An Enhanced Fall Detection Approach in Smart Homes Using Optical Flow and Residual Autoencoder
International Journal of Advanced Trends in Computer Science and Engineering
When focusing on the elderly people, falling is considered a major health problem that can lead to serious injuries; sometimes it can cause the death of them. So, fall detection is an important service for the elderly healthcare to improve the daily life of them and decrease the costs of monitoring these individuals. The fall detection is considered as a problem of detecting anomalies because falling is an abnormal activity. To deal with such issue, we present (OFSRAE) which is unsupervised
... detection framework based on deep learning techniques to detect the elderly people falls. Our proposed framework consists of two stages: the data preprocessing and the deep learning model. We applied the dense optical flow in the first stage to extract the information of the motion and direction of moving objects in the foreground. The deep learning model in the second stage is based on the convolutional long short term memory autoencoder (ConvLSTMAE) network, and the residual connections to extract spatial and temporal features of videos that are captured from thermal and depth cameras. The reconstruction error of an autoencoder is used to identify falls as anomalies recorded in such videos. We experimentally evaluated OFSRAE framework on the publicly available UR and thermal fall datasets which conserve the elderly privacy that is an important issue during monitoring. The experimental results show that the proposed framework detects falls with high performance compared to other deep learning models in the literature.