Detecting Falls with Recurrent Autoencoders and Body Acceleration Data

According to the world population statistics, the elderly proportion is continuously growing. With this growth, the concerns of this population increase. Falls are a common problem within the elderly population causing considerable health and financial impacts. In order to mitigate such impacts, automatic fall detection systems are more and more employed by the elderly. In this work, we propose a fall detection system based on the body acceleration data, where we investigate the use of recurrent-autoencoders to distinguish between normal activities of daily living and falls. We study different training strategies combined with either the use of the autoencoder as feature extractor or employing its reconstruction error for decision. Experiments on the UniMiB SHARE fall data set provides interesting insights about the designed system. .