Fall Detection: Leveraging Depth Information in Bayesian Networks

Deaths due to accidental falls is a pervasive issue across the healthcare system, affecting both professional settings such as hospitals and nursing homes as well as private residences where elderly people might live either alone or without supervision during part of the day. With them being the second leading cause of accidental deaths, just behind traffic accidents, there is significant incentive to develop measures capable of reducing the impact of such accidents. An accurate Fall Detection system presents itself as a viable solution, whereas an automated video surveillance system is capable of raising an alert when such an event occurs. For such an approach to be commercially viable, it must both have a high degree of accuracy and be trainable with easily obtainable data. The first is necessary in order to reduce the number of false positives, and thus not burden the health-care personnel with false alerts. The second would make it possible to implement the system without depending on datasets available for research purposes only, nor requiring a large time investment on creating a private dataset, with the privacy concerns this entails. We address this by introducing an Anomaly Detection approach based on Bayesian Networks. Our approach models a given video frame based on the relationship between simple features extracted from the image, and does not require any kind of private information nor class labels to work. This makes our model both privacy-preserving and with low data requirements. Furthermore, the model can be trained in just a few seconds. We achieve results that far surpass the current state-of-the-art when compared to other unsupervised approaches.