1-DCNN with Stacked LSTM Architecture for Human Activity Recognition Using Wearable Sensing Data
P. Krishnaleela, R. Meena Prakash · IETE Journal of Research · 2024
Human activity recognition (HAR) has grown more important in the domains of pervasive computing, human behavior analysis, assistive health care, and Human–Computer Interaction (HCI). While Deep Learning (DL)-based methods have demonstrated remarkable efficiency in the identification of activities, the management of time series data remains a challenge. Nevertheless, a large number of models currently in use direct the effective extraction of temporal and geographical information from data on human behavior. In this paper, a fusion of a 1-dimensional convolutional neural network (1-DCNN) and Stacked long short-term memory (SLSTM) is proposed for HAR. This model automatically extracts geographical information from the raw body sensor data. 1-D CNN model can extract local features and the SLSTM model can extract long-term relationships in sequence data. In addition, a batch normalization layer was added after the convolution layer to reduce internal covariate shift, accelerate training, and reduce the need for dropout. Moreover, a Max Pooling Layer was applied next to the pooling layer to reduce model parameters. The final step involves applying the learned features to the Softmax layer to fully recognize human activities. The model performance is evaluated on two standard datasets MHEALTH and WISDM and achieves classification accuracies of 99.67% and 99.55% correspondingly. The results demonstrate that the recommended model outperformed the state-of-the-art methods.