Anomaly Detection Based on Denoising Autoencoder and Variational Autoencoder with Obrid-Sensor
Tomohiro Hayata, Taiki Sunakawa, Yuki Horikawa, Atsushi Matsubara, Seiji Nishifuji, Shota Nakashima · 2022
Aging society is a serious problem around the world. The number of elderly people living alone is increasing. If they have an accident or gets ill at home, there is a risk that they may die because of lack early detection. Some existing systems use cameras to detect dangerous states of the people. Using this system is difficult in places where privacy should be protected like bathrooms. A system has been proposed to detect abnormal states using a sensor called Obrid-Sensor. This sensor can observe subject's state without invading privacy of because brightness information to be acquired is one dimension. A previous study is proposed a fall detection method with an autoencoder by estimating subject's state from standing data. Various noises cause a loss of performance in the autoencoder. Therefore, the method was proposed by using a denoising autoencoder which is more robust to noise, a variational autoencoder which is more expressive. As a result, recall of all methods was 100 %, and fall detection could be performed without false negative. The denoising autoencoder was the most suitable for anomaly detection.