Ambient Sensor-based Abnormal Human Activity Detection Using Autoencoder

Among the challenges related to human activity recognition systems that are deployed in ambient assisted living environments, is the ability to identify abnormal behavior patterns. This laborious task has led to advances in deep learning in monitoring, particularly of the elderly, which has helped improve the performance of these systems.This work proposes a novel deep neural network architecture that consists of a specific autoencoder (AE) where each layer of the deep network is represented by a long short-term memory (LSTM) model. Then, two distance measures, the mean square error (MSE) and the Mahalanobis distance (MD), are applied to classify the desired activity as normal or abnormal according to a predetermined threshold. We compared the proposed approach with other state-of-the-art methods in the context of abnormal activity detection. The experimental results achieved interesting performances.