Recurrent Neural Network for Human Fall Motion Prediction

Abstract

Falls pose a significant health risk, especially for older adults, where one in three people over 65 experiences a fall each year. For those over 85, the consequences of falls can be severe, leading to life-threatening injuries and a marked decline in quality of life. To address the critical need for fall prevention, this study proposes a prediction approach using Recurrent Neural Networks (RNNs) to recognize patterns in human motion that may indicate an impending fall. By utilizing the CAUCA fall dataset—carefully designed to detect abnormal fall movements—we implemented and assessed different RNN architectures, including Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU). Our findings show that the GRU model performed best, achieving an accuracy of 0.716, MPJPAE of 32.921 pixels, MPJVE of 22.457 pixels per frame, and Euclidean distance metric outperforming the other models, with RNN closely following at an accuracy of 0.716, MPJPAE of 41.872 pixels, MPJVE of 21.44 pixels per frame. These promising results suggest that RNN models, particularly GRU, can serve as valuable tools in predicting falls, offering a foundation for future technology that can help prevent falls before they happen.