multi-task learning model that processes raw inertial data to simultaneously detect falls and identify people. The model achieves over 98% accuracy in fall detection across four datasets, with less than 1.6% false positives, and identifies people with an average accuracy of 79.6%. It operates in real-time, requiring no retraining for new subjects, making it suitable for practical implementation.