Incorporating Attention Mechanism Into CNN-BiGRU Classifier for HAR
Ohoud Nafea, Wadood Abdul, Ghulam Muhammad · IEEE Access · 2024
Human activity recognition (HAR) has become an increasingly important problem over the past few years. The HAR system is primarily intended to assist healthcare providers and elderly care facilities using other technologies such as the Internet of Things (IoT). Standard pattern recognition techniques have developed significantly, but they continue to depend on heuristic and manual feature extraction, which undermines the generalization performance. Automatic high-level feature extraction has become possible using deep learning to optimize performance. For sensor-based HAR, deep learning techniques have also been applied in a variety of fields. The proposed methodology uses convolutional neural networks (CNN) and recurrent neural networks (RNN) to extract the spatial and temporal features. We propose to employ an attention mechanism based BiGRU. Attention mechanisms can improve the model performance by focusing on particular features or timesteps which are valuable for distinguishing different forms of activity. Various datasets have been implemented for these models in which data collection is performed using diverse methods including accelerometers, sensors, and gyroscopes. Comparative analyses of the proposed models demonstrate their effectiveness and advantages. The UCI-HAR, WISDM, and mHealth datasets were used in experiments to evaluate the proposed method, and the results indicated a high accuracy of 94.94%, 98.57% and 99.18%, respectively.