It is very important to predict falls before hitting on the ground and to trigger fall protective airbags timely to prevent serious physical and psychological injuries to the elderly. A lightweight fall prediction algorithm based on bi-directional long short-term memory neural networks is proposed. A deep learning model instead of manual feature extraction at the terminal is used to extract deep features of accelerometer data automatically to improve the generalization capability of the model. Firstly,the data window selected according to the fall landing moment and lead time is adopted as the input. Secondly,a lightweight bi-directional long short-term memory neural net-work is designed to extract acceleration features and predict the fall. Finally,the model is lightened with the help of TensorFlow Lite framework. Experimental results show that the new algorithm achieves an accuracy of 96.92％,a sensitivity of 95.73％,and a specificity of 98.15％ on the SisFall fall public dataset with a lead time of 215 ms,and the TensorFlow Lite model for deployment to embedded ter-minals occupies only 62.2 kB in memory size with 1.2 ms running time,which is expected to be deployed in embedded terminals for re-al-time fall prediction. The proposed algorithm achieves higher prediction accuracy and longer lead time,which is more suitable for re-source-constrained embedded devices and provides a further reference for the development of fall prediction and wearable fall protection devices for the elderly.
关键词
跌倒预测/深度学习/双向LSTM/前置时间/可穿戴设备/保护气囊
Key words
fall prediction/deep learning/bi-directional LSTM/lead time/wearable device/airbag
维普
万方数据