A New Approach for Thermal Vision based Fall Detection Using Residual Autoencoder
Faten A. Elshwemy, Reda Elbasiony, Mohamed T. Faheem Saidahmed
Abstract
This paper focuses on falling of the elderly people which is considered as one of the most critical issue that can face them in their life. To deal with such issue, we propose a new approach named a Spatio-temporal Residual AutoEncoder (SRAE) model. This model is an unsupervised fall detector based on utilizing the deep learning technique to detect falls of the elderly people. Our proposed model uses autoencoder based on convolutional neural network, convolutional long short term memory (ConvLSTM) network, and residual connections to extract spatial and temporal features of videos captured from thermal cameras. The reconstruction error of an autoencoder is used to detect falls recorded in such thermal videos. Furthermore, SRAE model is tested on the publicly available thermal dataset where thermal images conserve the privacy of the elderly under observation which is a very important issue. The obtained results show that the our proposed model detects falls with high receiver operating characteristic area under curve (ROC AUC) (97%) ,and precision recall area under curve (PR AUC) (93%) compared to denoising autoencoder (DAE), convolutional autoencoder (CAE), and convolutional long short term memory autoencoder (CLSTMAE) introduced in the literature.