A Fall Detection Method Based on Channel Attention and Transformer for Wearable Sensor Data

Abstract

Fall detection has become an important research topic in healthcare monitoring systems due to the increasing aging population. In this paper, a fall detection method based on channel attention and Transformer is proposed for multivariate sensor data. The proposed model employs a one-dimensional convolutional neural network (1D CNN) to extract local temporal features, followed by a Squeeze-and-Excitation (SE) module to enhance channel-wise feature representation. A Transformer encoder is then introduced to capture long-range temporal dependencies and model the dynamic characteristics of fall events.

Experimental results on a public dataset demonstrate that the proposed method outperforms several baseline models, including CNN, CNN+LSTM, and CNN+Transformer, in terms of accuracy, precision, recall, and F1-score. The improvements indicate that the integration of channel attention and global temporal modeling effectively enhances fall detection performance.

Furthermore, the proposed model maintains a relatively efficient structure, making it suitable for relatively practical applications in wearable devices and smart healthcare systems. The results confirm the effectiveness and practicality of the proposed approach.