Accidental falls have an enormous human cost, especially for elderly people, so there is need for automatic fall detection techniques for timely warnings.
We have used smart wearable devices and deep learning on embedded to detect falls with recurrent neural networks. The implementation challenge was due to limited computing and memory resources and to the necessity of battery life for continuous use (24x7).
Datasets with simulated falls by volunteers have been collected: seven carry positions, seventeen different activities, forty volunteers, over five thousands tracks. Manual annotations on videos, basic for training, have been done.