Fall detection based on improved capsule network
Yanshi Liu, Minghui Yao · 2024
Aiming at the limitations of traditional machine learning in data feature extraction, an integrated Capsule Network and multi-layer Bidirectional Gated Recurrent Unit (Bidirectional Gated Recurrent Unit) were proposed. BiGRU and the BACN fall detection and recognition model of Attention mechanism. In this model, the capsule network is responsible for capturing the spatial features, while the bidirectional GRU module effectively extracts the hidden temporal features of the data. At the same time, the attention mechanism is used to further highlight and screen the fine-grained features to improve the accuracy of model recognition. Through repeated experimental training and model optimization, we determined the optimal hyperparameters of the network and constructed an efficient fall detection model. The accuracy of the model on Sisfall, Mobiact fall data set and self-collected data set reached 98.3%, 97.8% and 94.4% respectively, which fully proves its effectiveness in practical application.