OneFall-GAN: A one-class GAN framework applied to fall detection

Abstract

One of the most common health risks for senior citizens is a falling event, and to reduces the risk of death, a fall needs to be quickly reported. Thus, automatic fall detection systems were proposed to mitigate the falling problem, most of them relying on image detection. However, identifying a fall from a sequence of images is challenging because the contextual information present in some activities of daily living (ADL) are similar to fall events, and the recognition model has to deal with very high-level information. Furthermore, the publicly available datasets are generally unbalanced, having only synthetic data of fall events. Thus, approaches that do not need explicit labels to train a model are desirable. As falls are rare events, a possible solution is to use a one-class approach, and Generative Adversarial Networks (GAN) have presented competitive results in this context. This paper proposes OneFall-GAN, a One-Class GAN that uses only ADL for training and can identify a fall as an anomaly on the learned representations. We trained our neural network using the data captured by an RGB camera and pre-processed by a pre-trained Mask R-CNN that can segment and binarize a person in a scene. Thus, our model has robust temporal information used in the training phase and once trained, the discriminator is specialized on classifying ADL events. Consequently, when a fall event is presented, it is capable of classifying them as an anomaly. Using this approach, we are turning a binary classification problem into a one-class approach. Our experimental results show that our OneFall-GAN framework is a competitive alternative for fall classification as an anomaly detection task using only visual data, particularly because it achieved the best performance when compared to current state-of-the-art approaches, reaching an accuracy of 99.02%, 0.97 precision, 1.0 recall on PRECIS HAR and 98.75% accuracy, 0.87 precision, 0.99 recall on UP Fall.