Gutiérrez, JesúsMartín Gutiérrez, SergioRodriguez, Victor2024-10-112024-10-112023-06-14J. Gutierrez, S. Martin, V. Rodriguez Human stability assessment and fall detection based on dynamic descriptors IET Image Processing. 17, 3177–3195 (2023), Wiley. Print ISSN: 1751-9659. EISSN: 1751-9667. https://doi.org/10.1049/ipr2.12847IET Image Processinghttps://doi.org/10.1049/ipr2.12847https://hdl.handle.net/20.500.14468/24027This is the accepted manuscript of the article. The registered version was first published on Image Processing. 17, 3177–3195 (2023), Wiley. Print ISSN: 1751-9659. EISSN: 1751-9667, is available online at the publisher's website: https://doi.org/10.1049/ipr2.12847 Este es el manuscrito aceptado del artículo. La versión registrada fue publicada por primera vez en Image Processing. 17, 3177–3195 (2023), Wiley. Print ISSN: 1751-9659. EISSN: 1751-9667, está disponible en línea en el sitio web del editor: https://doi.org/10.1049/ipr2.12847Fall detection systems use a number of different technologies to achieve their goals, contributing, this way, to better life conditions for the elderly community. The artificial vision is one of these technologies and, within this field, it has gained momentum over the course of the last few years as a consequence of the incorporation of different artificial neural networks (ANN’s). These ANN’s share a common characteristic, they are used to extract descriptors from images and video clips that, properly processed, will determine whether a fall has taken place. However, these descriptors, which capture kinematic features associated to the fall, are inferred from datasets recorded by young volunteers or actors who simulate falls. Given the well documented differences between these falls and the real ones concerns about system performances in the real-world, out of laboratory environments, are raised. This work implements an alternative approach to the classical use of kinematic descriptors. To do it, for the first time to the best of our knowledge, we propose the introduction of human dynamic stability descriptors used in other fields to determine whether a fall has taken place. These descriptors approach the human body in terms of balance and stability, this way, differences between real and simulated falls become irrelevant, as all falls are a direct result of a fail in the continuous effort of the body to keep balance, regardless of other considerations. The descriptors are determined by using the information provided by a neural network able to estimate the body center of mass and the feet projections onto the ground plane, as well as the feet contact status. The theory behind this new approach and its validity is studied in this article with very promising results, as it is able to match or over exceed the performances of previous systems using kinematic descriptors in laboratory conditions and, given the independence of this approach from the conditions of the fall, real or simulated, it has the potential to have a better behavior in the real-world than classic systems.eninfo:eu-repo/semantics/openAccess33 Ciencias Tecnológicasartículomonocular human stability assessmenttemporal convolutional neural networkfall detectionfall identificationdynamic descriptorhuman balance