Logo: to the web site of Uppsala University

 Falls can affect individuals of all age groups, leading to severe health problems, including serious injuries and death. Vision-based deep learning Fall Detection Systems (s) (FDS) have shown high accuracy and reliability in detecting falls.

This study investigates the application of a large foundational pre-trained Vision Transformer (ViT), specifically Video Masked Autoencoder with dual masking (VideoMAE V2), which was initially trained on a general human action dataset, for the downstream task of fall recognition using transfer learning. The objective is to evaluate the model’s generalization capabilities against state-of-the-art models and assess its performance in real-world scenarios. Additionally, we explore the impact of fine-tuning different components of the model’s architecture on its performance.

We test the model on the simulated benchmark dataset fall detection dataset University of Rzeszow Fall Detection (URFD) and on the Real-world Fall Dataset (RFDS) dataset, which includes unsimulated real-world scenarios. Under the given experimental settings, The model achieved 94% accuracy in identifying falls and non-falls. In real-world scenarios, it identified 97% of falls but only 5% of Actions/Activities of Daily Living (ADL) scenarios, resulting in an overall accuracy of 51%.

These results highlight the need for diverse real-world fall datasets to improve the robustness of FDS.