Multi-Head Attention-Based 1DCNN-LSTM Networks for Human Activity Recognition
Liang Long, Lizuo Jin · 2025
Human Activity Recognition (HAR) has wide applications in areas such as sports training, medical rehabilitation, military, and video games. Human motion recognition based on wearable sensors has intensified. Currently, the predominant methodologies employed are deep learning approaches such as Convolutional Neural Network (CNN) and Recurrent Neural Networks (RNN). However, their performance still needs to be improved in recognizing actions with similar features.This paper proposes a 1DCNN-LSTM structure based on a multi-head attention mechanism. Building upon a 1D CNN, the approach incorporates a bidirectional Long Short-Term Memory (BiLSTM) network with residual connections, thereby enhancing the model's ability to recognize complex features. The multi-head attention mechanism is employed to refine and optimize the feature representations, thereby improving the network's overall accuracy. On the open-source dataset KU-HAR, this network achieved an accuracy of 98.18 %, outperforming state-of-the-art models (DeepConv, AM-DLFC, ResNet), delivering excellent experimental results.