Transformers and CNNs in Ensemble: Advancing Multimodal Fall Detection
Sarita Sahni · Journal of Information Systems Engineering & Management · 2025
In today’s scenario, where the detection and prevention of falls are important for the safety of elderly and high-risk people, it is also important to have a high reliability in detecting such cases. This study introduces a novel Fall Detection System (FDS) that combines Transformer and Convolutional Neural Networks (CNNs) in an ensemble fashion for multimodal analysis. The proposed architecture utilizes Transformer encoders for the long-range dependencies of modality-specific and cross- domain feature representations whereas local features are obtained from three sensor modalities using CNNs. An adaptive fusion mechanism combines the learned features from different modalities to enhance detection accuracy. This mechanism effectively integrates relevant information from each modality, allowing the model to capture and utilize complementary insights from all available sensors. Extensive evaluations of publicly available datasets SisFall and KFall, demonstrate the effectiveness of the proposed model. Our approach achieves superior performance, with an accuracy of 0.9632 ± 0.020 on the SisFall dataset and 0.9684 ± 0.0208 on the KFall dataset, highlighting its robustness in fall detection.