When Vision Transformer Meets Fall Detection: A Transferable Representation Learning
Chongxiao Qu, Lei Jin, Yongjin Zhang, Changjun Fan, Shaojie Xia, Shuo Liu · 2023
Fall-related injuries, especially among the elderly, are a major global health concern due to the current aging trend. Utilizing dependable fall detection systems can mitigate the negative outcomes associated with falls. This paper examines fall detection using the UP-Fall Detection dataset, which is a publicly available dataset collected from 17 volunteers using various multi-modal sensors. We employ a Vision Transformer-based (ViT) model to distinguish the images in the dataset depicting various activities including falls. To reduce computing complexity, we fine-tune a pretrained ViT model from Torchvision. During the process, the learned representative features in the pre-trained model are used directly, and we just need to modify the classifier and learn its parameters by training. Simultaneously, we implement a state-of-the-art (SOTA) model from the literature on the same dataset and compare it with our proposed method by conducting extensive experiments. The outcomes of the experiments demonstrate that our approach surpasses the SOTA method on this dataset, as indicated by enhancements in crucial performance metrics such as accuracy, precision, recall, and F1-score.