Transfer Learning for Efficiency in Elderly Fall Detection with Limited Data Samples

In light of significant demographic shifts worldwide, elderly fall detection is an ongoing, vital area of research. Deep learning, known for its effectiveness in healthcare applications, is challenged by limited accessibility to substantial datasets, especially in the case of fall detection. Moreover, training deep learning models is both time-consuming and costly. To address these issues, in this paper, we implemented a sample size technique called N×Subsampling and utilized transfer learning with MobileNetV2. Our study leveraged the public URFD database, and the obtained experimental results demonstrated a notable achievement: an accuracy range of 94.74% to 98.94%, using only a 15% training subset consisting of 732 images of activities of daily living and 369 images of fall scenarios.