Federated Transfer Learning for Vision-Based Fall Detection

The incidence of human falls has emerged as a growing public health concern, particularly among the elderly and individuals with disabilities. Fall detection has assumed a vital role in healthcare research, aiming to mitigate the adverse consequences of falls, such as severe medical complications, prolonged treatment, hospitalization, and potential permanent disabilities. Fall detection methods encompass auxiliary equipment-based and computer vision-based approaches, which have gained prominence due to the increasing effectiveness and resilience of the Internet of Things (IoT). Nevertheless, computer vision-based methods give rise to significant privacy concerns, as monitoring human movements may entail capturing sensitive personnel images. To address this concern, this paper proposes a privacy-preserving approach for fall detection, utilizing Federated Learning in conjunction with Transfer Learning to train models. This study implemented various well-known pre-trained models, including VGG16, VGG19, InceptionV3, InceptionResNetV2, and Classic CNN, in a federated environment using a dataset comprising 30 videos. Among these models, InceptionResNetV2 achieved the highest test accuracy of 97.38%.