ELDERLY INDIVIDUAL FALL DETECTION SYSTEM USING DEEP LEARNING

Abstract

Falls often harm the elderly, end in serious injury, may require hospital treatment and unfortunately can be deadly. Ensuring elderly people receive rapid medical care following a slip or fall has become more difficult now that there are people that are more elderly. For this reason, there is now a greater desire for systems that can automatically spot falls to give swift care. In this research work, a CNN-BiLSTM based approach using skeletal key points for elderly fall detection is proposed. Our process includes two parts: first, we use Google’s MoveNet Thunder to extract the body pose from each video frame and then, we pass the series of skeletal keypoints through a network consisting of Convolutional Neural Networks for understanding spatial features and Bidirectional Long Short Term Memory networks for learning the motion itself. The two-stage process makes it possible for the model to recognize falls accurately by observing both how someone moves and his or her body position in the sequence. Publicly available benchmark dataset, URFD, was utilized for experiments in this work. No standard guidelines have been made for the design of falling detection systems, despite their many forms. For this reason, we have gathered a range of works for this topic to provide a summary of where research on human position based fall detection algorithms currently is. Based on the URFD dataset, the proposed hybrid approach performs better than existing methods, mainly because it can represent complex fall events well. We have found that the model obtains an accuracy of 95.80%, a precision of 93.90%, a recall of 94.78%, a specificity of 93.75% and an AUC of 0.9311.